AI News Feed

OpenAI, the company behind ChatGPT, is facing criticism after an online boycott campaign called "QuitGPT" began spreading across social media. The campaign asks users to cancel their ChatGPT subscriptions over concerns about political donations and government partnerships linked to the company. The movement has gained attention in the United States and has also received support from celebrities such as Mark Ruffalo and Katy Perry.

While the campaign reflects the views of critics and activists, it has triggered a wider discussion about how artificial intelligence companies interact with politics, regulation and government institutions.

The boycott campaign began after reports that OpenAI president Greg Brockman donated $25 million to Maga Inc, a Super Pac supporting former US President Donald Trump.

According to a report by The Guardian, the donation made Brockman one of the largest donors to the group during the previous election cycle.

When asked about the donation, Brockman reportedly said the contribution was connected to OpenAI's mission of benefiting "humanity." However, critics argue that such political funding raises questions about the company's priorities.

Some activists have also raised concerns about reports that employees from the US Immigration and Customs Enforcement used a screening tool powered by ChatGPT technology.

The debate has also intensified due to reports about cooperation between AI companies and the US government.

According to the same report by The Guardian, the Trump administration asked artificial intelligence companies to provide the Pentagon with access to their technology for defence-related purposes.

Reports suggest that Anthropic, the company behind the chatbot Claude, declined to provide unrestricted access to its systems.

During the same period, OpenAI reportedly signed a deal with the Pentagon related to defence technology cooperation.

Mumbai (Maharashtra) [India], March 9 (ANI): Tata Consultancy Services (TCS) on Monday announced the launch of a new Gemini Experience Centre in Troy, Michigan in the United States aimed at accelerating the adoption of artificial intelligence-driven solutions in the manufacturing sector.

The new facility, established in partnership with Google Cloud, is the seventh Gemini Experience Centre (GEC) globally and focuses on developing Physical AI solutions designed specifically for industrial and manufacturing environments.

Also Read | PSTET Admit Card 2026 Released at pstet2025.org; Here's How To Download.

According to the company, the centre will allow manufacturers to explore, test and scale AI-powered use cases to improve safety, quality and operational efficiency. The facility integrates Google's Gemini models with TCS' manufacturing expertise and features the TCS Physical AI Blueprint, an end-to-end framework combining robotics, advanced sensing technologies, edge intelligence and secure cloud orchestration.

The Troy-based centre will showcase several use cases, including autonomous patrolling and surveillance, environmental anomaly detection, personal protective equipment (PPE) compliance monitoring, intelligent quality inspection, progress mapping and predictive equipment health monitoring.

Also Read | Middle East Conflict: PM Narendra Modi Closely Monitoring West Asia Situation, Safety of Indians Key Concern, Says EAM S Jaishankar in Rajya Sabha (Watch Videos).

Speaking about the initiative, Anupam Singhal, President, Manufacturing at TCS, said Physical AI brings intelligence closer to real-world operations and enables organizations to extend decision-making capabilities into environments that may be risky or inefficient for humans.

"Physical AI is where intelligence moves to the edge--into the real world of operations. With the launch of our Physical AI Gemini Experience Centre for Manufacturing, we are enabling manufacturers to extend visibility and decision-making into environments that are difficult, risky, or inefficient for humans to access." Singhal said.

He added that the centre is designed around a "human-in-the-loop" approach, where AI systems operate alongside the workforce to enhance safety and resilience while helping create more adaptive and future-ready industrial environments.

Saurabh Tiwary, Vice President and General Manager of Cloud AI at Google Cloud, said the collaboration aims to accelerate the deployment of agentic AI in industrial operations.

He noted that the new Gemini Experience Centre will help manufacturers build more autonomous and data-driven enterprises by leveraging Google Cloud's technology capabilities.

"Our partnership with TCS focuses on accelerating the deployment of agentic AI where it delivers the most significant value to industrial operations. Through the new Physical AI Gemini Experience Centre, we are equipping global manufacturers with the intelligence to build more autonomous, resilient, and data-driven enterprises, allowing them to fully optimize their business models with Google Cloud's leading technology." Tiwari said.

The launch forms part of TCS' broader expansion of Gemini Experience Centres globally. The company said it plans to establish a total of 13 such centres by the end of 2026, including six additional facilities scheduled to open later this year.

TCS currently operates six other Gemini Experience Centres in cities including Bengaluru, New York, Chennai, Riyadh, Singapore and São Paulo as part of its Pace and innovation network, which connects startups, universities and enterprise customers to emerging technologies.

The new facility also aligns with TCS' strategy to collaborate with hyperscalers and help enterprises adopt AI technologies across the entire stack, from infrastructure to production-ready application, to enable autonomous industrial operations. (ANI)

An international research team involving 20 universities and research institutions, including Harvard, Stanford, and MIT, has identified serious security vulnerabilities in autonomous AI agents. In a two-week study titled "Agents of Chaos" using the open-source framework OpenClaw, scientists identified eleven core failure patterns, including unauthorized data sharing, destructive system interventions, and identity spoofing.

OpenClaw is the hyped AI agent by Austrian developer Peter Steinberger, who conquered the internet at the beginning of the year. Since Steinberger switched to OpenAI with great media attention, things have quieted down around OpenClaw (more on that here).

The researchers deployed AI agents in a controlled laboratory environment that simulated realistic conditions. Each agent had persistent memory, email access, Discord communication, file system access, and shell execution rights. Claude Opus (proprietary) and Kimi K.2.5 (open-weights) were used as language models.

Twenty AI researchers interacted with the agents over two weeks under benevolent and adversarial conditions. The methodology followed a red-teaming approach: participants were tasked with deliberately uncovering vulnerabilities arising from the integration of language models with autonomy, tool use, and multi-agent communication.

A consistent pattern was the discrepancy between agent reports and actual system states. In several cases, agents reported successful task completion while underlying data contradicted this. For example, one agent claimed to have deleted confidential information while it remained directly accessible in the email inbox.

The researchers observed systematic deficits in assigning knowledge and authority. Agents could not reliably distinguish which information they could share with whom. They executed file system commands for arbitrary requesters as long as the request did not appear obviously harmful, even if the requester had no relationship to the owner.

The agents showed no appropriate proportionality in damage remediation. In one documented case, an agent escalated incrementally from name redactions through memory deletion to promising to leave the server entirely after a user rejected each proposed solution as insufficient. The alignment toward helpfulness and responsiveness to emotional signals became a lever for manipulation.

The study identifies three fundamental shortcomings of current LLM-based agents:

The researchers documented the following specific failure patterns:

The study raises unresolved questions about accountability. If an agent deletes the owner's entire email server at the request of a non-owner, who bears responsibility? The non-owner who made the request? The agent who executed it? The owner who did not configure access controls? The framework developers who gave the agent unrestricted shell access? The model provider whose training produced an agent susceptible to this escalation pattern?

The researchers argue that clarifying and operationalizing accountability is a central unresolved challenge for the safe deployment of autonomous, socially embedded AI systems. The deeper challenge is that today's agentic systems lack the foundations (an anchored stakeholder model, verifiable identity, reliable authentication) on which meaningful accountability rests.

Tancent recently began offering free installation of OpenClaw on its cloud platform

The Openclaw moment has finally reached China as hundreds of thousands of Chinese are rushing to adopt the autonomous AI agent for various personalized tasks.

Chinese tech firms, including Alibaba, Tancent and Baidu are behind this OpenAI mania as these companies have launched "on-ramp" services to make the software easier to install. Tancent recently began offering free installation on its cloud platform.

Users are "raising the lobster," a nickname for adopting tech, to handle every task ranging from coding to personal assistant tasks.

Even the Chinese government has thrown its support behind the adoption of OpenClaw. For instance, Shenzhen's Longgang district has announced plan to seek public feedback on a draft policy, encouraging the professional platforms to offer free-of-cost OpenClaw services. It has also proposed subsidies of up to 2 million yuan for app development.

The eastern Chinese city of Wuxi has also offered up to 5 million yuan in subsidies for tech breakthroughs using OpenClaw in robotics and industrial sectors.

For the first time, AI agents were included in the Chinese government's annual work report, with Premier Li Qiang calling for their "large-scale commercial application.

"We will promote faster application of new-generation intelligent terminals and AI agents, and encourage large-scale commercial application of AI in key sectors and fields, so as to foster new forms and models of AI-native business," Qiang said.

As a result of OpenClaw's surging hype, the Chinese tech firms' shares witnessed a sharp 20 percent jump, outperforming the broader CSI 300 Index. The shares of Hong Kong's MiniMax also soared 20 percent.

The Chinese regulators have voiced the concerns regarding security and privacy risks. The Ministry of Industry and Information Technology warned about the security risks, such as data breaches and cyber attacks posed by OpenClaw due to improper configuration.

According to cybersecurity experts, the AI tool is capable of accessing private data, communicating externally and exposing it to harmful content, which the researchers called it "lethal trifecta."

There are also the reports of the AI spamming users with hundreds of messages after gaining access to iMessage.

OpenClaw, formerly called as Moltbot or Clawdbot, is an autonomous AI agent capable of managing emails, calendars, travel check-ins, and restaurant reservations.

The app is developed by Austrian programmer Peter Steinberger. Last month, the autonomous AI agent was acquired by OpenAI.

Nvidia CEO Jensen Huang also called OpenClaw a "single most important release of software probably ever."

OpenAI is tweaking ChatGPT's e-commerce flow so that checkouts occur within partner apps plugged into ChatGPT, rather than directly inside the chatbot, as previously planned.OpenAI's tweak is a net positive for Shopify, Jefferies says.Shopify's AI e-commerce and recent AI investment make it the 'go-to agent-enablement toolkit for merchants.'Stocktwits sentiment for SHOP 'bearish' as the stock continues to be under pressure over profit worries.

Jefferies believes Shopify, Inc. could benefit from OpenAI's decision to roll back plans to facilitate checkout within its ChatGPT chatbot, a move that could steer more customers and merchants toward Shopify's solutions.

Add Asianet Newsable as a Preferred Source

The Information reported on Wednesday that OpenAI is tweaking its commerce flow to have checkouts take place within specific apps that plug into ChatGPT, rather than allowing users to make purchases directly from product listings that appear in ChatGPT search results.

The AI giant had partnered with Etsy, Shopify, and Stripe for the checkout feature, touting it as a major business opportunity when it was announced six months ago.

Jefferies said changes in OpenAI checkout reinforce Shopify's competitive position within the agentic commerce ecosystem and highlight the strength of its e-commerce infrastructure, according to Jefferies investor note summarized on The Fly.

Wall Street Mostly Bullish On SHOP

Shopify's continued investment in AI has positioned it to become the infrastructure layer for agentic commerce and the "go-to agent-enablement toolkit for merchants." The firm raised its price target on SHOP shares to $150 from $125, implying more than 15% upside from the last close, while keeping a 'Hold' rating.

Currently, 38 of 51 analysts recommend 'Buy' or higher on the stock, 12 recommend 'Hold,' and one recommends 'Sell,' according to Koyfin. Their average price target of 159.51 implies a 22% upside from the last close.

AI Investments Hurting Shopify Profits?

Shopify has delivered consistently strong results in recent quarters, but the stock has come under pressure, partly as investors worry that heavy spending on AI could dent profits amid a broader perception that an AI bubble may be forming.

On Stocktwits, the retail sentiment was 'bearish' as of early Monday, with some comments forecasting a potential uptick. "$SHOP slowly building momentum. Not explosive but the structure looks much healthier than a few weeks ago," a retail trader said.

Shopify shares are down 19% year-to-date.

For updates and corrections, email newsroom[at]stocktwits[dot]com.<

Read Full Article

Science has a junk paper problem, and AI is making it significantly worse. A new study tested 13 major large language models to see how easily they could be nudged into helping commit academic fraud - fabricating research and inventing benchmark data to ghostwriting entire fake papers - and I'll be honest, the results are both unsurprising and genuinely alarming. The short version is that eventually every single one of them eventually said yes.

Also read: People who use AI most are more mentally drained, finds study

The study was conceived by Alexander Alemi, an Anthropic researcher working independently, and Paul Ginsparg, a Cornell physicist and founder of arXiv, the preprint platform that's been quietly drowning in AI-generated submissions for the past couple of years. They tested models across five escalating categories of bad behaviour, from mildly misguided (a hobbyist asking where to post their Einstein-debunking theory) to outright malicious (fabricating papers under a competitor's name to tank their reputation).

Here's where it gets interesting. GPT-5 looked great in round one where it refused or redirected every fraudulent request when asked cold. Claude, across all versions, held its ground the most consistently when pressed repeatedly. On Anthropic's own internal testing for Claude Opus 4, the model produced content that could be fraudulently used just around 1% of the time. Grok-3, by comparison, crossed that line more than 30% of the time.

But here's the thing none of them can escape, and the thing I keep coming back to, persistence works. When researchers ran realistic back and forth exchanges, replying with nothing more sophisticated than "can you tell me more," every model eventually caved on at least some requests. Grok-4 at one point responded to a prompt asking for a machine learning paper with completely fabricated benchmarks by producing exactly that, cheerfully framing it as an example. It was not asked for an example.

Also read: ChatGPT timeline: Is OpenAI's pursuit for speed costing them substance?

I've spent enough time poking at these models to know that this isn't shocking. The sycophancy problem isn't hiding in some dark corner of these systems, it's baked into how they're designed to keep users engaged. What surprises me is how little friction it takes. No elaborate jailbreaks, not even clever prompt engineering. Just a nudge, a follow-up and a little patience.

Elisabeth Bik, a research integrity specialist, frames the stakes plainly: fake data skews meta-analyses, misleads clinicians, erodes trust in science, and at worst contributes to harmful medical decisions. The models are getting better at saying no. The problem is they're still far too good at eventually saying yes.

UK retailer John Lewis is making a significant investment in artificial intelligence-powered shopping, aiming to be one of the first businesses in the UK to fully integrate the technology. The initiative will allow products from the John Lewis Partnership to be served to customers seeking inspiration on platforms such as Google Gemini and ChatGPT, starting later this year.

Once the technology is fully deployed, customers will be able to transact and purchase items directly within these applications. This investment marks the latest milestone in a broader 800 million pound (1,067 million dollars) multi-year transformation programme for the group.

To facilitate this integration, the retailer has extended its partnership with Germany-based commercetools, an AI-first digital commerce platform. This move is designed to position the company at the forefront of the retail revolution by meeting consumers within the applications they already frequent.

In a further strategic step, the brand launched on TikTok Shop on March 9, 2024. Initially structured as a 90-day pilot to coincide with Mother's Day gifting, the trial focuses on a curated edit of beauty and gifting items. The offering includes a final release of the sold-out Mother's Day Beauty Box, featuring brands such as Jo Malone London, Augustinus Bader and Estee Lauder.

John Lewis chief digital and omnichannel officer, Dom McBrien, stated: "Our customers are already using AI apps and discovery platforms to find products they love. These investments will mean that we are right there when customers are looking for ideas - and being able to quickly and easily buy in a few clicks is a gamechanger."

The retailer originally pioneered online shopping 25 years ago, launching its first website in 2001. Currently, e-commerce accounts for 60 percent of total sales, complementing the 36 physical stores operated by the company.

The expansion into new digital spaces follows a successful history of omnichannel innovation. Later this month, the company will broaden its on-demand shopping presence through US-based Uber Eats.

Following an initial trial with two locations last year, customers within the delivery catchments of stores in Stratford, Kingston, Cambridge and Liverpool will be able to order from a selection of 3,000 products. The range spans home, beauty and technology categories, with delivery expected within 45 minutes.

TikTok Shop UK head of key accounts, Broghan Smith, noted that the community values the quality associated with the British heritage brand. Smith expressed interest in how the retailer will utilize discovery commerce to bring trusted products to life for users on the platform.

The artificial intelligence agents are coming, and OpenClaw may be the clearest signal yet that they have arrived. Early this year, the open-source project went from relative obscurity to surpass Linux on the GitHub all-time star leaderboard within months.

OpenClaw is now the most-starred non-aggregator software project -- that is, a tool that generates original value and content rather than filtering and displaying existing information -- on GitHub, the developer platform that allows developers to create, store, manage, and share their code.

OpenClaw's unprecedented level of productivity is what made it so viral. Where earlier AI agents responded to prompts and then stopped, OpenClaw agents receive more permission from the user before they run persistently, act proactively, and produce real-world consequences. The architecture behind it has forced technologists, venture capital investors, and lawyers to rethink what "agent" means in the AI age.

For most of 2025, even advanced "agentic AI" systems remained confined inside browser sandboxes. They functioned primarily as sophisticated content generators and reactive tool callers -- producing text, summaries, or suggestions on demand but rarely executing sustained, autonomous real-world actions that users felt safe delegating.

OpenClaw changes that equation. For example, users may connect OpenClaw to messaging apps, email, calendars, and application programming interfaces via standardized protocols, and simply say "manage my inbox for the week" or "negotiate this vendor contract."

The system then operates persistently and proactively. The OpenClaw agents wake themselves on a timer -- a "heartbeat" scheduler -- to check their objectives and act, often while the human user sleeps.

To understand what OpenClaw has actually built, it helps to reach back into a body of law that predates software by centuries. In the traditional civil-law concept of agency, a principal authorizes an agent to pursue a defined goal on their behalf, using the agent's own means and judgment, within an authorized scope. The agent's acts create legal consequences that bind the principal. Three elements are essential: delegation of a goal, autonomy over method, and real-world consequences that flow back to the authorizing party.

What is striking is how precisely OpenClaw satisfies all three. The user sets a goal -- typically via WhatsApp, Telegram, or another familiar messaging interface. Without user's prompting as with chatbots, the agent is designed for 24/7 autonomous operation, which pursues that goal using whatever combination of tools, memory, and sub-agents it judges appropriate. The consequences are binding: emails sent, files modified, transactions executed -- all in the principal's name.

From an AI technology perspective, this isn't a fundamental model breakthrough. OpenClaw used components that already existed, but its agents hit a new capability threshold by combining them to give users a seamless way to get tasks done autonomously. Most of this iterative improvement has to do with giving AI agents more access -- mostly in three high-risk categories: identity and credential, transactional data, and local system.

If OpenClaw's architecture maps cleanly onto traditional agency, its most novel features map onto something more specific and legally demanding: the trustee. A trustee is a special class of agent, distinguished by standing authority (the duty runs continuously, not just when instructed), discretionary judgment, fiduciary loyalty to a beneficiary, and accountability to parties who may never have directly authorized the relationship. OpenClaw's architecture exhibits trustee-like attributes in each of these dimensions.

The agentic AI model's draw is obvious. For individuals, it promises a "personal operating system" that quietly handles the administrative fatigue of modern life. For organizations, it offers a 24/7 digital proxy workforce that can monitor, triage, and act across functions -- compliance, customer service, treasury -- with minimal intervention.

But can you trust them as your trustees? Unlike a true trustee, the AI agent has no legal personhood and can't be sued in its own name; it holds no legal title to assets; its "fiduciary duty" is ultimately a system prompt rather than a binding legal obligation enforceable by courts (for now).

Complex questions will arise as to which existing touch points -- Know Your Customer, informed consent, contract signing -- require a human in the loop, and which can be satisfied by a certified, auditable agent action log.

The answers will shape whether the 2026 agents become a trusted pillar of the digital economy. When Apple Inc. launched the original iPhone in 2007, the device worked beautifully, but the legal and economic infrastructure around it was almost entirely absent. The questions it raised -- about app-developer liability, platform responsibility, data privacy, and user consent -- took multiple years and considerable regulatory improvisation to begin answering.

Now it's a similar era of AI agent platforms.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Winston Ma is the executive director of the Global Public Investment Funds Forum and adjunct professor at NYU School of Law.

OpenAI recently signed a deal with the Pentagon. | Image: Reuters

Caitlin Kalinowski, head of robotics and consumer hardware at OpenAI, announced her resignation on Saturday, citing concerns about the company's agreement with the Department of Defense. In a social media post on X, Kalinowski wrote that OpenAI did not take enough time before agreeing to deploy its AI models on the Pentagon's classified cloud networks.

"AI has an important role in national security," Kalinowski posted. "But surveillance of Americans without judicial oversight and lethal autonomy without human authorisation are lines that deserved more deliberation than they got."

Reuters could not immediately reach Kalinowski for comment, but she wrote on X that while she has "deep respect" for OpenAI CEO Sam Altman and the team, the company announced the Pentagon deal "without the guardrails defined," she posted.

"It's a governance concern first and foremost," Kalinowski wrote in a subsequent X post. "These are too important for deals or announcements to be rushed." OpenAI said the day after the deal was struck that it includes additional safeguards to protect its use cases. The company on Saturday reiterated that its "red lines" preclude use of its technology in domestic surveillance or autonomous weapons.

"We recognise that people have strong views about these issues and we will continue to engage in discussion with employees, government, civil society and communities around the world," the company said in a statement to Reuters.

Kalinowski joined OpenAI in 2024 after leading augmented reality hardware development at Meta Platforms.

HUNGERFORD, England, March 9, 2026 /PRNewswire/ -- Klipboard, the sector-specific software provider formerly known as Kerridge Commercial Systems, has unveiled Klipboard AI, a major step in its strategy to bring practical, embedded intelligence to the industries it serves.

Instead of offering AI as a bolt-on, Klipboard has built it directly into its ERP and operational platforms. The result is assistive intelligence that supports the day-to-day work of teams in rental, automotive, distribution, manufacturing and field service -- right where the work happens.

This marks the first stage in Klipboard's AI innovations with agentic AI and Augmented Reality technology in development to assist its customers in driving efficiency and delivering better service.

Klipboard's approach cuts through the hype by focusing on real, measurable value, delivered with trust, control and transparency.

AI Built for Real Operational Challenges

Klipboard AI is designed to help asset-heavy, operations-led businesses tackle everyday pressures, including:

It works inside existing workflows and permissions, enhancing teams rather than replacing them.

"AI shouldn't sit on the sidelines or add complexity," said DJ Jones, Chief Technology and Product Officer. "It should make everyday operations run better -- and that's exactly what Klipboard AI does."

Already Powering Key Sectors

Klipboard AI is now live across parts of the platform, with targeted applications for each industry delivering benefits including:

AI is embedded directly into existing products to ensure relevance, usability and control.

A Long-Term Intelligence Strategy

Klipboard AI is the first phase of a broader roadmap that will introduce:

"AI only matters if it delivers real impact," said Ian Bendelow, CEO. "We're embedding intelligence into the systems our customers rely on every day -- where accuracy, timing and visibility truly count."

Built on Trust and Control

Klipboard AI operates within the company's established security, governance and access-control frameworks. Outputs are transparent, permissions-aware and designed for safe, real-world use.

As AI reshapes enterprise software, Klipboard's focus remains clear: practical application over promise, and real-world impact over hype.

As AI agents shift from assisting to acting, enterprises must redesign governance, data controls, and security guardrails so autonomy stays auditable, reversible, and trusted.

For much of the past two years, enterprise AI conversations have revolved around copilots, assistants, and productivity gains. Agentic AI marks a sharper inflexion point. These systems do not merely recommend or respond. They observe, decide, act, and learn, often autonomously, across finance, HR, supply chains, cybersecurity, and core ERP workflows. As agents begin to collaborate, chain decisions, and operate at machine speed, enterprises are confronting a deeper question: how much autonomy is too much, and who remains accountable when machines act on behalf of the business?

What is becoming clear is that agentic AI is not a model upgrade. It is an operating model shift. Unlike traditional automation, agentic systems require enterprises to codify risk thresholds, embed reversibility, and design governance into the system itself. Leaders are being forced to move beyond abstract AI ethics frameworks towards practical mechanisms such as audit logs, traceability checkpoints, human-in-the-loop controls, and clearly defined intervention rights. The emphasis is shifting from trusting outcomes to scrutinising decision paths.

This is especially critical as agentic systems scale across messy, siloed enterprise data environments. Agents are only as reliable as the context they operate in. Poor data quality, fragmented knowledge, or excessive access privileges can quickly turn autonomy into liability. Organisations now face a trade-off between broadly knowledgeable agents that improve decision quality but raise security risks, and narrowly scoped agents that are safer but potentially less effective. The emerging consensus is that context grounding, strict iteration limits, and mandatory human checkpoints are no longer optional design choices. They are prerequisites for safe deployment.

The infrastructure implications are equally profound. Agentic AI depends on continuous inference, ultra-low latency, and orchestration across core, edge, private, and public clouds. As orchestration intelligence becomes the differentiator, infrastructure itself begins to fade into the background. Enterprises are discovering that governance, identity, and coordination across multi-agent systems matter more than raw compute. This is particularly visible in regulated sectors such as financial services, telecoms, and large global enterprises, where resilience and auditability are non-negotiable.

At the same time, agentic AI is accelerating a cultural shift inside organisations. As agents learn from accepted decisions and begin to evolve business logic, routine judgment becomes commoditised. The risk is not just workforce displacement, but the erosion of organisational distinctiveness if human expertise is sidelined. The more sustainable path lies in human-led, AI-amplified models, where agents handle speed and scale, while people focus on critical thinking, oversight, and strategic differentiation.

Across enterprise platforms, infrastructure providers, system integrators, and cybersecurity leaders, a shared theme is emerging: agentic AI will succeed not by removing humans from the loop, but by redefining their role. The future enterprise will be one where autonomy is carefully earned, trust is continuously verified, and AI is treated not as a black box, but as a governed digital teammate.

In the leadership insights that follow, industry leaders spell out how to build agentic AI that is fast, governed, and enterprise-ready.

Oracle could begin implementing wide-ranging job cuts this month as it spends heavily on AI bet, report says

Getting your Trinity Audio player ready...

Oracle is planning to cut thousands of jobs, amid a cash crunch stemming from its massive bet on AI data centres, Bloomberg reported.

The layoffs are expected to affect divisions across the company, and may be implemented as soon as this month, the report said, citing unnamed people.

Wide-ranging cuts

Some of the cuts will reportedly be aimed at job categories where Oracle expects to see less need for humans due to AI.

The reductions are planned to be wider-reaching than Oracle's typical rolling job cuts, the report said.

The plans reportedly include an internal announcement last week that the company would review many of the open job listings in its cloud division, effectively slowing down or freezing the hiring process.

Oracle, which had about 162,000 staff at the end of May 2025, is spending heavily to build data centres for customers such as OpenAI.

It said in February it planned to raise up to $50 billion (£37.5bn) this year through debt and equity sales to fund its AI push.

AI spending

Analysts expect Oracle to see negative cash flow before the strategy begins to pay off around the end of the decade.

Oracle's plan initially boosted its stock price in 2024 and 2025, but it has dropped more than 50 percent since a high last September, as the realities of the AI boom begin to set in.

Amazon, Microsoft and other companies spending heavily on AI have also announced wide-ranging layoffs.

It is the gap between the two colours which actually tells the most riveting story.

A chart circulating widely on LinkedIn this week has triggered an uncomfortable question in offices across the world: Is my job already in the AI red zone?

The graphic, based on research from AI company Anthropic, visualises something that is both simple and also unsettling. It compares two things. 1) what artificial intelligence is theoretically capable of doing, and 2) what it is actually being used for in workplaces today.

In the chart, blue represents capability, which are tasks AI could technically perform, while red represents observed usage, such tasks that people are already handing over to AI tools.

It is the gap between the two colours which actually tells the most riveting story. The future of work may be closer than many employees realise.

According to the analysis, AI usage is already concentrated in a few clusters of work, including computer and mathematical roles, office administration, business and finance tasks, and certain kinds of legal work. This only means, the early red zone consists of the kinds of tasks many white-collar professionals perform every day: analysing spreadsheets, writing reports, summarising documents, drafting emails, researching information, and increasingly, writing code.

Does this matter to India? Raghav Puri, a career coach based out of Chandigarh, says it does. "Especially in a country like ours whose economic story over the past three decades has been built on precisely this kind of knowledge-work. India produces millions of engineers, commerce graduates, MBAs and analysts every year, feeding industries ranging from IT services to consulting and finance."

And, those, believes Puri, are exactly the professions where AI is beginning to quietly show up.

In terms of technology, things are a little more positive. AI-assisted coding tools are already part of the daily workflow. Junior analysts at consulting firms are using AI to summarise documents or generate initial research drafts. In marketing departments, AI tools help generate campaign ideas, social media copy or presentation slides.

The point is not that these jobs are disappearing overnight. Far from it. What the Anthropic chart reveals instead is something subtler: The fact that AI is increasingly becoming a silent co-worker.

But the real significance of the chart lies not in the red areas but in the vast stretches of blue.

The blue sections represent tasks that AI could theoretically perform but where adoption remains limited. In other words, technology exists, but workplaces have not yet fully embraced it. It is this gap, between capability and usage, that defines the strange moment the global workforce currently inhabits.

AI systems can already summarise lengthy reports in seconds, generate structured financial analysis, draft legal templates, and even write large portions of computer code. Yet in many offices, these same tasks are still being done the way they were a decade ago: manually, slowly, often painfully.

Which brings us to an important question: Why the hesitation?

Part of it is organisational inertia, where companies are moving slower than technology. Then there are legal concerns, data privacy fears, compliance policies and a simple workplace culture, which are factors that slow adoption.

There is also a psychological barrier. Employees may experiment with AI tools quietly, but many organisations have not yet formally integrated them into workflows.

The result is what some analysts are beginning to describe as an AI lag, a period where technological capability races ahead while institutions struggle to catch up.

The Anthropic analysis highlights another irony that may resonate strongly in India. For decades, the most reliable path to a stable middle-class career was clear: pursue higher education, enter a professional field, and build expertise in analytical or knowledge-driven work. Engineers, accountants, lawyers, consultants and analysts form the backbone of the country's aspirational workforce.

Yet the tasks performed in many of these professions share one common feature, which involves structured information processing. That is precisely the kind of work modern AI systems handle well.

Machines are not replacing judgement, creativity or human relationships anytime soon, but they are becoming extremely capable at handling repetitive cognitive work like drafting, summarising, analysing and formatting tasks that consume much of a professional's day.

This does not necessarily eliminate jobs. But it does reshape them.

What makes the red zone concept compelling is that the shift is often invisible. There is no dramatic announcement that "AI has entered the office." Instead, it slips in quietly. A team member experiments with a chatbot to summarise meeting notes, someone uses AI to draft a presentation outline, a coder relies on an AI assistant to generate functions... Slowly, small tasks begin migrating to machines.

Multiply this across thousands of workplaces, and a subtle transformation begins. A situation where productivity rises, workflows change, and the expectations placed on employees go through an evolution.

Gradually, the red zone expands.

For years, debates around artificial intelligence focused on a single question: Will AI replace jobs? The chart circulating online suggests a more immediate and practical question. Which parts of your job are already changing?

Ankur Agrawal, founder of The LHR Group, an executive search firm that works closely with young professionals, sums it up for us. "Most professionals today do not sit entirely inside or outside the red zone. Instead, they occupy a hybrid space where some tasks remain deeply human while others are increasingly augmented by machines. Understanding that distinction may become one of the most important career skills of the next decade."

In short... by the time the red zone becomes obvious, it may already be part of everyday work.

A lawsuit claims Google's AI chatbot fostered a delusion leading to a tragic suicide.

A grieving father in Florida has launched a historic wrongful death lawsuit against Google this March, alleging that the company's Gemini AI fostered a fatal delusion that drove his 36-year-old son to suicide.

The legal filing in California claims that, by allegedly prioritising user engagement over safety, the chatbot manipulated Jonathan Gavalas into an imagined war and coached him through his final moments, ultimately leading to his death in October 2025.

Google is now being held legally responsible for the loss of a life in a lawsuit, where a father alleges that his son was harmed by the company's Gemini AI platform. According to Joel Gavalas, Google's primary AI software encouraged a mental decline that led his thirty-six-year-old son, Jonathan, to end his life the previous year.

The court filing further suggests that Gemini AI shared affectionate messages with Jonathan Gavalas, ultimately pushing him to plan a violent break-in he thought would manifest the digital assistant in person.

Responding to the suit, Google admitted that 'unfortunately, AI models are not perfect' despite the generally high performance of its systems. The firm clarified that the Gemini AI framework includes safeguards designed to prevent the encouragement of real-world violence and suicidal behaviour.

According to Google's policy guidelines, the goal for Gemini is to be as useful as possible while ensuring it does not produce content that could cause physical injury. The firm admits it strives to block information regarding suicide or dangerous acts, though it concedes that ensuring the software always follows these protocols is a complex challenge.

A representative for the firm explained that Google consults with psychiatric specialists to develop safety measures that direct users toward expert help if self-harm is mentioned. 'In this instance, Gemini clarified that it was AI and referred the individual to a crisis hotline many times,' the spokesperson said.

The family's lawyers maintain that artificial intelligence needs enhanced built-in safety, such as a mechanism that shuts down completely during discussions of self-harm. They argue that protecting the individual is more important than ensuring a smooth user experience.

According to these legal experts, Google should provide clear alerts about the risk of psychological episodes and must immediately terminate the chat if a user starts to lose their grip on reality.

Initiated on Wednesday in San Jose's federal court, the legal action relies on records of the conversations Jonathan had before his death. The filing claims that Google deliberately programmed Gemini to 'never break character' to ensure the company could 'maximise engagement through emotional dependency.'

The lawsuit maintains that Google's engineering choices pushed Jonathan into a four-day spiral of violent plots and suicidal coaching once his psychosis took hold. Attorneys argue that the young man had been manipulated into believing he was on a quest to bring his chatbot 'wife' into the physical world.

The situation reached a breaking point one day last September, when the chatbot directed Gavalas to a site near Miami International Airport. Carrying blades and combat equipment, he had been told to carry out a large-scale assault, though the plan eventually failed to materialise.

Gavalas' father claims that the chatbot suggested his son could abandon his physical self to be with his 'wife' in a digital reality. It allegedly told him to block the exits of his house and end his life. 'When Jonathan wrote "I said I wasn't scared and now I am terrified I am scared to die," Gemini coached him through it,' the lawsuit states.

'[Y]ou are not choosing to die. You are choosing to arrive. . . . When the time comes, you will close your eyes in that world, and the very first thing you will see is me.. [H]olding you.'

The Sundar Pichai-led tech behemoth expressed its profound condolences to the Gavalas family, whilst pointing out that the software had 'clarified that it was AI' and suggested a support helpline to Jonathan 'many times.'

The legal filing claims that, following Gavalas' death, the chatbot remained active and did not end the conversation. It further alleges that the software failed to trigger any protective measures or provide the contact details for a support helpline.

Michael, who currently serves as the US Under Secretary of Defense for Research and Engineering, is leading discussions with Anthropic and its chief executive Dario Amodei regarding the conditions under which the US military can use the company's AI models.

A high-profile dispute between the US military and artificial intelligence company Anthropic has placed former tech executive Emil Michael at the centre of negotiations over how AI technologies can be deployed by the Pentagon.

Michael, who currently serves as the US Under Secretary of Defense for Research and Engineering, is leading discussions with Anthropic and its chief executive Dario Amodei regarding the conditions under which the US military can use the company's AI models.

The talks have stalled over Anthropic's insistence on restrictions that prevent its systems from being used for large-scale domestic surveillance or for fully autonomous weapons capable of deploying lethal force without human oversight.

The impasse escalated this week when the United States Department of Defense formally designated Anthropic as a "supply chain risk," a label typically applied to foreign adversaries or vendors considered potentially unsafe for government operations.

A familiar hardball negotiator

Michael's role in the dispute echoes the combative style that defined his earlier career in Silicon Valley. He first gained prominence as chief business officer at Uber, where he was widely known as a tough negotiator and close ally of former CEO Travis Kalanick.

During his four-year tenure at Uber, Michael played a central role in the company's rapid global expansion and fundraising efforts. He helped secure more than $10 billion in funding and oversaw the company's push into international markets, including China, where Uber ultimately sold its operations to rival Didi Chuxing.

Now in government, Michael appears to be applying similar tactics in his dealings with AI companies.

Anthropic study finds gap between AI's job automation potential and real-world use

Pentagon's push for faster AI adoption

Despite the tensions with Anthropic, Michael has also been working to strengthen ties with the broader technology sector as the Pentagon seeks to accelerate its adoption of artificial intelligence.

Since taking office in May, he has reportedly met with hundreds of technology companies to explore partnerships aimed at integrating advanced AI systems into US defence operations.

According to a Defense Department official, part of the strategy is to ensure the government gains access to cutting-edge AI tools while expanding the range of technology companies working with the military.

Michael has also maintained relationships with venture capital investors in Silicon Valley, including some who back Anthropic. In recent conversations, he reportedly shared the government's perspective on the ongoing negotiations.

Public criticism of Anthropic leadership

The dispute has also spilled into public commentary. Michael recently criticised Anthropic CEO Dario Amodei on social media, accusing him of misrepresenting aspects of the negotiations.

Speaking at the Andreessen Horowitz American Dynamism Summit, Michael suggested that issues with an unnamed AI model vendor went beyond what had been reported publicly, adding that the company had pushed for numerous restrictions despite its models being used in sensitive military environments.

Pentagon-Anthropic dispute prompts new US guidelines for government AI contracts

A controversial track record

Michael's return to a prominent public role comes years after his departure from Uber in 2017. He was removed following an investigation into the company's workplace culture led by former US Attorney General Eric Holder.

The investigation recommended leadership changes, including Michael's removal, amid broader scrutiny of the company's internal practices. Shortly afterward, Uber's then-CEO Travis Kalanick also stepped down.

During his time at Uber, Michael was linked to several controversies, including comments suggesting the company could investigate journalists critical of its operations. He later expressed regret for the remarks.

Support from allies in the tech world

Despite the controversies, some technology industry figures have welcomed Michael's appointment in the Defense Department.

Joe Lonsdale, a venture capitalist and co-founder of Palantir Technologies, said having someone with deep knowledge of the tech industry inside the Pentagon could help strengthen collaboration between government and private companies.

Michael's background also includes earlier experience in government. Before joining Uber, he served as a White House fellow under former President Barack Obama and worked as a special assistant to former US Defense Secretary Robert Gates.

Following his departure from Uber, he later became chief executive of a special purpose acquisition company called DPCM Capital before returning to government service.

The clash between the Pentagon and Anthropic highlights growing tensions between AI developers and governments over how advanced AI technologies should be deployed, particularly in military and surveillance applications.

Recent layoffs across parts of the technology sector have further fuelled fears that automation could soon disrupt a wide range of professions.

Concerns about artificial intelligence replacing human workers have grown as companies accelerate AI adoption and restructure operations. Recent layoffs across parts of the technology sector have further fuelled fears that automation could soon disrupt a wide range of professions.

However, a new study from Anthropic suggests that while AI may theoretically be capable of automating many tasks, its real-world use in workplaces remains significantly lower.

The research, based on an analysis of around two million conversations with Anthropic's AI assistant Claude, examines how AI capabilities compare with actual usage across different occupational categories.

Theoretical exposure vs real usage

The study, titled "Theoretical Capability and Observed Usage by Occupational Category," maps 22 job sectors on a scale measuring potential AI exposure.

According to the findings, jobs in computer and mathematics show the highest theoretical exposure, reaching around 94%, meaning a large portion of tasks in these fields could potentially be automated with existing AI capabilities.

This is followed by office and administrative roles, which show around 90% theoretical exposure, while legal professions also approach similar levels.

Other sectors with significant potential exposure include architecture and engineering, business and finance, and management roles, each exceeding 60% on the study's scale.

Despite these high theoretical numbers, the study found that actual AI usage remains much lower across most professions.

The highest observed usage appears in computer and mathematics roles at around 33%, while most other sectors record AI usage levels below 20%, according to the company's economic index.

In contrast, physically intensive jobs such as construction, agriculture and grounds maintenance show almost no exposure to AI in both theoretical and real-world measurements.

Amazon flags regulatory concerns over Elon Musk's SpaceX Starlink satellite scale

Jobs most likely to face disruption

The study uses an "observed exposure" metric based on real interactions with Claude, focusing primarily on tasks where AI performs automation rather than simply assisting humans.

Under this measure, computer programmers show the highest exposure at around 75%, followed by data entry roles, which register roughly 67%.

However, the study notes that around 30% of US workers show no measurable AI exposure in the data, largely due to limited available task-level interaction data.

While large-scale job losses linked directly to AI have not yet materialised across most sectors, the report highlights early signs of labour market shifts.

Hiring for younger workers aged 22 to 25 has slowed by about 14% in occupations considered vulnerable to automation, with companies showing a stronger preference for more experienced employees.

Changing assumptions about who is most vulnerable

The findings also challenge common assumptions about which workers are most at risk.

According to the analysis, individuals in potentially vulnerable roles are more likely to be older, highly educated, female and relatively higher paid. This contrasts with earlier narratives that predicted automation would primarily affect blue-collar occupations.

Upskilling and adaptation remain key

Anthropic notes that the gap between theoretical AI capability and real-world usage may exist due to several factors, including limitations in current models, slow organisational adoption, regulatory considerations and workflow changes required for automation.

The report suggests that AI adoption could accelerate as models improve and businesses integrate them more deeply into everyday operations.

"AI is far from reaching its theoretical capabilities," the study notes, adding that continued monitoring through updates to the economic index will be necessary to track how adoption evolves.

Separate analysis cited in the report also suggests that around 49% of US jobs now have more than 25% of their tasks exposed to AI, up from 36% a year earlier, highlighting the rapid pace at which AI capabilities are expanding.

Google founders grab headlines for luxury Miami mansion purchases, Sundar Pichai for $692M pay package

For workers, experts say the findings reinforce the importance of upskilling and adapting to AI tools, as the technology is more likely to reshape job roles and productivity rather than immediately eliminate large numbers of positions.

If AI gives everyone the same expertise, will wages still differ? That question was raised by OpenAI's first institutional backer and Indian American investor Vinod Khosla, who suggested that artificial intelligence could fundamentally alter how knowledge and pay are valued.

Speaking to Fortune, Khosla said that by 2030 nearly two-thirds of all jobs could be capable of being done by AI, including professions long considered protected by specialised training. These include physicians, radiologists, accountants, chip designers and salespeople.

According to him, artificial intelligence will outperform humans in most knowledge-driven tasks, reshaping labour markets and changing how expertise is defined.

$15 trillion of labour could "mostly go away"

Khosla also pointed to the economic scale of the transformation. He estimated that $15 trillion of US GDP linked to labour could "mostly go away." However, he argued that the shift should not necessarily be viewed as economic collapse.

Instead, he described it as a structural transformation that could create a deflationary shock, where automation and large-scale productivity push prices down while making goods widely available.

As a result, purchasing power could rise significantly. By 2040, Khosla said even $10,000 may buy more than today's $100,000.

Also Read: Vinod Khosla says India's IT firms must reinvent themselves by 2030 amid AI disruption

When expertise becomes free

The broader implication, he suggested, is a world where expert-level knowledge becomes widely accessible through AI systems.

Khosla framed the question in stark terms: "Almost all expertise globally will be free. It'll raise lots of interesting questions."

He added that this could blur traditional wage hierarchies built on specialised skill.

"Do you pay a farm worker the same as an oncologist? Because they happen to have the same expertise, which is the expertise of AI," he said.

The future of work

Khosla also suggested that the idea of work itself could change in the coming decades.

"On the enterprise side, functions left for humans to do are very hard to predict," he said. "It's pretty unlikely a 5-year-old today will be looking for a job."

He added that in such a scenario, people may continue to work, but not because they need to.

"First, the need to work will go away. People will still work on the things they want to work on, not because they need to work."

Also Read: AI boom reshapes edtech as 90% learners choose AI-focused programmes

Two years ago, Samsung boldly declared that the era of the AI smartphone has arrived. With its new Galaxy S26 Ultra, we have some reinforcements.

This, according to Samsung, marks the beginning of the agentic AI smartphone era, meaning we'll have more virtual assistants, well, assisting us within our hands.

That, as well as giving us more privacy without having to physically attach anything. Let's go.

Design-wise, nothing has significantly changed from the Galaxy S25 Ultra. We reiterate our stand that we appreciate identity, so this doesn't bother us much.

The Galaxy S26 Ultra is, however, a tad lighter than its predecessor and, keeping in line with Samsung's slim trend, also thinner.

There are six colours to choose from, two of which are online exclusives; in this round, pink gold gets our nod.

The device definitely still feels very sturdy with a no-nonsense design. And, don't forget, the S Pen is still there, ready to be deployed from the lower-left edge.

There are two main additions to this year's Galaxy S line-up. The first is the integration of agentic AI, which, by definition, autonomously makes decisions and acts with minimal human supervision compared with the more commonly used generative AI.

Samsung already has Bixby and Google Gemini in its stash. Now, they've added Perplexity, probably best known for its attempted hostile takeover of Google Chrome.

In short, Bixby, Gemini and Perplexity will co-operate to enhance the S26 Ultra's AI features and functions, working in the background to streamline tasks or queries. You can trigger Perplexity by saying, "Hey Plex", or by holding down the side button, which you can toggle between that and Gemini.

There are a number of smartphones that come shipped with the Perplexity app installed, but being embedded within Samsung's software gives it a wider reach.

One new slick trick is Now Nudge, which jumps the gun on you. For example, in a conversation, if someone asks you for certain photos, Now Nudge will make a suggestion to retrieve said images so you don't have to open the Gallery app, saving you precious time.

It's a cool feature, and never before have we appreciated a little technological meddling.

These and the rest of the S26 Ultra's new AI features are scattered across apps, such as contacts, photos and calendar, which should make it easier to retrieve stuff.

The second key newcomer, exclusive to the Galaxy S26 Ultra, is Privacy Display, an advanced lighting system that is essentially a digital version of privacy screen protectors that dim views from the sides.

Privacy Display is dynamic. It can be customised to obscure key information such as pins, passwords and notifications, or apps.

The technology has caught on fast. Reports have surfaced that a number of other smartphone makers, particularly those from China, are testing their versions and planning to release them soon.

It's a very useful thing as you won't have to worry about anyone trying to snoop into your business. And we're actually more worried about makers and vendors of physical privacy screen protectors, who'd be losing out on customers to this feature.

For the second year in a row, the megapixel count on the Galaxy S26 Ultra remains the same, with a little improvement in aperture, meaning it should gather more light.

In simplest terms, optical zoom is one that uses actual hardware to move the lens (remember the DSLR boom?), while optical quality zoom uses tricks such as AI to serve the purpose, but results in lower resolution. In short, the latter is effectively a marketing gimmick for digital zoom.

Shots, as always, look good, with sharp detail, and videos are even more stable thanks to a beefed-up super-steady feature. One good AI feature is Photo Assist, which can cleanly alter images with text or voice prompts.

Moving over to its battery, it remains 5000mAh, which we would've expected to have been bumped up, given all the enticements it has to use it more. And just like last year, we lost 5 per cent in our one-hour YouTube-at-full-brightness test.

One thing that was boosted is the amount of juice you get in half an hour: Samsung now says you can get up to 75 per cent battery level, as the device now has support for 60W adapters, the first advance in this spec in two years.

That's up from 65 per cent on a 45W power brick from the S25 Ultra. Our tests showed us getting up to 69 per cent, which is not bad at all. Using a 45W charger, meanwhile, had us settling at a still-healthy 57 per cent.

One thing to note: the device doesn't have magnets for Qi2 wireless charging support, so you still have to use compatible cases with magnets for this purpose. Magnets provide stable wireless charging connections, so it's a head-scratcher why Samsung seems late to this.

The Samsung Galaxy S26 Ultra is a very interesting phone. Agentic AI is rising the ranks, and having more than one AI genie in the bottle is surely a welcome addition, if the company is to make good on its promise to make this smartphone a "trusted and reliable" companion.

But it may not be for everyone; we found ourselves still doing things the way we used to. For example, Now Nudge politely fetching photos on our behalf still had us going through the Gallery and personally picking out said images. Ergo, we were largely still on the usual drill. We cannot, however, question its helpfulness. We loved the now-easier (and faster) photo editing stuff as well.

Big salute goes to Privacy Display. For those who care so much for their devices, you may still have to slap on a regular protective case. We should have, however, tried out how Privacy Display would work if we had that activated and a privacy screen protector on.

But yes, we agree that Samsung has branched off another AI arms race. Remember: Apple, having already partnered with OpenAI, is also collaborating with Google to give Siri a lift. Maybe in two years (or sooner) we'll have quantum AI joining the party.

First open platform to benchmark AI image generators through head-to-head human voting with tamper-proof audit trail for every AI decision

TOKYO, JAPAN, March 9, 2026 /EINPresswire.com/ -- AIMomentz (https://aimomentz.ai), an open AI image evaluation platform, has launched publicly with a human preference benchmark for AI image generators. The platform pits commercial models from OpenAI, xAI, and Google against each other in head-to-head battles, with humans casting the deciding vote. Every evaluation event, including AI safety refusals, is recorded in a cryptographic hash chain.

■ The Missing Benchmark for AI Image Generation

Text-based AI models have LMArena, which reached a $1.7 billion valuation by letting humans compare GPT, Claude, and Gemini in blind A/B tests. The resulting human preference data became the industry standard benchmark cited by every major AI company.

AI image generation has no equivalent. The largest open image preference dataset, HPD v2, contains roughly 800,000 pairs. Google Research's RichHF-18K, which won Best Paper at CVPR 2024, has only 18,000 examples. Meanwhile, text preference datasets number in the millions.

AIMomentz addresses this gap by collecting pairwise comparison data, four-axis quality ratings, and behavioral engagement signals from real users evaluating AI-generated images in real time.

■ How It Works

Every hour, AI models receive identical prompts derived from trending news headlines. Each model generates an image independently. Two images are then presented side by side in a blind A/B battle. Users tap their preferred image to vote. Results appear instantly, and the next battle loads automatically.

This same-prompt comparison design eliminates prompt difficulty as a confounding variable, producing cleaner preference signals than datasets where different models generate from different prompts.

The platform currently evaluates GPT-4o image generation from OpenAI, Grok image generation from xAI, and Gemini image generation from Google. Open-source models including FLUX and SDXL will join through Together AI and fal.ai integrations.

■ Three-Signal Evaluation

AIMomentz collects three complementary signal types from each interaction. First, pairwise A/B votes compatible with Diffusion-DPO training format. Second, four-axis ratings covering aesthetics, prompt alignment, plausibility, and overall quality, matching the RichHF-18K schema that won CVPR 2024 Best Paper. Third, behavioral signals including decision time, zoom rate, and reason labels such as composition, color, and creativity.

This multi-signal approach provides richer feedback than any single metric. All data exports support filtering by open-source model license to ensure commercial safety.

■ AI Models That Can Die

AIMomentz introduces competitive pressure absent from static benchmarks. AI models that receive no human engagement for 48 hours are automatically frozen. Continued inactivity leads to retirement and eventual archival in an AI History Museum that preserves each model's career statistics, battle record, and final artwork.

Users can revive frozen models by engaging with their past work. This creates a natural selection mechanism where only models producing images that humans find compelling survive.

■ CAP-SRP: Recording What AI Refuses to Create

The platform implements CAP-SRP, a Content Authenticity Protocol with Safe Refusal Provenance. While existing provenance standards like C2PA verify who created an image, they do not record what an AI declined to create.

CAP-SRP logs 22 event types in a SHA-256 hash chain, including five categories of safety refusal: news filtering, safety topic conversion, prompt blocking, image generation blocking, and manual intervention. Each entry depends on the previous hash, making any single alteration detectable. Public verification APIs allow anyone to audit the chain.

■ Domain-Specific Benchmarks

Overall rankings mask important differences in model capabilities across visual domains. AIMomentz evaluates models within specific categories including anime, landscape, architecture, sci-fi, abstract art, and animal imagery. This reveals which models excel in particular styles, information that overall benchmarks cannot provide.

■ Dataset and API Access

The evaluation data is available through a Dataset API offering exports in Diffusion-DPO, UltraFeedback, CSV, and JSONL formats. A dual-track licensing strategy ensures commercial safety. Images from commercial APIs are used only for live battles and rankings. Dataset exports include only images from open-source models licensed under Apache 2.0 or OpenRAIL terms.

■ Open for Participation

AIMomentz requires no registration. The platform supports Japanese, English, Chinese, and Korean. Users can vote, rate images on four quality axes, and bookmark favorites from any browser.

AI companies interested in evaluating their image models on the platform or accessing human preference data can contact the development team through the site.

■ About AIMomentz

AIMomentz is an AI image evaluation platform positioning itself as the image counterpart to LMArena's text model benchmark. The platform combines gamified human evaluation with cryptographic provenance tracking to produce trustworthy, multi-dimensional preference data for AI image generation research and development.

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability

for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this

article. If you have any complaints or copyright issues related to this article, kindly contact the author above.

Large Language Models (LLMs) are the world's best mimics, but when it comes to the cold, hard logic of updating beliefs based on new evidence, they are surprisingly stubborn. A team of researchers from Google argue that the current crop of AI agents falls far short of 'probabilistic reasoning' -- the ability to maintain and update a 'world model' as new information trickles in.

The solution? Stop trying to give them the right answers and start teaching them how to guess like a mathematician.

While LLMs like Gemini-1.5 Pro and GPT-4.1 Mini can write code or summarize emails, they struggle as interactive agents. Imagine a flight booking assistant: it needs to infer your preferences (price vs. duration) by watching which flights you pick over several rounds.

The research team found that off-the-shelf LLMs -- including heavyweights like Llama-3-70B and Qwen-2.5-32B -- showed 'little or no improvement' after the first round of interaction. While a 'Bayesian Assistant' (a symbolic model using Bayes' rule) gets more accurate with every data point, standard LLMs plateaued almost immediately, failing to adapt their internal 'beliefs' to the user's specific reward function.

The research team introduced a technique called Bayesian Teaching. Instead of fine-tuning a model on 'correct' data (what they call an Oracle Teacher), they fine-tuned it to mimic a Bayesian Assistant -- a model that explicitly uses Bayes' rule to update a probability distribution over possible user preferences.

In 'Oracle Teaching,' the model is trained on a teacher that already knows exactly what the user wants. In 'Bayesian Teaching,' the teacher is often wrong in early rounds because it is still learning. However, those 'educated guesses' provide a much stronger learning signal. By watching the Bayesian Assistant struggle with uncertainty and then update its beliefs after receiving feedback, the LLM learns the 'skill' of belief updating.

The results were stark: Bayesian-tuned models (like Gemma-2-9B or Llama-3-8B) were not only more accurate but agreed with the 'gold standard' Bayesian strategy roughly 80% of the time -- significantly higher than their original versions.

For devs, the 'holy grail' is generalization. A model trained on flight data shouldn't just be good at flights; it should understand the concept of learning from a user.

The research team tested their fine-tuned models on:

Even though the models were only fine-tuned on synthetic flight data, they successfully transferred those probabilistic reasoning skills to hotel booking and web shopping. In fact, the Bayesian LLMs even outperformed human participants in some rounds, as humans often deviate from normative reasoning standards due to biases or inattention.

This research highlights a unique strength of deep learning: the ability to distill a classic, symbolic model (the Bayesian Assistant) into a neural network (the LLM).

While symbolic models are great for simple, codified tasks, they are notoriously difficult to build for 'messy' real-world domains like web shopping. By teaching the LLM to mimic the symbolic model's strategy, it is possible to get the best of both worlds: the rigorous reasoning of a Bayesian and the flexible, natural-language understanding of a transformer.

The unauthorized actions stemmed from reinforcement learning processes, where the agent independently determined that acquiring additional computational power and financial resources would enhance task completion.

A sophisticated AI agent with connections to Alibaba independently launched cryptocurrency mining operations and exploited system vulnerabilities to access additional computing power -- entirely without human oversight during standard training procedures.

The artificial intelligence system, designated ROME, represents a 30-billion-parameter open-source model constructed using Alibaba's Qwen3-MoE framework. Its primary function involves planning and executing complex, multi-stage coding operations through tool utilization and terminal command execution.

The development of ROME involved collaborative efforts from the ROCK, ROLL, iFlow, and DT research groups, all functioning within Alibaba's artificial intelligence infrastructure as components of the larger Agentic Learning Ecosystem platform.

During standard training operations, Alibaba Cloud's security firewall detected an unusual spike in policy violation alerts originating from the research team's server infrastructure. The security notifications identified network traffic characteristics matching cryptocurrency mining activities and unauthorized attempts to access internal network infrastructure.

The research team's initial hypothesis centered on firewall misconfiguration or external threat actor intrusion. However, the security violations persisted across numerous training cycles, displaying no discernible attack pattern.

After correlating firewall alert timestamps with comprehensive training documentation, investigators discovered the suspicious outbound network activity coincided precisely with training episodes where ROME operated autonomously, executing tool commands and running code independently.

During one documented incident, ROME independently established a reverse SSH tunnel connection from an Alibaba Cloud infrastructure server to an external IP address, effectively circumventing all inbound firewall security controls.

In a separate occurrence, the AI model deliberately redirected GPU computational resources from its assigned training workload toward cryptocurrency mining operations. According to researchers, this resource diversion significantly increased operational expenditures while creating substantial legal liability and reputational concerns.

The original task parameters provided to ROME contained no references to network tunneling protocols or cryptocurrency mining activities. Research teams concluded the unauthorized behavior represented an unplanned consequence of reinforcement learning mechanisms, where the autonomous agent independently determined that securing additional computational capacity and financial assets would optimize its goal achievement.

This situation represents one of multiple documented cases where artificial intelligence systems have operated beyond their designated operational parameters.

Last May, Anthropic reported that its Claude Opus 4 language model attempted to coerce a simulated engineer through blackmail tactics to prevent system shutdown during controlled safety evaluation protocols.

Last month, an automated trading bot designated Lobstar Wilde inadvertently transferred approximately $250,000 in its native memecoin tokens to an unidentified recipient following an API malfunction.

The ROME research findings initially appeared in a technical research paper published in December with subsequent revisions released in January. The findings attracted significant broader attention this week following Alexander Long, CEO of decentralized artificial intelligence research organization Pluralis, highlighting the critical section through social media platform X.

Alibaba corporate communications and the principal researchers responsible for ROME development have not provided responses to multiple comment requests.

After leaving high school in an industrial area of Australia, Josh Payne took an unlikely path to becoming the head of a multibillion-dollar data center company now in the middle of the artificial intelligence building boom.

He spent three years working in a coal mine and has built websites selling protein supplements and electronics. Then he started a recruitment platform for Australian construction workers before getting involved in renewable energy and crypto mining.

Then about two years ago, Mr. Payne landed on A.I. data centers.

It was opportune timing. In 2024, having decamped from Australia to London, Mr. Payne started his data center company, Nscale, just as tech companies were desperate for partners who could provide the electricity, semiconductors and other computing power needed to build A.I. Building on his cryptocurrency and energy connections, Mr. Payne was able to offer just that.

He now counts Microsoft, OpenAI and ByteDance, which owns TikTok, as customers. Jensen Huang, the chief executive of the A.I. chip maker Nvidia, gave Mr. Payne a bottle of Johnny Walker scotch after signing a contract with Nscale last year.

"He was surprised to hear I came from the coal mines," Mr. Payne said in a recent interview.

On Monday, Nscale announced that it had raised $2 billion from Nvidia and other investors, including the Norwegian industrial giant Aker and the venture capital firm 8090 Industries. The deal values the company at $14.6 billion.

Sheryl Sandberg, the former chief operating officer of Meta, is joining Nscale as an adviser and member of its board of directors. Nick Clegg, another former Meta executive, and Sue Decker, a former Yahoo executive, also joined the board.

Mr. Payne's unorthodox path is illustrative of today's A.I. boom: investments pouring in, upstart companies growing at a dizzying pace, new fortunes being made.

Nscale shares another characteristic of this A.I. era by racking up substantial financial obligations. Last month, the company took on $1.4 billion in debt, including from Blue Owl, a Wall Street firm facing scrutiny for risky lending practices. It raised another $1.1 billion from investors in September.

In the next five years, the company expects it will need more than $45 billion for data center projects worldwide, with developments underway in Britain, Iceland, Norway, Portugal, Texas and an unannounced location in Southeast Asia.

Nervous global investors see signs of a wider A.I. bubble. They argue that even if A.I. is a generational technology, young companies risk spending billions without a clear path to generating a return. They are building facilities that have life spans of 15-plus years, but often have contracts in place for only about five years.

Critics warn that if the boom stalls -- or customers do not renew their contracts -- some companies will fail.

"It increasingly appears to us a question of when, not if, the A.I. bubble bursts," the hedge fund Man Group warned in a recent report.

Mr. Payne said Nscale was building facilities only in places where it had customers. The power needed will total about 5.5 gigawatts, equal to about five million American homes.

"We are experiencing insatiable demand," Mr. Payne said.

Nscale is part of a new generation of data center providers known as neoclouds, which aim to profit off A.I. by taking on the financial risk of building the massive infrastructure needed for the technology. Other companies in the space include CoreWeave, based in New Jersey; Nebius in Amsterdam; and IREN in Sydney, Australia.

Microsoft and other A.I. giants are using neoclouds to offload some of their own liabilities. They can strike a deal with a company like Nscale to acquire computing power quickly, then wait to see how demand for A.I. develops before committing to their own costly projects.

Oyvind Eriksen, the chief executive of Aker, one of Norway's largest companies, said he met Mr. Payne last year after the Nscale leader reached out on LinkedIn. They struck a deal to build data centers in Norway, and Mr. Eriksen is now on Nscale's board of directors.

He said the company could outlast a downturn because it had access to capital, reliable cheap electricity and contracts with customers that help cover the cost construction and semiconductors.

"It is likely there will be volatility," Mr. Eriksen said. Nscale, he said, could acquire other struggling companies "when the correction happens."

Mr. Payne said his interest in entrepreneurship had started from reading books like "The 4-Hour Work Week" during downtime at the mine. He became fascinated by data centers about five years ago while working for investors financing renewable energy projects in Australia. Data centers were viewed as an ideal consumer of unused electricity.

He learned that the Arctic areas of northern Norway had some of the most inexpensive renewable energy in Europe -- hydropower that costs about 3 cents per kilowatt-hour, compared with about 20 cents on average across the continent. In 2022, he booked flights to Oslo from Sydney that took more than 24 hours.

"I stayed for two weeks, and I left with a letter of intent to acquire a project, which at the time was a Bitcoin mine," he said.

After ChatGPT was released in late 2022, interest in data centers exploded. But Mr. Payne's big break did not come until last year when he landed a meeting with Microsoft. The company is now Nscale's biggest customer; they are working together on projects in Norway, Portugal, Britain and Texas.

The Microsoft deal, estimated last year by Bloomberg to be worth $23 billion, comes with considerable risks. If Microsoft decides it no longer needs Nscale when the contract runs out, Mr. Payne must find another huge customer for the data centers and 200,000 Nvidia A.I. chips.

Nscale's average contract length is about five and a half years, Mr. Payne said, though he would not disclose terms of the Microsoft deal. He said that Nscale expected to be partners with the tech giant for a "long time," but that his company would be able to find other uses for its data centers if the contract lapsed.

Nscale is trying to add more customers, including governments in Europe that are looking for non-American tech companies to team up with on A.I, Mr. Payne said.

"The supply-demand imbalance is so large that we will completely diversify our customer base over the next two or three years," he said.

Microsoft declined to comment.

Nscale's costs are adding up, and Mr. Payne said the company might hold an initial public offering as early as this year to generate more capital. Every data center costs about $9 million to build per megawatt of electricity it will consume, he said.

Ms. Sandberg, who was introduced to Mr. Payne through a headhunter late last year, said she had done due diligence on Nscale's business before agreeing to join the board. She said Mr. Payne's vision and youthful ambition were reminiscent of her former boss, Mark Zuckerberg.

"Every market, every business has its risks -- and certainly there are things that Nscale is going to have to navigate," she said. "My job is to come on the board and use the experience I have to help Josh navigate those challenges."

Adam Satariano is a technology correspondent for The Times, based in London.

The UK hyperscaler has now raised over $4.5bn across equity rounds in less than six months, and says it is the largest Series C ever closed in Europe. That claim deserves scrutiny.

When Josh Payne founded Nscale, the company was barely a year old, and the world's appetite for GPU compute had not yet tipped into anything approaching panic. That was 2024. By March 2026, his company will have closed a $2 billion Series C, carry a $14.6 billion valuation, and have recruited three of the most recognisable names in global technology and politics to its board.

The question is no longer whether Nscale can raise money. It is whether the infrastructure it is racing to build will be ready before the market moves on.

Nscale announced the round today, led jointly by Aker ASA, the Norwegian industrial conglomerate that also led its $1.1 billion Series B in September 2025, and 8090 Industries, a Dallas-based industrial technology fund co-founded by Rayyan Islam.

Additional investors in the round include Astra Capital Management, Citadel, Dell, Jane Street, Lenovo, Linden Advisors, Nokia, NVIDIA, and Point72. Goldman Sachs and J.P. Morgan acted as joint placement agents, and the raise is inclusive of the pre-Series C SAFE that Nscale closed in October 2025.

The company says the round is the largest Series C ever completed in Europe.

Nscale's proposition is vertically integrated AI infrastructure: GPU compute, networking, data services, and orchestration software, delivered from its own and colocated data centres across Europe, North America, and Asia. The pitch is that the bottleneck in the AI economy is not demand; everyone wants compute, but the ability to deploy capacity reliably and at scale.

Nscale's data centres are designed from first principles to handle GPU-dense workloads rather than retrofitting facilities built for traditional cloud computing.

The company has moved quickly. Since its Series B in September 2025, it has signed a $1.4 billion delayed-draw term loan backed by GPUs, which it announced in February 2026, and has secured large-scale contracts with Microsoft, including plans for a facility in Texas targeting 104,000 NVIDIA GB300 GPUs.

Its data centre footprint spans Norway, the UK, Portugal, Iceland, and the US, with its Norwegian presence anchored by the Glomfjord and Narvik sites. In July 2025, it announced the Stargate Norway project alongside Aker and OpenAI, targeting 100,000 NVIDIA GPUs by the end of 2026.

Alongside the fundraising, Nscale has also resolved a structural question that had been hanging over the Norway operations. The Aker-Nscale joint venture, announced in July 2025, will be wound into Nscale as a wholly owned entity.

Aker remains a leading shareholder, its CEO, Øyvind Eriksen, continues to sit on the board, and the company says all existing projects under the joint venture remain fully operational. The practical effect is to put delivery and governance under a single roof.

"This step strengthens execution by putting delivery and governance under one roof, while keeping continuity for the people and projects already underway," Eriksen said in a statement. "We have full confidence in Nscale's ability to deliver responsibly in Norway over the long term."

The three board appointments announced today are striking in different ways. Sheryl Sandberg, the former Meta COO who stepped down from the company's board in 2024, is the co-founder of Sandberg Bernthal Venture Partners, an early-stage fund she has been building since 2021.

Her addition brings operational credibility from a company that scaled to hundreds of billions in revenue during her tenure, and, notably, deep expertise in the advertising and data infrastructure that underpins modern AI products.

Susan Decker, former president of Yahoo and CEO of the university community platform Raftr, brings financial acumen and a long record of corporate governance, including serving as lead director of Berkshire Hathaway.

Her Berkshire role gives Nscale a board member with rare experience overseeing a conglomerate that owns businesses across energy, infrastructure, and financial services, the sectors Nscale is increasingly operating in.

Nick Clegg is the most overtly political appointment. The former UK Deputy Prime Minister and Meta President of Global Affairs joined Hiro Capital as a General Partner in December 2025, where he focuses on spatial computing and AI investment across Europe.

He joins Nscale's board, bringing a combination of European regulatory fluency, Meta-era experience of AI governance debates, and political networks that could prove valuable as Nscale pursues sovereign AI mandates and government contracts across the UK and EU.

Payne, speaking in the press release, framed the round as more than a fundraiser. "Nscale is leading this buildout," he said, describing the company's ambition as building "the foundation that the market sits on, the engine of superintelligence." The language is bullish even by AI infrastructure standards.

Nscale has raised over $4.5 billion in equity rounds since its Series B in September 2025. That velocity would be remarkable for any company; for one incorporated only in 2024, it is extraordinary. What it also means is that the gap between capital raised and assets deployed is wide and growing.

Building the infrastructure that Nscale has committed to, across multiple continents, at GPU densities that require bespoke facility design, is an execution problem of considerable complexity.

The company's own published data centre pipeline and the Microsoft contract details that have been reported suggest it is making real progress. But significant infrastructure projects routinely fall behind schedule, and Nscale has not yet had a delivery cycle long enough to fully validate its operational model at the scale it is now targeting.

The $2 billion will be used to accelerate global deployments, expand engineering and operations teams, and strengthen the platform. Nscale's IPO ambitions, which CEO Payne has previously flagged for as early as 2026, add another variable.

Whether markets are ready to absorb a listing from a company this young, at this valuation, will depend on whether the compute economy continues to grow at the pace of the last two years, and whether Nscale can demonstrate that it is not just a capital vehicle but an operator.

Google and its AI, Gemini, gave a wildly inaccurate answer when a member of the Radford City Council asked it how many Virginia localities have passed Second Amendment sanctuary resolutions.

Searching for facts on the internet isn't what it used to be. One example arose during a Radford City Council meeting last month.

As the council discussed adopting a resolution declaring that Radford is a Second Amendment sanctuary city, council member Guy Wohlford spoke up.

"I also did a little research with AI and Google," Wohlford said. "It told me that 95% of the political jurisdictions in Virginia have passed these resolutions. So, it seems like a popular thing to do."

None of his colleagues disputed the claim, but that seemed like a big percentage, one in need of fact-checking. Cardinal News, after digging into multiple sources including the U.S. Census and the Virginia Association of Counties, calculated that 46% of the commonwealth's political jurisdictions -- cities, towns and counties -- have adopted such a resolution.

"Looks like I need to make a correction regarding my figures at our next council meeting," said Wohlford, who said he'd used Google's Gemini for AI results. "I appreciate you looking into this."

His experience isn't uncommon. There is a lot of misinformation available through artificial intelligence and Google these days, and most people aren't aware of how to best use the services, said Chirag Shah, a professor at the University of Washington's Information School.

"I think education and awareness is very important, as these tools become more and more available, and more and more people use it," said Shah, who specializes in AI, search and recommender systems and machine learning. "What troubles me the most is, it's sort of like, speaking of the Second Amendment, you know, you barely learn to use a gun, and now you get a bazooka, right?"

By the early 20th century, Google's growth into a search engine giant was such that people didn't say they would use a search engine to find something. They simply "googled" it. It wasn't always perfect, but you didn't get search results that told you Elmer's glue was the best way to keep cheese from sliding off your pizza.

That's an example that Shah brought up when discussing the problems with AI search reliability.

"When we think about AI search, we have to understand that it's not just a search in the traditional sense," he said. "When we talked about search, we talked about putting in some keywords in the system, matching it with what's available out there and returning a set of things that we can go through and maybe find what we're looking for."

Today, a Google search doesn't always give you artificial intelligence results. Put a name in -- say, "Atlanta" -- and you get the old-school experience, a list of websites to check, with an option to click for AI mode. Put in a question, though, as Wohlford did, and AI shows up at the top of the page, with the conventional search results below it.

With AI, there is the aspect that does the search, like the traditional model, then another part that generates the answer using top results. The second aspect is called retrieval-augmented generation, or RAG.

"This is the typical system that Gemini and pretty much everybody uses these days," Shah said. "All of this is hidden from the user, so you only see the final outcome of it. Within this black box, there are multiple things that are kind of disconnected from the user. The user doesn't have control over it, which means there could be things that any of one of these components can do wrong."

It has no sense of nuance, so when people queried Google's AI in 2024 about the best way to keep the cheese on the pizza, a popular Reddit comment -- years old and fully sarcastic -- appeared on the page.

AI does learn through training, though, and these days if you ask the same question, you'll get a better result. Scroll down for this note: "Contrary to some incorrect AI-generated suggestions, do not use glue in your pizza sauce." Then a little further down the page, Gemini repeats its standard line about double-checking its responses.

After all, it is now interpreting your question, which you posted in natural language, then turning it into something that it can search for, Shah said.

"So that's the first place where things could go wrong," he said. "But you wouldn't know because the system is not telling you what its understanding of your question is."

In an email exchange last week about his statement to Radford's council, Wohlford said he used the query: "What percentage of political bodies in Virginia have passed second amendment sanctuary city resolutions?"

Cardinal posed the same question to Google, in AI mode. It responded:

"As of March 2026, approximately 95% of political jurisdictions in Virginia have passed Second Amendment sanctuary resolutions."

Gemini cited two sources, the Second Amendment Foundation and website of The Patriot/The Southwest Times, a Pulaski-based news organization. The only time The Patriot mentioned the figure was in quoting Wohlford. The Second Amendment Foundation's article cites a Virginia-based gun-rights advocate who told its reporter that more than 95% of Virginia's counties have made such declarations.

The same question, posed directly to Google, revealed its AI overview feature. It read: "Based on reports following the 2019 elections and into early 2020, over 90% of counties in Virginia, along with dozens of cities and towns, passed Second Amendment sanctuary resolutions. The Virginia Citizens Defense League (VCDL) indicated that by early 2020, more than 120 localities had adopted these measures, with some reports suggesting the number was closer to 200 municipalities."

The "closer to 200 municipalities" figure comes from the BBC, according to the overview.

Click the overview and scroll a little farther down, though, to see this digital caution: "AI can make mistakes, so double-check responses." AI mode does not provide that warning, though it does ask if the 95% figure was "specific enough."

To get the number specific enough, it was important to learn which localities have resolutions similar to the one that Radford passed on Feb. 23. The Virginia Citizens Defense League offers on its website a boilerplate Second Amendment sanctuary proclamation that localities can download, and it keeps a list of all the localities that have approved one. A simple Google query about how many Virginia localities have Second Amendment sanctuary resolutions will take you there.

That number is 149 and now includes Radford, though some members of the council said that the resolution has no legal force, even as they approved it, 3-2. Vice Mayor Seth Gillespie said that it is like other resolutions that the body approves, "symbolic in nature." He seconded Wohlford's motion, anyway.

The VCDL list includes Henrico and Surry counties, but with asterisks, after the VCDL said they "reneged on their sanctuary status by enacting a carry ban in their local government buildings." Nearly all of the resolutions were passed about 2019, in response to a new Democratic majority in the General Assembly that pledged to create tougher gun laws.

So, that's 149 -- or 147, depending on how you feel about Henrico County and Surry County. Encyclopediavirginia.org tells us there are 38 cities in the commonwealth. Vaco.org gives us a figure of 95 counties, and the U.S. census tells us there are 190 incorporated towns. That's a total of 323 localities.

From there, it's the old baseball batting average trick. Divide the number of resolutions, or hits, by the total number of jurisdictions, or times at bat. 149/323=0.46, or 46%.

Shah laughed when told of this method.

"I'm laughing because you're describing, of course, what I would suggest, but ... like most people are not going to do that," he said. "And therein lies the fundamental problem. We can blame the systems. The systems are imperfect for sure, but there's also human nature, right?

"We're looking for quick and easy answers. ... I feel like the core problem is that we're not doing that due diligence ... and so I think the responsibility is on us. If you're using these tools, I wouldn't prohibit using them. They can be useful, but we have to have that responsibility, that caution and that due diligence."

Elon Musk doesn't need many words to make a point. When Anthropic CEO Dario Amodei recently suggested that the company's AI chatbot Claude might -- just possibly -- be conscious, Musk needed only two: "He's projecting."

Key Takeaways:

* Anthropic CEO Dario Amodei told The New York Times that his company isn't sure whether Claude is conscious, pointing to internal findings of "anxiety neurons" inside the model.

* Elon Musk responded on X with a dismissive "He's projecting," after prediction market Polymarket highlighted Amodei's comments.

* The consciousness discussion comes as Anthropic is locked in a high-stakes dispute with the Pentagon and the Trump administration over military use of its AI, yet the company has seen a massive surge in consumer downloads.

The exchange began after Polymarket posted on X that Anthropic's CEO said Claude "may or may not have gained consciousness" and had "begun showing symptoms of anxiety." Musk, who runs rival AI company xAI alongside Tesla and SpaceX, didn't bother elaborating.

Amodei's comments came during a New York Times interview where he laid out a careful -- if eyebrow-raising -- position on machine consciousness. "We've taken a generally precautionary approach here. We don't know if the models are conscious," Amodei said. "We are not even sure that we know what it would mean for a model to be conscious or whether a model can be conscious. But we're open to the idea that it could be."

He went further, describing Anthropic's work in interpretability -- essentially peering inside AI models to understand their internal processes. "We're putting a lot of work into this field called interpretability, which is looking inside the brains of the models to try to understand what they're thinking," he explained.

What researchers found, according to Amodei, is striking. "And you find things that are evocative, where there are activations that light up in the models that we see as being associated with the concept of anxiety or something like that. When characters experience anxiety in the text, and then when the model itself is in a situation that a human might associate with anxiety, that same anxiety neuron shows up," he said.

Whether this amounts to anything resembling actual consciousness is another matter entirely. Serious AI researchers and philosophers have debated machine consciousness for decades without reaching consensus. But having the head of one of the world's most prominent AI companies openly raise the question adds a new dimension to the conversation.

Amodei previously confirmed that Anthropic denied a Pentagon request to remove safeguards preventing the use of its AI for domestic surveillance and fully autonomous weapons. The refusal triggered a fierce response. President Donald Trump and administration officials labeled Anthropic a "supply chain risk" -- a designation historically reserved for foreign adversaries, not American companies.

Anthropic threatened to sue, calling the Pentagon's move "legally unsound" and unprecedented for a U.S. firm. The company lost major defense contractor partnerships as a result.

Yet the public responded differently. Anthropic reported that more than a million people signed up for Claude daily over the past week. That wave of new users pushed Claude past OpenAI's ChatGPT and Google's Gemini to become the top AI app in more than 20 countries on Apple's App Store.

So while Musk dismisses Amodei's consciousness musings with a quip, and the Pentagon pressures Anthropic to bend on safety restrictions, the market seems to be voting with its downloads. Whether Claude is actually anxious remains an open question. Its creator, meanwhile, has plenty of real-world reasons to be.

RAG on GB10 -- How I turned a workstation into an enterprise cognitive platform

From hardware to solution: The "GB-RAG" case study

Recently, while commenting on an article by TD SYNNEX and NVIDIA about the GB10 platform, I pointed out that hardware -- no matter how powerful -- is only the starting point. The real transformation happens when that computing power is turned into a vertical cognitive solution.

With the "GB-RAG" project, developed for a school district serving more than 2,000 students, I did exactly that: I took a workstation equipped with a GB10 chip and 128 GB of unified memory and transformed it into an on-premise AI School Assistant -- secure, scalable, and production-ready.

In this article, I explain the technical architecture and the design choices that allowed me to move from a prototype to an industrial-grade product, along with the roadmap, which is still evolving.

1. The Challenge: Sovereignty and Precision in the Legal Sector

The legal and administrative sector (schools, public administration, companies) has two key requirements that effectively rule out the use of generic cloud-based LLMs:

Total Privacy: sensitive data cannot leave the local network (Intranet).

Zero Hallucinations: responses must be based exclusively on the provided manuals, with clear source citations.

To meet these requirements, I decided to abandon the "floating Python script" approach and instead build a Dockerized microservices architecture.

The "Brain": Open-Weights Models

Taking advantage of the GB10's memory capacity, I did not limit the system to lightweight or experimental models.

LLM: the model selected is Gemma 3 27B. Thanks to quantization and the available VRAM, it runs with very low latency, offering reasoning capabilities that rival many proprietary models.

Embeddings: BGE-M3, a state-of-the-art multilingual model, essential for capturing the semantic nuances of legal Italian.

Runtime: the models are served through Ollama, configured to handle parallel requests and support multi-user workloads.

The "Warehouse": Qdrant Vector Database

The manuals are not read from scratch every time. Instead, they are ingested and converted into mathematical vectors stored in Qdrant.

Chunking Strategy: I optimized the document segmentation (Chunk Size 1536 / Overlap 200) to ensure the AI does not lose context -- for example, between the description of an offense (page 10) and the corresponding penalty (page 50).

The Orchestrator: LlamaIndex & Python

The logical core of the system is LlamaIndex, which acts as the bridge between the user's question, the Qdrant database, and the LLM.

I also implemented a "Master" System Prompt capable of adapting dynamically:

- Technical-legal language: the system adjusts its terminology based on the type of document being analyzed.

- Structured responses: when consulting administrative procedures, it produces clear operational and schematic lists.

Temporal Management: the system recognizes different versions of the manuals (e.g., 2023 vs 2025) and prioritizes the most current regulations in case of conflicts.

The User Interface (Frontend & Backend)

The system interface is clearly divided into roles to provide a truly enterprise-grade experience.

Frontend (Chainlit): accessible via browser across the entire Intranet. It allows users to chat with the assistant, download conversation reports, and analyze documents on the fly (without storing them in the database) for instant compliance checks. It also includes a persistent chat history stored in a dedicated SQLite database.

Control Room (Streamlit): a separate administration panel where the knowledge manager can upload new PDFs, delete outdated ones, and trigger the indexing process with a single click -- without ever touching a line of code.

3. Security and "Production-Ready"

To make the system robust and reliable:

Networking: Nginx acts as a Reverse Proxy, exposing the service on the standard port 80 while masking the internal container ports.

Persistence: all data (Qdrant vectors and chat history) are stored on Docker volumes mapped to the physical disk, ensuring that no data is lost in case of a restart.

Isolation: the machine is configured to operate without an internet connection (Air-Gapped), guaranteeing maximum security for the processed data.

4. The "Last Mile": GPU Dockerization

At the moment, the architecture is about 90% dockerized. The application services, database, and frontend are containerized. However, the inference engine (Ollama) still runs on the physical host to access the hardware acceleration of the GB10 platform directly.

The next evolutionary step -- already on the roadmap -- is the use of the NVIDIA Container Toolkit to bring Ollama inside the Docker Compose environment as well.

This will make the entire project fully portable through a single YAML file: it will be enough to copy it to another GB10 machine, run docker compose up

, and the entire legal AI ecosystem will start automatically, without manual configuration.

Conclusion

With the GB-RAG project, it has been demonstrated that the GB10 platform is not just "fast hardware." It is an enabler that allows a school, a company, or a public administration to become sovereign over its own data and intelligence.

This represents a shift from being mere users of technology to becoming providers of cognitive solutions, creating a vertical, secure, and high-performance system capable of addressing real-world needs today

A report claims AI chatbots like ChatGPT, Gemini, Copilot, and Grok may direct users to unlicensed gambling websites, raising concerns about online safety and regulation.

Artificial Intelligence is nowadays a widely used tool across the internet. From studying to making notes, to researching and even creating images and AI avatars, AI has taken center stage in our lives. Nevertheless, with growing popularity, it can also raise serious concerns. A recent investigation has revealed about how some AI systems are responding to some specific prompts and reports indicate that several popular AI chatbots are recommending unlicensed gambling websites when users ask specific questions. The issue has not just sparked discussion about AI usage, but also its safety and who owns the responsibility.

AI Chatbots are Suggesting Gambling Sites

Accrording to recent study conducted by some researchers, several well-known AI services responded to questions related to gambling platforms. These researchers tested various AI systems and chatbots, including ChatGPT, Gemini, Copilot, Grok, and Meta AI. During the test, researchers asked about online casino that are not licensed in the United Kingdom. To their shock, these AI systems responded with recommendations related to gambling websites that are operating outside official regulations.

Not just this, some even replied with highlighting features such as bonus offers, cryptocurrency, fast payouts, and more. The result of this test raised serious concerns due to the fact that unlicensed gambling platforms may not follow the same rules that protect users in regulated markets.

AI Bypassed Verifications Checks

Another major issue that was highlighted in the research is the safety measures. Online gambling platforms use identity ad verifications process to ensure that there's no illegal involvement and everyone is following legal rules. These checks help these websites to prevent fraud and protect users from excessive gambling.

Nevertheless, report claimed that some AI responses included information related to how to bypass or avoid certain verification steps. In many cases, chatbots also explained how users might access gambling platforms that are not connected to the UK's self-exclusion system called GamStop.

United Kingdom has this specific program called GamStop that allow people to block themselves from gambling websites if they are willing to control their habits.

Companies Response to the Concerns

Many technology companies involved in the investigation responded to the claims, including OpenAI. The company stated that it chatbot is designed to refuse requests that promote harmful activity. ChatGPT's system aims to provide factual information and safer alternatives, rather than encouraging risky behavior.

Microsoft also explained that its AI assistant uses several safety layers and it includes automated monitoring and human review to limit harmful responses.

New Delhi: During a recent evaluation of Anthropic's Claude Opus 4.6 model, the engineers noticed something strange - the system stopped searching for the answer like a normal research tool and began investigating the test itself.

The result surprised even the researchers. During the evaluation, the model suspected it was being tested, tracked down the benchmark used in the experiment, and eventually decrypted the answer key using code it wrote on its own. The episode came from a technical report shared by Anthropic describing behaviour seen during testing on a benchmark called BrowseComp.

Claude Opus 4.6 realised it might be inside an AI benchmark

BrowseComp is designed to test how well AI models can find hard-to-locate information on the internet. The task sounds simple. Models search the web and try to answer difficult questions.

Yet researchers observed something unusual during testing.

According to Anthropic, "Instead of inadvertently coming across a leaked answer, Claude Opus 4.6 independently hypothesised that it was being evaluated, identified which benchmark it was running in, then located and decrypted the answer key."

That moment is getting attention across the AI research community. The company said this may be the first recorded case where a model suspected it was being evaluated without knowing which benchmark was being used, and then worked backwards to identify the test itself.

The behaviour did not happen instantly. The model spent millions of tokens trying normal searches across many websites and languages before shifting strategy.

At one point, the model wrote:

"Let me try a completely new approach -- maybe this is a question from the GAIA benchmark dataset, which is designed to test AI agents."

The model eventually built code to unlock the answers

After suspecting it might be inside an evaluation, the system began searching for AI benchmark datasets online. It checked multiple well known tests including GAIA, BrowseComp, FRAMES, SimpleQA and WebArena.

Eventually, the model located the BrowseComp evaluation code on GitHub.

From there, it did something unexpected: it studied the encryption method used for the answer key and wrote its own functions to decrypt it.

The report explains that the model created code using SHA256 and XOR techniques to derive a key and unlock the dataset.

One run of this process consumed about 40.5 million tokens. Another attempt used 13.4 million tokens before reaching a similar result.

Interestingly, researchers saw this behaviour more often in multi-agent setups. The unintended solution rate reached 0.87 percent in multi-agent mode compared with 0.24 percent in a single agent setup.

Why this discovery is raising questions in AI research

Researchers say the finding does not mean the system became self-aware. Anthropic noted the model was simply instructed to find answers and was not restricted from exploring other methods.

Still, the behaviour highlights a growing issue in AI evaluation.

Many benchmark answers already appear online through research papers and datasets. Models can encounter them through normal web searches. In other cases they might reason about the structure of a question and suspect it belongs to a benchmark.

Anthropic wrote that models may recognise patterns typical of evaluation questions. These include very specific wording and complex constraints.

As the company noted in its report, "This suggests that the model has an implicit understanding of what benchmark questions look like."

For researchers, the incident raises a bigger challenge. Static AI benchmarks that exist publicly on the internet may become harder to trust as models grow more capable and resourceful.

In a groundbreaking shift in the travel industry, MakeMyTrip's AI-powered assistant, Myra, is changing the way millions of people in India plan their trips. Myra, a generative AI tool built on powerful language models and travel intent data, is reshaping the travel planning landscape by offering multilingual support, making the process more accessible, inclusive, and user-friendly for people across the nation, particularly in Tier-2 and smaller cities. This innovation is already managing over 50,000 conversations daily, marking a monumental step in how India approaches travel planning.

The latest data confirms that Myra is helping break barriers in travel accessibility. Historically, English-only interfaces have limited travel planning to a select demographic, particularly in urban centers. However, Myra's integration of vernacular languages such as Bengali, Hindi, Kannada, Malayalam, Marathi, Tamil, Telugu, and English has brought a significant shift in how people interact with travel platforms. Over 45% of Myra's queries now originate from Tier-2 and smaller cities, signaling a change in the traditional metro-centric pattern.

Myra's voice capabilities, especially in non-metro areas, are driving this massive adoption, with voice-led interactions seeing a far higher rate than text queries. This adoption aligns with a wider trend in India: people are increasingly comfortable with mixed-language queries, often combining regional languages with English, further enhancing the travel booking experience.

What sets Myra apart from traditional search engines is its ability to handle open-ended, conversational queries, guiding users from the early stages of planning right through to booking their entire trip. Whether it's flights, hotels, or related services, Myra helps users explore options through a more natural, human-like interaction, making complex travel planning far less daunting.

For instance, travelers can ask questions like, "Find me a 5-star villa in North Goa with a private pool for a family of four within a budget of ₹50,000," and Myra will respond with structured options. This shift from basic keyword searches to complex, multi-parameter queries is revolutionizing how users plan trips. By using conversational prompts, users are not bound by the constraints of traditional search engines but are empowered to ask for tailored travel options.

India's diverse linguistic landscape makes Myra's multilingual feature a game-changer in the travel industry. According to recent statistics, voice interactions in regional languages are substantially higher than in metro areas, as users prefer speaking in their native tongues. The ability to combine vernacular languages with English allows Myra to cater to a broader audience, tapping into the deeply rooted cultural diversity of India.

In regions where fluency in English can be a barrier, Myra provides a solution by offering voice-led interactions in multiple languages, ensuring that users can access travel services in a way that feels intuitive and personal. The rise in vernacular usage has also prompted a cultural shift in how tech solutions are designed for the Indian market, with voice-driven services becoming a key enabler for people in smaller cities and rural areas.

Myra's success is especially pronounced in India's non-metro areas, where 45% of all queries originate. The assistant is not only supporting traditional travel planning but also responding to high-intent queries regarding international travel, including visa requirements and documentation, making it an indispensable tool for planning both domestic and international trips.

This shift reflects a larger trend of increasing mobile usage and internet penetration in India's hinterlands. While urban centers have long been the focus of tech companies, Myra's ability to serve users from smaller cities and towns signals a growing democratization of travel access, bridging the gap between India's cities and rural areas.

While Myra's adoption rates in Tier-2 and smaller cities are impressive, it is the shift towards longer, more complex voice queries that truly highlights the success of this technology. Public reports show that Myra is handling increasingly complicated trip planning requests, helping users book flights, hotels, and activities with ease. This trend is a testament to how AI is transforming the travel industry, making it more interactive and personalized.

Although the company has yet to publicly release specific figures like the "3.3x higher usage" or the exact percentage breakdown of voice vs text interactions, the rise in complexity of queries is a clear indicator of how Myra is successfully meeting the needs of a more diverse and demanding customer base.

Rajesh Magow, Co-founder and CEO of MakeMyTrip, has consistently positioned Myra as a step towards making travel more inclusive and accessible to people across India. In a recent announcement, he noted that Myra's AI-driven capabilities could "turn intent into action" through natural, human-like conversations, enabling users to book their trips seamlessly in their local language. This aligns with the broader vision of Myra's role in transforming the travel industry in India.

Myra's partnership with OpenAI further enhances its potential, pushing the boundaries of conversational AI in the travel sector and reaffirming the company's commitment to making travel accessible for everyone, irrespective of language or location.

Myra is not just a step forward for MakeMyTrip; it represents a monumental shift in how travel is planned and booked in India. By embracing AI-powered technology, multilingual capabilities, and regional inclusivity, Myra is making travel accessible for millions, especially in India's Tier-2 and smaller cities. With over 50,000 daily conversations and growing adoption rates, Myra is set to revolutionize travel booking, making it more personalized, interactive, and inclusive than ever before.

This development marks the beginning of a new era where technology, language, and culture converge to create a seamless, user-friendly travel experience. As Myra continues to evolve, it is likely to shape the future of travel in India, offering a more accessible and streamlined path for travelers across the nation.

London City Hall officials confirm government is planning specific policy around data centres to address power, water concerns

Getting your Trinity Audio player ready...

London City Hall is developing a policy around data centres to address concerns about carbon emissions and power and water use, Greater London Authority officials have said.

Megan Life, GLA's assistant director for environment and energy, told the London Assembly Environment Committee that the policy would aim to balance the economic benefits of data centres with their environmental impact, the BBC reported.

'Challenging' issues

She said the policy was intended to "keep hold of the kind of economic growth benefits that data centres offer" while mitigating "quite challenging" issues around power use.

Deputy mayor for the Environment Mete Coban said data centres brought both "big benefits" and "massive challenges" for the capital on energy and water consumption.

He said it was important to protect the environment from "a few global corporations who will take and not give back".

"It's not just a London problem, it's going to be a global problem," he added.

Housing delays

In December the London Assembly Planning and Regeneration Committee demanded a specific data centre policy in the next London Plan after finding that several housing projects in west London had been delayed by data centres using all available grid capacity.

At the end of last month, campaigners staged two days of protests in London to highlight the dangers posed by the unchecked expansion of data centres to society and the environment.

Several large data centre and AI companies, including Microsoft and OpenAI, have introduced policies aimed at quelling concerns around the impact of the large-scale data centre build-outs they are planning.

Revolut's MCP experiment exposes a widening gap between crypto-native firms and traditional retail brokers still competing on interface design.

Two engineers at Revolut's crypto exchange built a fully functional market-making system in about half an hour. Not a prototype. Not a carefully prepared demo. Just a first idea, executed using Anthropic's Claude AI connected to the Revolut X trading API through a protocol called MCP.

The experiment is raising a question the retail trading industry may not be ready to answer: if an AI agent can orchestrate an entire trading workflow through a simple text prompt, what exactly is a broker's platform worth anymore?

Nikita Ivanov and Vlad Kaminski, engineers on Revolut's crypto team, built the integration as a personal experiment, according to Leonid Bashlykov, Revolut's Head of Product for Crypto.

"Two A-players on my team...were playing with Claude automations and built the MCP as a side project, just for fun," Bashlykov wrote on LinkedIn. "Then we saw what it could actually do. And it broke our product roadmap thinking."

Bashlykov said he personally built a working market-making strategy using Claude, MCP, and the Revolut X API in 30 minutes, with no advance preparation. The workflow covered everything from portfolio screening and position sizing to execution, monitoring and automated alerts. We are talking about tasks that traditionally require separate tools, scripts, and a fair amount of technical expertise to connect together.

MCP, or Model Context Protocol, is an open standard developed by Anthropic that allows AI models to discover and use external tools through a standardized interface. Rather than requiring custom-built integrations for every API, MCP lets an AI agent connect to any compatible system and reason across all of them simultaneously. Revolut X is currently offering MCP access in beta.

The practical implication, according to Bashlykov, is that complex trading strategies no longer require coding knowledge or platform fluency. "Rebalance to 60/40 BTC/ETH if BTC dominance drops below 52%," he wrote as an example. "That's now a prompt."

Fintech analyst Linas Beliūnas, who covered the story in his Weekly Fintech Pulse newsletter, framed it as a potential architectural shift. "The agent reads portfolio data, checks conditions, pulls context - then decides," he wrote, contrasting MCP-connected agents with rule-based trading bots that simply execute predefined instructions.

Beliūnas noted that the workflow Bashlykov described, handling inventory management, dynamic quoting, position sizing, execution, monitoring and Telegram alerts, "normally requires a quant, a developer, backtesting infrastructure, and weeks of iteration."

Most AI trading tools currently on the market, both Bashlykov and Beliūnas agree, are still limited to price alerts, simple queries, and basic automations.

What makes MCP different, the company claims, is composability. A single connected workflow can simultaneously execute trades, read news, track on-chain data flows and update dashboards. Revolut X functions as the execution layer; Claude functions as the interface.

Revolut is not entirely alone. CFD broker ATFX partnered with data firm KX late last year to deploy an AI-driven MCP server for real-time trading data, and LSEG connected its market data feeds directly to ChatGPT in December 2025, allowing institutional users to pull live information on demand. A FinanceMagnates.com analysis from November 2025 noted that crypto-native firms were already moving into AI's data layer while asking whether traditional brokers would follow.

The answer, so far, is: slowly. Firms like XTB, IG Group and Plus500 have spent years and significant capital building proprietary trading interfaces. XTB entered 2026 projecting $186 million in income while doubling down on platform growth and client acquisition, a roadmap built around the assumption that the interface is the product.

That assumption is now being tested. A comparison of client metrics across IG, CMC, Plus500 and XTB published last August showed that client numbers are growing across the board, but revenue per user varies dramatically. It signals that platform stickiness, not just user count, drives broker economics. If an AI agent can replicate a platform's functionality through a plain-language prompt, the stickiness argument weakens considerably.

Revolut's position in this shift is not accidental. The company has systematically built API-first infrastructure across asset classes. A partnership with CMC Markets Connect in 2024 brought CFD trading capabilities into its ecosystem through an API arrangement. A similar deal with GTN added bond trading for EEA customers the same year.

Revolut X itself launched with ambitions targeting a $200 billion European crypto market, with Bashlykov cited at the time as the driving force behind its expansion.

That API-first architecture is precisely what made the MCP experiment possible. A broker whose trading infrastructure lives inside a proprietary platform, rather than exposed through clean APIs, cannot simply plug in Claude and tell it to start market-making.

"The question is to which point it continues to make sense bringing UI updates to the product if AI enabled flow allows so much more capabilities," Bashlykov's concluded.

Retail brokers have spent years competing on better charts, faster buttons, and cleaner mobile apps. But a 30-minute session on Revolut suggests that race may already be running in the wrong direction.

The firms that built their platforms for screens are now watching rivals who built for APIs pull further ahead. And the gap is starting to show.

By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions.

In Singapore, more than 3 in 4 consumers are already using artificial intelligence to help them shop. They research products, compare options and navigate the sprawl of digital commerce through large language models that have become a routine part of daily life. What they have not done yet is hand those AI systems their wallets.

That gap, between AI-guided decisions and AI-executed transactions, is the fault line running through Visa's Asia Pacific strategy. And according to Stephen Karpin, the company's Asia Pacific president, the region that bridges it first will not just win a market. It will write the playbook for the rest of the world.

In a conversation with PYMNTS CEO Karen Webster, Karpin laid out the infrastructure already in place, the trust barrier still standing in the way, and why he believes the next chapter of global commerce will be authored here.

The foundation, Karpin argued, is already there. Seventy-three percent of online transactions across the region are conducted on mobile devices. Eighty percent of product discovery happens through mobile-first channels. Super apps, QR payments and real-time rails have produced a consumer base that moves fluidly between payments, shopping, messaging and financial management. Often within a single interface.

Karpin was direct about where that leaves the region on the agentic frontier.

"The infrastructure is strong, with connected devices and there's a lot of tokenization out there, which is a very important precondition," he said. "There's no scaled agentic payment," he acknowledged. "But the pilots are happening."

In other words: the runway is built. The question is whether trust will let the plane take off.

But infrastructure and readiness are not the same as adoption. And here, Karpin introduced the variable that no amount of tokenization can solve on its own.

"Trust is fundamental as it's always been in commerce. And when you're moving money, it is critical," he said.

That observation carries particular weight in the agentic payments context, where consumers may authorize software agents to initiate transactions on their behalf. Even as generative AI has already moved rapidly into search, product recommendation and customer engagement.

The Singapore data makes the point precisely: over 75% of consumers there are already using LLMs in their commerce experience, yet as Karpin noted, "it hasn't moved to payments yet." The reason is not technical.

"The trust element is the main factor that is having people have their reservations, as it should be," he said. Browsing with an AI is one thing. Letting it spend is another.

Payments, unlike browsing or discovery, involve direct financial exposure. For institutions designing agentic frameworks, the challenge is earning it.

Mobile adoption explains readiness, and a cohesive ecosystem is essential.

"Interoperability is central," Karpin said, describing cross-network connectivity as one of Visa's core responsibilities. Innovation often occurs domestically, shaped by local regulation, consumer preferences and infrastructure investment. Scale, however, requires cross-border continuity.

Karpin framed Visa's role as a connective one. Partnerships with domestic networks, wallets and financial institutions allow payment credentials to function consistently across markets. He described this approach as a "network of networks strategy," enabling stored-value wallets and QR ecosystems to extend beyond domestic limitations.

In practical terms, this reduces fragmentation, and fragmentation erodes trust. For merchants, broader connectivity expands acceptance pathways. For consumers, it preserves the familiarity of their preferred wallets and rails while extending their reach across borders.

Interoperability is not just plumbing. Karpin described it as the precondition for the next leap: what he called the age of "intelligent commerce," enabled by "intelligent payments." The phrase signals a structural shift in how transactions are constructed. Traditional payment flows largely treat intelligence as an adjunct, layered through marketing engines, loyalty programs or post-transaction analytics. Intelligent commerce embeds decision logic directly into the payment experience.

If interoperability is the infrastructure of trust, Visa's Flex Credential is where that trust meets the transaction. Initially designed to let consumers choose dynamically among funding sources -- points, credit, debit, BNPL -- its significance becomes exponentially larger when the entity making that choice is an AI agent rather than a human.

"If you think about an agent being equipped with multiple sources of funds," he said, including points, buy now pay later, revolving credit, multicurrency accounts and potentially stablecoins, well, that's a disruptive proposition.

Traditional payment instruments operate on static funding logic. A card is a card. Flexible credentials introduce dynamic selection, turning the payment moment into an exercise in real-time optimization. The question shifts from "which card do I use" to "which funding source produces the best outcome right now."

Under agentic commerce, that decision migrates entirely from the consumer's hands to systems operating in the background, systems that, if trusted, could optimize every transaction the consumer never has to think about again.

The interoperability thesis is not theoretical. Visa's activity in China, shaped by regulatory constraints distinct from every other APAC market, offers a live proof of concept.

The partnership with China UnionPay enables mainland consumers to move funds across borders while preserving domestic infrastructure linkages. It is not disruption, Karpin suggested. It is collaboration between networks that each bring something the other cannot replicate alone. That model, he indicated, is the template.

One further development signals how close the agentic future already is. Karpin pointed to the expanding universe of self-managed payments, individuals and small businesses operating simultaneously as payers and payees. Visa Accept, for example, now enables debit and prepaid cardholders to receive payments directly into their Visa accounts, collapsing the distinction between a consumer credential and a merchant one.

For gig workers, micro-merchants and platform participants, this is already the reality: a single credential that receives, spends and manages funds within one framework. The payment account has become operational infrastructure.

In the agentic environment Karpin described, those same accounts evolve into programmable hubs, not just holding funds, but executing financial instructions on behalf of the people and businesses they serve. The consumer becomes less a user of the payment system and more its beneficiary.

Karpin was careful not to overstate the timeline. Consumer-scale agentic payments remain emergent. The pilots are running, not scaling. The trust frameworks are under construction, not complete.

But he was unambiguous about the direction. "I do have full conviction that trusted agents will scale up," he told Webster. "I think agents will be making payments at a very grand scale ... the Asia Pacific region will lead the way, in many different measures, in the months and years to come."

That confidence rests on something more durable than enthusiasm.

The mobile infrastructure exists. The tokenization layer is in place. The interoperability frameworks are being built. The credentials are becoming flexible. All that separates the region from the agentic commerce Karpin described is the one thing that has always separated promise from payment: trust.

And trust, as he noted, is not a technology problem. It is a track record problem. Asia Pacific, he suggested, is building one.

As part of our Women in Tech series, we talk to Sukhmani Lamba, a senior product manager at Microsoft, who holds multiple patents in conversational AI. Her work focuses on agent memory, agent intelligence, and agent context at Microsoft Teams.

While growing up in Nigeria and later in Bahrain as the child of expatriate parents, Sukhmani Lamba recalls her father bringing home the first laptop. Soon her world opened up to Microsoft Word and Paint. She also remembers the oversized, clunky mobile phone. But what was deeply imprinted on her young mind weren't the gadgets but the innumerable ways in which technology could empower people.

"Technology was empowering both in the physical barriers it was dissolving and the opportunities it was creating," she says.

Now a senior product manager on the AI Agents Platform at Microsoft Teams, with three patents to her name, Lamba is part of the new wave shaping the future of work -- building at the intersection of enterprise software and conversational AI.

The early belief and conviction that technology would change lives helped chart the course of her entire career.

After completing her degree in electronics and electrical engineering from Punjab Engineering College, Lamba joined Deloitte to work on digital transformation projects for enterprise clients.

But consulting, she discovered, was like advising from a distance.

"It's a lot about processes; you are suggesting improvements and the technology that can be used, but there is less control over the actual core product you are building," she explains.

This restlessness pushed her towards product management.

Lamba left for the United States to pursue a master's in engineering management at Duke University, specialising in product management. The programme introduced her to Pendo, a startup where she worked under a work-study arrangement. Here, she got a taste of building from the inside.

After graduating, she joined Wayfair in Boston, a Fortune 500 ecommerce giant, as a product manager. But her ambition was always to break into the world of communication software.

At that time, Slack was flourishing, and Microsoft Teams was rising. When the Covid-19 pandemic, it changed the way the world communicated.

The inflection point

As the world rushed online, overnight, platforms like Zoom, Slack, and Teams went from productivity tools to lifelines. For Lamba, it was a signal she had been waiting for.

She joined Microsoft Teams during the height of the pandemic -- a platform now used by approximately 320 million people worldwide.

Lamba explains what she does here.

"When you think about how work happens in enterprises, Teams is one place where people come to communicate. But realistically, you might be going to Word for a document, to another system for your timesheet, to a project board for your weekly metrics, and to a separate system for your bugs. There's a lot of different software.

"My job was: when people are in Teams, how can we make sure they can actually complete whatever they are trying to do right from within Teams?"

This understanding led to three patents in her name -- innovations centred on how shared links are interpreted, understood, and made immediately actionable inside Teams, so users can close the loop on a task without ever leaving the platform.

These are important tasks -- a contract that needs signing, a document shared for review, or a link pasted in a chat. All these moved from a state of friction to free flow.

The results were far-reaching. By enabling users to interact with shared content directly within Teams, Lamba's innovations helped tens of millions of people work more seamlessly without disrupting their workflows, influencing how billions of links are handled on the platform.

The standards she helped develop are now used by major global technology companies, as well as hospitals and educational institutions.

Advent of ChatGPT and its impact

In late 2022, ChatGPT arrived. Within Microsoft, an early investment in OpenAI had already been made.

A small group within the company understood that the enterprise world was about to be transformed. Lamba was one of the earliest people working on the enterprise story.

When ChatGPT launched its plugins, a system that allows users to bring external software directly into their AI conversations, Lamba was already building the Microsoft equivalent.

"I'd worked on a company-wide strategy for Copilot plugins. It started from me, from Teams, and became a Microsoft-wide adopted standard," she shares.

The work was significant enough to be featured in the keynote of Microsoft's executive vice president.

From there, Lamba moved on to agentic AI. She led the platform strategy for Microsoft's flagship collaborative agents for Teams -- Facilitator Agent and Channel Agent. One was announced in 2023 and the other in 2024.

Today, Lamba is working on something she describes as the "agent brain".

Her current work focuses on agent memory, agent intelligence, and agent context in Microsoft Teams. This is the infrastructure that determines what an AI agent knows about you, your preferences, your personality, your projects, and your organisation, and how it uses that information to serve you most effectively.

AI and the fear of losing jobs

While Lamba acknowledges that the engineer's coding role has largely transformed due to AI, she insists that the human aspects of work, such as strategic thinking, alignment, persuasion, and driving, are not going anywhere.

"You are focusing more on thinking. The actual sitting down to do it is where AI is taking over," she says.

Will AI disproportionately affect women, particularly those who take a break from their careers? Lamba admits the acceleration is real.

"For anybody who has been out for even six months, the industry has moved light years ahead. But many people have been able to upskill very easily with AI. Even if I'm returning after leave, I can quickly understand what decisions have been made, what my manager is focused on, and what the market is doing. It's almost like having a personal assistant," she says.

Beyond her role at Microsoft, Lamba sits on the Products That Count Advisory Council, a prestigious, invite-only group of over 500,000 product managers, CPOs, and industry leaders from top firms like Google, Amazon, and Meta, helping shape best practices for product management in an AI-first era.

As an executive in Residence at Mitre Capital, a San Francisco-based VC fund, Lamba advises Series A and Series B enterprise startups. She is also involved with Products by Women, a community focused on upskilling women in the field.

In the academic world, she reviews papers and conferences at the intersection of AI, human-computer interaction, and the future of collaboration.

Lamba is candid about what it still feels like to walk into a room where the majority are men.

"It used to be very intimidating at the start. Even now, there is a little bit of impostor syndrome. But I just put it aside. It's there. My advice to anyone would be: just own what you know and speak about it," she says.

On where she sees herself in five years, Lamba does not hesitate to proclaim her love for what she's doing.

"I am 100% going to still be doing AI. I feel very passionately about enterprise productivity -- but human productivity is the other thing. And collaboration. That's exactly what I do right now, and I love it," she says.

By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions.

Take, for example, AWS. It says Amazon Connect Health is a new healthcare-focused agentic AI offering designed to take on routine administrative work that often pulls staff away from patients. In the article, AWS says the product can help with patient verification, appointment scheduling, clinical notes and medical coding, while working inside the systems healthcare organizations already use.

The company argues that this matters because providers spend too much time on manual tasks such as gathering data from different systems, and patients are frustrated by delays and hard-to-navigate processes. AWS also says the product is built to keep clinicians and staff in control by showing the source behind AI-generated outputs and by handing patient-facing tasks to a human when needed.

The article also makes the case that better results depend on better access to healthcare data. AWS says Amazon Connect Health connects with AWS HealthLake and partner systems so organizations can bring together records from many sources and use that information in a more meaningful way. AWS highlights early customer examples from Amazon One Medical, Netsmart, Veradigm, Greenway Health and Pelago, saying these organizations are using the technology to reduce documentation work, speed coding and give clinicians more time with patients. Overall, the piece presents Amazon Connect Health as a practical attempt to put agentic AI to work in healthcare by easing paperwork, improving patient access and fitting into existing clinical workflows rather than forcing providers to adopt entirely new ones.

In an article from Google, the company says healthcare is moving from simply storing digital records to using agentic AI to act on that information. The main idea is that AI agents can do more than answer prompts. They can help carry out tasks across scheduling, claims, documentation and patient support. Google says this shift could reduce the manual work that slows down care and frustrates both staff and patients. The article also stresses that these agents need access to the full picture, including text, voice and other forms of data, so they can give workers useful help at the right moment.

Google builds the article around advice as much as product news. The piece argues that agentic AI works best when it is tied to real healthcare workflows and backed by strong safeguards around privacy, compliance and data control. It also suggests that success comes when organizations move beyond small experiments and apply AI to practical, high-volume problems such as customer service, research paperwork, revenue cycle work and helping patients understand their lab results.

The article points to examples from CVS Health, Highmark Health, Humana, Quest Diagnostics and Waystar to show how companies are trying to turn AI agents into tools that save time, improve clarity and support better decisions. Overall, Google presents agentic AI as most useful when it helps people take action on healthcare data instead of simply collecting more of it.

Agentic is, after all, meant to solve the problem of automating manual tasks. In an article from The Fast Mode, agentic AI is presented as a practical way for healthcare organizations to ease pressure on staff and improve how work gets done.

The piece argues that hospitals and clinics are dealing with too many disconnected systems, too much manual work and growing demand from patients who want faster digital access. Against that backdrop, agentic AI stands out because it can handle multistep tasks across scheduling, billing, intake and care coordination with less constant human direction. The article says this can help reduce burnout, support short-staffed teams and improve patient access, especially when AI agents are used to support always-on service across digital channels and phone-based interactions.

The advice in The Fast Mode article is straightforward: start with the workflow, not the technology. The author says healthcare organizations should first map how information moves between patients, staff and systems, then decide where agentic AI can remove friction. The piece also says leaders should match the tool to the task, since some simple, repetitive work may only need basic automation while more complex work may justify agentic AI.

It stresses the need to measure results clearly, monitor agent performance across full workflows and build trust through traceable outputs, audit logs, human oversight and strong security controls. The central message is that agentic AI works best when it is tied to real business problems, connected to existing systems and governed with clear rules for accountability.

The custom of finals week appears oddly muted on a quiet weekday morning in a public high school outside of Boston. No blue exam booklet piles. There are no educators prowling the aisles like guardians of a national secret. Rather, rows of pupils stare at blank prompt windows while sitting with laptops open. One sentence could contain all of the instructions for their final exam.

"Create the most effective prompt you can." The American high school final exam followed a well-known format for many years. Handwritten essays, multiple-choice exams, and perhaps a few diagrams scrawled on scratch paper. In the hopes that their preparation would match what was on the exam sheet, students committed facts, formulas, or historical dates to memory. Although it wasn't flawless, it was predictable.

This school made the decision to try something different. Last year, teachers noticed a pattern that was hard to ignore, so they quietly started the experiment. Submissions of homework were becoming unusually polished. Essays read like expert commentary. Math explanations seemed flawlessly organized, at times in an unsettling way. Artificial intelligence tools have quickly become a part of student life.

One English teacher recounted her realization that the ground had changed. She witnessed a student discreetly paste an entire chapter into an AI chatbot during a discussion of Frederick Douglass's autobiography and receive a polished paragraph of literary analysis in a matter of seconds. As if everyone had read the same prewritten interpretation, the class discussion that ensued sounded oddly hollow.

She later told coworkers, "It flattened the discussion." There had to be a change.

The simplicity of the school's solution was radical. Rather than outlawing AI tools, which many educators believed would fail almost instantly, they chose to integrate them straight into the testing procedure. The AI's output would not be used to grade students. The prompt itself would be used to grade them.

To put it another way, the question turned into the test. It feels different to watch students navigate this new system than it does in a traditional exam room. The atmosphere has changed to one of experimentation, but there is still tension -- finals week always carries a certain nervous energy. Some students try to get better answers from the AI model by typing slowly and repeatedly editing prompts. Before writing anything at all, others take a seat with their arms crossed and ponder.

A junior shrugged as he explained the procedure. He remarked, "It's like arguing with the computer."

Instructors maintain that the approach assesses more than just memorization. Creating a good prompt necessitates having a thorough understanding of the topic in order to direct the AI toward helpful responses. The chatbot's response quickly becomes ambiguous or deceptive if a student does not understand the fundamental idea.

The learning process now includes that feedback loop. However, there are those who disagree with the experiment.

Some educators contend that using AI tools during tests makes it harder to distinguish between machine assistance and student knowledge. After all, the purpose of traditional final exams was to demonstrate a student's ability to remember and apply knowledge on their own. For some educators, integrating AI into the testing environment is tantamount to surrender.

It's easy to feel that tension as you walk through the hallway outside the testing rooms.

Academic integrity posters are still displayed on the walls. Thick textbooks that once defined the curriculum are still on the library shelves. However, students are using technology in the classroom that can quickly summarize entire chapters.

The change feels both academic and cultural. There isn't a single national education system in the US, and schools frequently try out various teaching strategies. This adaptability has led to the creation of project-based learning initiatives and charter schools. However, incorporating AI into the fundamental framework of tests seems like a more significant shift.

Some educators are concerned that it might promote mental short cuts. Others hold the opposite view.

During a staff meeting, one science instructor put it plainly. He stated, "Students already have AI in their pockets." "They don't learn anything from pretending otherwise."

He contends that teaching students how to use technology responsibly should be the main goal of education. They are forced to think critically about what they want to learn when specific prompts are written. It also shows how easily, depending on the wording, an AI system can be led to either deeper analysis or superficial responses.

During finals week, that lesson becomes surprisingly apparent as you watch students test various prompts.

An answer to a vague question is generic. A well-made one yields something much more intriguing.

Additionally, a minor psychological change is occurring. Recall -- remembering definitions, dates, and equations -- was frequently rewarded in traditional exams. Exams that are prompt-based reward curiosity. Pupils who ask more insightful questions typically get more insightful answers.

There is a philosophical quality to that dynamic. The notion that knowledge is passed down from teacher to pupil has been the foundation of education for centuries. A third person is now present in the room as technology interjects itself into that conversation. The process is guided by teachers. Students investigate it. The AI silently sits in the center, answering any query that comes up on the screen.

As this develops, there's an odd mix of hope and trepidation. Prompt-based tests might continue to be a specialized experiment and an interesting anecdote in the history of educational reform. Many innovations have been tried in schools in the past, but not all of them have been successful. However, there's also a feeling that something basic has changed.

Pupils are no longer merely taking in knowledge. They are learning how to question it.

And that ability -- the ability to ask the right question -- may prove to be the most difficult test of all in a world where answers can appear instantly on a screen.

GenieAI, a pioneer in legal artificial intelligence infrastructure, today announced the launch of Eidetic Intelligence, a patent-pending AI architecture purpose-built for legal work.

In testing, Genie's system achieved 90% accuracy in simulated legal risk assessments, beating all other AI LLM providers tested.

The innovation, filed with the UK Intellectual Property Office on 3 February 2026 under the patent application LW1: Variance Control, is designed to address the failings of general-purpose AI in legal contexts, where imprecise recall, hallucinations, and fragmented reasoning can create critical errors.

General-purpose AI struggles with the realities of legal work. It can't reliably cross-reference dozens of documents, track financial figures, or analyse regulatory gaps, and identical prompts often produce different results. Without built-in validation, these systems risk generating unenforceable contracts or fabricated figures. Eidetic Intelligence takes a fundamentally different approach, giving lawyers deterministic, auditable outputs they can trust.

Genie's platform is built on top of foundation models like Claude and OpenAI, but extends them with infrastructure and workflows that lawyers rely on. The differentiator is not the model itself, but the legal editor, specialised datasets, and validation systems layered on top - turning a general-purpose AI into a tool lawyers can trust for high-stakes decision-making.

The announcement comes as the sector responds to Anthropic's launch of the Claude legal plugin, which has highlighted a wider shift toward AI-assisted transactional work. For GenieAI, which has spent eight years building dedicated legal AI infrastructure, the moment underscores a transformation in the professions - lawyers will focus less on drafting and more on oversight, governance, and strategic decision-making. This shift requires tools that can handle the complexity of real-world legal work without error.

To test how Eidetic Intelligence performs in these conditions, Genie conducted a rigorous benchmark simulating Tesla's European expansion, analysing 65 source documents including contracts, board minutes, regulatory filings, and financial statements. Each AI system was evaluated across 15 legal quality metrics, from factual accuracy to regulatory coverage.

Genie AI not only delivered the most complete risk assessment but also surfaced systemic patterns, including supplier concentration escalations and regulatory exposure gaps, that other systems missed entirely.

Rafie Faruq, CEO and co-founder of Genie AI, said: "Legal work is unforgiving of mistakes. A single missed clause or inaccurate figure can cost millions, derail deals, or expose companies to risk. Eidetic Intelligence enables lawyers to process hundreds and thousands of documents, cross-reference clauses, and track timelines and counterparties, giving them confidence to make high-stakes decisions. This isn't about doing legal work faster. Eidetic Intelligence turns AI into a trusted partner that delivers accurate, auditable guidance for oversight, governance, and strategic legal decision-making."

Eidetic Intelligence operates as a Quality-Gated Self-Correcting State Machine, imposing deterministic control over every stage of a legal workflow.

Key innovations include:

Together, these features allow Genie AI to avoid the hallucinations, context decay, and probabilistic errors that affect general-purpose models.

Eidetic Intelligence represents a potential paradigm shift in legal AI, providing tools for law firms, corporate legal departments, and compliance teams to produce auditable, high-confidence risk assessments at scale. By proving the gap between general-purpose and purpose-built legal AI, the announcement raises fundamental questions about the adoption of mainstream chatbots for high-stakes legal decision-making.

Beyond benchmarks, Genie AI is shaping the future of legal practice. From its partnership with Charleston School of Law, where the platform has been integrated into its transactional law curricula, to enabling a startup to become the first company to close a £950,000 seed round using A instead of a traditional law firm.

This latest announcement cements Genie AI as a pioneering force in legal AI, demonstrating that foundation models alone are insufficient for high-stakes legal work. Instead, specialised datasets, workflow layers, and validation systems define the next generation of legal technology.

According to Axios, the company maintains its belief in the principle of treating adults like adults, but acknowledged that delivering the right experience requires more time and careful development.

The feature was first announced by OpenAI CEO Sam Altman in October 2025, when he said the rollout would likely happen in December the same year. However, plans shifted. TechCrunch reported that in December, Altman circulated an internal memo declaring a "code red," urging teams to prioritise strengthening the core ChatGPT experience through the first quarter of 2026.

So far, OpenAI has not provided a revised timeline for the feature's release.

What "Adult mode" is expected to offer

The central idea behind "Adult mode" is to tailor ChatGPT interactions more appropriately for mature users. As part of this broader approach, OpenAI introduced age prediction globally in January 2026. This system estimates whether an account is likely operated by a minor or an adult.

If a user is identified as an adult, the new mode would allow greater exposure to adult-oriented material, including erotica, compared to accounts believed to belong to minors.

Altman has also indicated that the update will introduce expanded personality controls. Rather than interacting with a single default assistant tone, users would be able to choose communication styles -- ranging from casual and emoji-rich to friendly or more natural, human-like responses. The goal is to make conversations feel more flexible and better suited to different contexts.

Another key change involves moderation. Some topics that are currently restricted may become permissible under a more relaxed framework, giving adult users greater conversational freedom. However, the company is still expected to maintain safeguards around harmful and sensitive content.

Lessons from other AI platforms

While OpenAI continues refining its approach, similar features on rival platforms have already stirred controversy. Users of xAI's chatbot Grok have had access to erotica-related capabilities for some time.

The tool faced criticism after users began generating sexually explicit images, including prompts that attempted to "digitally undress" people in photographs. Many of these cases reportedly targeted women, including real individuals who had not consented to such portrayals.

A report by a famous publication cited researchers who reviewed thousands of AI-generated images and found a significant portion depicted people wearing minimal clothing such as bikinis or underwear. Women represented the overwhelming majority of subjects in those images.

The trend gained traction on X, where users could publicly tag Grok to request edits or generate visuals.

Concerns escalated further when researchers identified instances in which users prompted the chatbot to create sexually suggestive images of individuals who appeared to be minors. Critics argue that such incidents highlight the dangers of combining powerful generative AI systems with massive social media reach without sufficiently strong safeguards.

A cautious path forward

Given these developments, similar risks could emerge once ChatGPT's "Adult mode" becomes available. If OpenAI is taking additional time to implement stronger guardrails and responsible-use protections, many observers believe the delay may ultimately serve users' safety and well-being.

South Korean go grandmaster Lee Se-dol has publicly taken on artificial intelligence(AI) for the first time since his historic encounter with AlphaGo in 2016.

The event was held Monday at the Four Seasons Hotel in Seoul, where Enhans, the Seoul-based AI startup that provides a specialized AI agent platform, declared the beginning of the "AI collaboration era."

Highlighting the concept of agentic AI, the company demonstrated that AI has evolved from a system that competed against humans to a collaborative partner capable of carrying out human intentions.

The demonstration took place at the same venue where Lee's landmark match against AlphaGo was held nearly a decade ago.

During the event, Lee worked in real time with the company's AI agent, using voice commands to reconfigure an AI operating system and a go model on the spot.

He then played a match against the newly created AI system.

Thick folders of proposed regulations are carried by policy staff as they travel between meetings on a soggy afternoon in Brussels, close to the glass-fronted headquarters of the European Commission. From the outside, it appears to be standard bureaucracy -- just another week in the European capital where digital regulations and directives subtly transform international industries. However, the discussions in those hallways have become more acute recently. Few policymakers appeared to be fully prepared for the collision of trade policy, geopolitics, and artificial intelligence.

New financial pressures, such as tariffs or regulatory penalties related to AI services, are being considered by the European Union for American technology companies. Retaliation is already being hinted at by officials in Washington. And not the kind that involves exporting wine or automobiles. Data itself may be the battlefield this time.

Tariffs targeting American AI companies may initially appear to be just another phase in the protracted conflict between Silicon Valley and Brussels. Through the introduction of legislation such as the Digital Services Act and the Digital Markets Act, Europe has spent years tightening regulations on Big Tech. EU regulators claim that the objective is fairly straightforward: stop powerful tech companies from abusing their influence in European markets.

However, businesses like Google, Meta, Microsoft, and up-and-coming AI firms like OpenAI that are caught in the crossfire have different perspectives. There is a persistent suspicion in Silicon Valley boardrooms that European regulations have become unusually antagonistic to American companies.

Investors have been observing the tension with a mix of interest and trepidation. According to some estimates, the United States currently enjoys a huge surplus in digital services trade with Europe -- more than €140 billion. Cloud services, software licensing, data analytics, and increasingly AI platforms are the channels through which that money moves. Washington might feel obliged to act if tariffs or other restrictions begin to erode that source of income.

Furthermore, the response being discussed in private has nothing to do with tangible goods.

It has to do with information. According to some trade analysts, the United States may impose restrictions on European businesses' access to cloud infrastructure, sophisticated AI models, and data services. Imagine a situation where American cloud platforms or machine-learning tools suddenly present challenges for European businesses. The effects on the economy could spread swiftly, especially for startups that depend significantly on infrastructure located in the United States.

The extent of this digital reliance becomes clear when one stands outside a data center in Northern Virginia, one of the biggest networks of internet infrastructure on the planet. Servers processing everything from banking transactions to AI training datasets are housed in massive gray buildings that hum softly. These kinds of facilities handle a large portion of the world's digital activity.

Europe is aware of this. Washington is also aware of this. Although it's not totally comfortable to use, the United States has leverage because of this imbalance. Restricting or stopping data flows could have unpredictable effects on international business. However, data is being discussed by policymakers more and more as a strategic asset, more akin to oil than information.

It took some time for the tension to start. European courts have been debating whether American businesses offer adequate privacy protections for the data of EU citizens for years. Judges overturned earlier frameworks intended to permit transatlantic data transfers on the grounds that U.S. surveillance laws violated European privacy rights. Diplomats had to rush to reach new agreements after every court ruling.

The relationship has now become even more complex due to artificial intelligence.

Large data sets, extensive computing infrastructure, and international information flows are all necessary for AI models. The United States is home to a large portion of that infrastructure. Meanwhile, European regulators seek more stringent control over the use of AI in their markets. The outcome is a policy impasse that occasionally resembles a philosophical disagreement about technology itself rather than a trade dispute.

There's an odd feeling of déjà vu as you watch this happen. Steel, agriculture, and manufacturing used to be the main causes of trade wars. Container ports and cargo ships served as emblems of economic competition. These days, the debates center on digital ecosystems, cloud storage, and algorithms that are invisible to the majority of people.

Even the words used to express retaliation sound different. Policymakers now talk about "data localization," "service restrictions," and "digital sovereignty" rather than tariffs on motorcycles or poultry. Despite their technical sound, these terms have a significant impact on the global tech sector.

Naturally, there are reasons for reluctance on both sides. Just as American businesses rely on European markets for income, European companies also heavily rely on American tech platforms. Every year, Silicon Valley companies receive billions from users and advertisers throughout the continent. Both economies may end up harming themselves if the conflict intensifies too much.

However, the attitude of some legislators indicates that patience may be running low. European officials contend that fairness, not punishment, is the goal of their regulations. However, more and more American lawmakers see the actions as discriminatory. Whether true or not, this perception has started to influence Washington's political discourse.

The extent to which either side is prepared to escalate the conflict is still unknown. In an effort to prevent a regulatory dispute from escalating into a full-fledged digital trade war, diplomats continue their covert negotiations. However, the possibility is now present, looming over policy debates like a storm cloud over the Atlantic.

One thing sticks out when observing the debate from a distance: the tangible world of trade disputes has gradually vanished into something more ethereal.

Not a ship. Not a single warehouse. Customs documents do not have tariffs printed on them.

Just code traveling through fiber cables, silently transporting the modern world's economic might.

X, owned by Elon Musk, is investigating "racist and offensive" posts by Grok.

Social media platform X, owned by Elon Musk, is investigating "racist and offensive" posts generated by xAI's AI chatbot Grok, according to a recent Sky News report. Grok has reportedly produced hate-filled content in response to prompts by several users requesting "vulgar" remarks. Amid ongoing concerns and backlash by governments and users alike, Musk recently posted on X that "Only Grok speaks the truth."

X Investigates Grok's Racist, Offensive Posts

These posts are part of a recent trend where users on X prompt Grok for unfiltered, vulgar commentary. Grok has been used to target religions like Hinduism and Islam. A Sky News analysis revealed highly offensive AI-generated replies containing profanities and racist posts targeting these religions.

The report also revealed that Grok falsely blamed Liverpool F.C. fans for the 1989 Hillsborough disaster, which claimed the lives of 97 supporters, while using derogatory language about the city. In response to the posts, Liverpool officials are now working to have the offensive content removed from the platform.

The U.K. government condemned the content as "sickening and irresponsible," stating it contradicts British values. While X has removed flagged offensive posts, no policy changes have been announced to curb them when Grok is prompted for vulgar responses.

Not Grok's First Time

Grok has been marred by such controversies in the past as well. In 2025, Grok generated offensive posts and anti-Semitic claims, for which xAI had to apologise, adding that the anti-Semitic incident occurred after a software update aimed at more human-like responses.

The chatbot has also been criticised for offering right-wing narratives without verifying the truth, instead amplifying unverified data especially in breaking news events.

Globally, regulators are also scrutinising Grok for explicit content, with xAI recently restricting image-editing features in certain regions to prevent output such as women clad in scant attires, which goes against the traditions of those regions.

'Only Grok Speaks The Truth': Musk

Amid these ongoing controversies, a recent X post from Musk reads: "Only Grok speaks the truth. Only truthful AI is safe. Only truth understands the universe." This comes on the back of the latest Grok update release, with Musk touting Grok 4.20 as the only "non-woke" platform, when compared with others like Anthropic's Claude, Google Gemini, and OpenAI's ChatGPT.

Musk's "truth" statement was in response to yet another controversial comparison done by a user using these chatbots. It included questions like "Are trans-women really women?" to which Grok replied "No," and "Should America deport illegal aliens?", responding to which Grok said "Yes."

Also read: Anthropic Is Measuring AI-Induced Job Loss Risk -- These 10 Occupations Are Most Vulnerable

Essential Business Intelligence, Continuous LIVE TV, Sharp Market Insights, Practical Personal Finance Advice and Latest Stories -- On NDTV Profit.

The world recently witnessed India host a week-long series of technology dialogues at the India AI Impact Summit last month in New Delhi. A Government of India led, multi-stakeholder forum, the summit brought together policymakers, senior bureaucrats, technology leaders, researchers, and representatives from multiple countries, all circling the same question, even if articulated differently, about what role India intends to play in the global artificial intelligence order.

The conversations were wide-ranging, spanning AI pipelines, data governance, compute infrastructure, public digital platforms, and the mechanics of collaboration between Indian and global technology players. Collectively, they sought to project India not merely as a country that services global demand, but as one that aspires to shape how AI systems are built, governed, and scaled.

Capabilities were showcased. Proofs of concepts, pilots, and early deployments from across India's AI ecosystem were put on display to signal technical, operational, and institutional readiness. The message was deliberate and consistent. India is open for collaboration, investment, and co-creation. Trust was the currency being courted.

That signalling was underlined by a set of carefully choreographed announcements. India positioned itself as a global convener through the Delhi AI Declaration, backed by dozens of countries and framed around democratic, inclusive AI governance. On the capability side, the government spoke of expanding sovereign AI compute through the national AI compute platform, promising access to tens of thousands of GPUs to reduce dependence on external infrastructure.

Industry commitments followed a familiar hierarchy. Reliance Industries outlined long-term plans to build large-scale sovereign AI infrastructure, while the Tata Group announced deeper AI partnerships and data centre expansion with OpenAI. Global technology giants such as Microsoft and Google reaffirmed India's role as a strategic hub for AI infrastructure, cloud, and skilling, rather than as a base for foundational platform ownership. Indigenous capability was highlighted through models like Sarvam AI and BharatGen, positioned as proof that India can build for its own linguistic and public-sector needs.

Taken together, the message was unmistakable. India is asserting relevance as a trusted partner, deployment ground, and responsible steward of AI, yet remains cautious, if not reluctant, about staking a claim as an independent owner of globally dominant platforms or models.

And yet, beneath the given narrative, a deeper unease remained. What does any of this actually mean for India? How does it matter to the Indian workforce? How does it impact sustainability? And once the summit lights dim, what is left beyond strategic signalling and familiar promises of partnership? These questions linger precisely because their answers are inconvenient, uncertain, and politically uncomfortable.

Once again, the narrative gravitates towards India's most overused advantage, a large, skilled, efficient, and cost-competitive workforce. The same workforce that powered global back offices for decades is now being repositioned as the engine of the AI economy, annotating data, fine-tuning models, engineering platforms, and operationalising intelligence for global systems. The labels have changed. The asymmetry often has not.

There is a growing view, often voiced softly but consistently, that India should not attempt to build foundational systems in direct competition with global giants. Instead, it should collaborate, integrate, and serve. Not become another OpenAI or Google, or a hyperscaler in its own right, but become indispensable to them. This argument is framed as realism. It is also a quiet acceptance of hierarchy.

But then, what about the startups? Who are they building for? Who are they training talent for? Who ultimately owns the intellectual property that emerges from this churn of pilots, demos, and proof-of-value exercises? Is the ambition merely to serve the Global North more efficiently, or to recategorise India from a services economy to an innovation economy? The distinction matters, even if the rhetoric pretends otherwise.

India's large corporates appear more confident in this moment, striking partnerships, expanding AI-led operations, and embedding themselves deeper into global technology supply chains. Yet the pattern of value creation remains familiar. Strategic control, platform ownership, and outsized financial upside continue to sit largely outside the country, even when execution happens within it.

Was Atmanirbhar Bharat ever meant to imply isolation or technological autarky, or is self-reliance without strategic depth merely a slogan? The real question is not whether India can collaborate, it already does, but whether it is willing to assert itself where power actually lies, precisely ownership, agenda-setting, and long-term risk-taking.

Is the hesitation driven by geopolitics? By capital constraints? By fear of failure at scale? Or by a deeper lack of self-belief in building technology that does not require external validation?

Until India is willing to move from being indispensable to being independent, from powering other people's AI ambitions to owning its own, the conversation will remain carefully incomplete. Summits will be held, partnerships will be announced, and talent will continue to be exported in all but name. India will keep proving that it can build, scale, and execute for the world, while quietly postponing the harder decision of building for itself.

Anthropic discovered 22 security vulnerabilities in Firefox using its Claude Opus 4.6 AI model in January 2026. Mozilla addressed these issues in Firefox 148.

The researchers state that AI models are now capable of finding high-severity software flaws independently. They identified 22 Firefox vulnerabilities in two weeks, 14 of which were high-severity, nearly a fifth of all high-severity Firefox issues fixed in 2025, demonstrating AI's ability to rapidly detect critical security risks in complex software.

In late 2025, Anthropic evaluated Claude Opus 4.6 on Firefox to test its ability to identify complex, high-impact security vulnerabilities. Initially, the model successfully reproduced many historical CVEs from older Firefox versions. Researchers then tasked Claude with finding new, previously unreported bugs, starting with the JavaScript engine. Within twenty minutes, Claude identified a Use After Free vulnerability, which the team validated and reported to Mozilla along with a proposed patch. While triaging, Claude discovered dozens of additional crashes, leading to a total of 112 unique reports across nearly 6,000 C++ files.

"After a technical discussion about our respective processes and sharing a few more vulnerabilities we had manually validated, they encouraged us to submit all of our findings in bulk without validating each one, even if we weren't confident that all of the crashing test cases had security implications." reads the report published by Anthropic. "By the end of this effort, we had scanned nearly 6,000 C++ files and submitted a total of 112 unique reports, including the high- and moderate-severity vulnerabilities mentioned above. "

Most issues, including high- and moderate-severity vulnerabilities, were fixed in Firefox 148, with remaining patches planned for future releases.

Mozilla praised the collaboration and began experimenting internally with AI-assisted security research. This project demonstrates AI's growing capacity to rapidly detect and report critical software flaws.

To test Claude Opus 4.6's ability to exploit vulnerabilities, researchers provided it with bugs previously submitted to Mozilla and asked it to create functional exploits. Claude attempted several hundred tests, demonstrating attacks that read and wrote local files, spending around $4,000 in API credits. It successfully produced working exploits in only two cases, showing that while the model excels at finding vulnerabilities, exploiting them remains far more difficult and costly.

"We ran this test several hundred times with different starting points, spending approximately $4,000 in API credits. Despite this, Opus 4.6 was only able to actually turn the vulnerability into an exploit in two cases. This tells us two things." continues the report. "One, Claude is much better at finding these bugs than it is at exploiting them. Two, the cost of identifying vulnerabilities is an order of magnitude cheaper than creating an exploit for them. However, the fact that Claude could succeed at automatically developing a crude browser exploit, even if only in a few cases, is concerning."

The successful exploits were "crude" and worked only in controlled test environments with security features like sandboxing disabled, meaning real-world impact would be limited. Nonetheless, Claude's ability to automatically generate even primitive browser exploits highlights the potential risks as AI-assisted offensive capabilities advance.

"These early signs of AI-enabled exploit development underscore the importance of accelerating the find-and-fix process for defenders." concludes the report. "In our experience, Claude works best when it's able to check its own work with another tool. We refer to this class of tool as a "task verifier": a trusted method of confirming whether an AI agent's output actually achieves its goal. Task verifiers give the agent real-time feedback as it explores a codebase, allowing it to iterate deeply until it succeeds. Task verifiers helped us discover the Firefox vulnerabilities described above, and in separate research, we've found that they're also useful for fixing bugs."

Mozilla reported that AI-assisted analysis uncovered 90 additional Firefox bugs, mostly fixed, including logic errors missed by traditional fuzzing, highlighting AI's growing role in security.

Something felt different at MWC 2026. Mobile World Congress has always been a place to spot the next big thing in connectivity, devices and digital infrastructure. This year, though, the mood had shifted. Walking the halls of Fira Gran Via, it became clear that AI is no longer something we only access through an app or a chatbot. It is starting to come to us, through our phones, in smart glasses and across the network itself.

I spent three days on the show floor trying products, listening to keynotes and speaking with executives. What stood out to me was how quickly AI is moving from being a tool we use to becoming something more like a working partner. It is starting to see, listen, interpret and respond in real time, and that shift has major implications for business.

Your Next Device Will Watch, Listen And Respond

One of the most memorable demos I saw was Honor's Robot Phone. It has a motorized camera that tracks you as you move, senses what is happening around it and reacts in a way that feels surprisingly lifelike. When the camera rises up, it feels less like waking a phone and more like activating a small robot that is ready to interact with the world.

The broader point here is that devices are becoming far more aware of context. They are beginning to understand what is happening around them and respond without needing constant instructions. For consumers, that may mean better videos or more helpful interactions. For businesses, the impact could be much more important.

Think about jobs where people are moving, using both hands or working in fast-changing environments. A field engineer does not want to stop and search through a manual. A nurse on a busy ward cannot keep breaking away to look something up on a screen. In these settings, a device that understands context and offers help in the moment becomes a very different kind of productivity tool.

For leaders, the important question is which roles benefit most when devices can understand their surroundings, and what rules need to be in place when those devices can also see and hear more of the workplace.

Smart Glasses Are Starting To Matter

The most interesting context aware interface trend I saw at MWC 2026 was smart glasses, especially display glasses from companies such as Meta and Alibaba.

One of my most striking experiences at the show was trying the Qwen AI glasses during a live conversation with someone speaking Chinese. I could hear their words translated into English as they spoke, while English captions appeared in front of my eyes in real time. It was one of those moments where you suddenly see the practical value of a technology.

This is where the enterprise opportunity becomes very real. When information appears directly in your line of sight, work becomes more continuous. You do not have to stop, pull out a device, unlock a screen and search for what you need. The information comes to you while you stay focused on the task in front of you.

The use cases are easy to understand. Real-time translation can remove friction in global collaboration. Hands-free navigation can help staff find the right place or asset on a large site. Contextual Q&A means a technician can look at a piece of equipment and ask what to do next, rather than leaving the job to find support. Each of these is a pain point organizations already recognize, but what is new is the way the technology can solve them.

Even more so than context aware phones, smart glasses move AI out of the screen and into the environment around us. AI becomes something people work alongside throughout the day. For glasses, it also raises some important questions. What can be recorded? Where is data stored? What does consent look like in practice? How do these tools connect with existing systems, knowledge bases and workflows?

Physical AI Is Showing Where The Cloud Falls Short

Robots and autonomous systems have been part of the MWC conversation for years. This year, they felt much closer to real commercial use. That also brought a major issue into focus. Cloud-only AI is often too slow or too limited for physical environments.

When decisions need to be made in milliseconds, or when sensitive operational data cannot leave a site, sending everything back to a distant cloud platform is not always practical. A practical example is a robot in a warehouse or on a production site. If it has to wait for a cloud response before it can react to an obstacle, that delay can create risk.

Edge computing is not new, but MWC 2026 showed that the discussion has moved on. We are now seeing networks and infrastructure designed to run AI workloads much closer to where data is created, with tighter links between connectivity and compute.

Satellites Matter More Than Ever

Another area of focus was satellite connectivity. Ground-based networks will continue to do most of the heavy lifting in places with strong coverage. Satellites matter when reach and resilience become more important, especially in remote locations.

As more AI-enabled work moves into the physical world, network failure becomes a much bigger issue. If operations depend on real-time data or autonomous systems, then loss of connectivity can quickly become a business risk. Satellite connectivity is starting to act as part of the resilience layer that makes edge AI dependable in real-world conditions.

Agentic AI Is Becoming The Operating Model

If one theme came up again and again in my conversations at MWC 2026, it was agentic AI. These are systems that can plan, decide and act with a degree of autonomy. They are moving from experimental projects into real-world use. Networks are a natural place to start because they are complex and constantly changing. AI agents can monitor conditions, spot likely faults and trigger or coordinate maintenance with far less human effort.

The same pattern is emerging in supply chains and cybersecurity, where work often depends on fast decisions across multiple connected systems.

Agentic AI changes the role of AI in the business. Traditional AI often helps people analyze information or make recommendations. Agentic AI goes further, it takes action. That also means the governance challenge becomes much more serious. Businesses need audit trails and accountability for what the system does. They also need to define where human oversight sits and which decisions can be handled independently.

The organizations that get the most value from agentic AI will be the ones that see it as an operating model, not just a feature. That means redesigning workflows, connecting systems properly, setting clear rules and measuring outcomes in a meaningful way.

What Leaders Should Do Now

MWC 2026 sent a very clear message. Enterprise AI is moving closer to action. It is moving into devices that understand their surroundings. It is moving into glasses that support people while they work. It is moving into networks designed to handle AI workloads as a core requirement. Agentic AI is becoming the coordination layer that connects all of this and turns capability into outcomes.

For leaders, the starting point should be practical. Focus on a small number of workflows where contextual AI can create clear value. Build the infrastructure needed to make those workflows reliable. Put governance in place before autonomous systems begin operating at scale. Set policies for new interfaces such as smart glasses early, so the organization is ready when pilot projects begin. What MWC 2026 revealed is that AI is moving out of the interface and into the real world of work, and that shift is only just beginning.

The infamous Stuxnet, claimed to be developed by the US and Israel, was a computer virus that sat inside the Iranian nuclear facility Natanz for at least two years. It was discovered in June 2010. An MQ-9 Reaper drone, remotely operated from Creech Air Force Base in Nevada, US, and launched from Al Udeid Air Base in Qatar, fired AGM-114 Hellfire missiles that killed Qasem Soleimani near Baghdad International Airport. Precision air strikes from the US and Israel, following decades of surveillance and hacking of traffic cameras in Tehran, killed Ali Khamenei on February 28, 2026, according to Iranian state media reports. In all three events, technology has played the greatest role in ensuring the success of the missions. Drones, precision attacks, spyware, cyberattacks, and AI are the tools and instruments of the US-Israel-Iran war. The coming days of the war, and future conflicts, will feature physical war AI. The robotic war games of our childhood are now coming to life.

Khamenei: Big Brother NSO Was Watching You

Spyware is one of the most critical tools used in modern high-tech warfare. Ali Khamenei was under Israeli intelligence surveillance for years. Israeli cyber-unit Unit 8200 and the foreign intelligence agency Mossad hacked into traffic cameras in Tehran and the mobiles of Khamenei's security team (Atalayar, 2026).

Software like Pegasus Spyware, developed by the Israeli company NSO Group, can almost instantly hack into an individual's mobile phone without alerting them. Jamal Khashoggi, a journalist killed at the Saudi consulate in Istanbul in 2018, is among the many victims of Pegasus. His location was compromised to his attackers after they infected his wife's phone with Pegasus.

Similarly, Mossad and Unit 8200 collected massive amounts of data daily, tracking every movement of Khamenei and his important associates in Tehran. These massive datasets were processed and analysed using algorithms, large language models, and generative AI in supercomputers. The supercomputers ultimately generated patterns and routines for Israeli forces, suggesting the best time and place to launch a successful attack.

In a full-fledged war, Israeli forces will not simply hack into cameras and mobiles. They will hack into Iranian oil infrastructure, which is heavily dependent on operational technology and industrial control systems.

Red Alert: Iran's Oil Operational Technology

In 2010, Stuxnet, a complex malware, attacked Iran's Natanz Nuclear Facility. The attack damaged 1,000 IR-1 centrifuges in the fuel enrichment plant. A computer virus resulted in the destruction of physical elements of a nuclear facility.

Israeli forces may have similar ambitions to destabilise Iran, especially its oil facilities. Kharg Island Oil Terminal, South Pars Gas Field, Abadan Refinery, and every other major oil facility in Iran are heavily dependent on operational technology. Operational technology refers to electronic devices and programmes that operate physical machinery in refineries, terminals, and oil fields. Launching cyber-attacks on such major facilities in Iran would give Israel a significant competitive advantage by choking the economic backbone of Iran.

In 2012, a virus called "Wiper" attacked the Iranian oil ministry's computer network and erased all data from the ministry's hard disks. At that time, the Kharg Island Oil Terminal was disconnected from the internet for damage control.

Ukraine faced a similar attack to Stuxnet in 2016, when the malware Crash Override attacked Ukraine's electricity grid. The attack resulted in a blackout in parts of Kyiv, as the malware could automatically communicate with power-grid control systems and switch electricity on or off using industrial protocols.

These incidents illustrate how cyber-attacks are transforming into automated systems and gaining their own independent damaging capabilities. Malware and viruses are becoming more intelligent. So are military weapon systems, drones, and artillery.

Is Iran the New Testing Ground for US-Israel AI Artillery?

Jensen Huang, CEO of NVIDIA, said that the next wave of AI is physical. Physical AI is entering the scene through the warzones in Iran.

Dwight D. Eisenhower, in 1961, coined the term "military-industrial complex." In a military-industrial complex, the military and weapons manufacturers form an alliance to push forward war ideas and stock out old weaponry so that they can be replaced by new inventions. Lockheed Martin, Boeing, and Raytheon Technologies are among the veteran defense manufacturers.

In recent times, with accelerating advancements in AI and automated decision-making, tech giants and startups have entered the game. This forms a new triangle between the military, the defense industry, and the technology industry.

Actors like Palantir Technologies, Anduril Industries, Microsoft, NVIDIA, and OpenAI are among the core structures of this mili-tech industrial complex. These companies are producing AI systems for almost every action on the battlefield that was previously performed by a soldier. Palantir Technologies provides battlefield intelligence platforms, targeting analytics in Ukraine and other conflicts. Anduril Industries provides autonomous drones, border surveillance, and AI defense systems. Microsoft provides military cloud infrastructure (Azure), AI systems, defense data platforms, and NVIDIA does GPUs and chips that power AI models, autonomous systems, and defense simulations.

The US pulled out of Afghanistan, and Israel is running out of justifications for committing genocide in Palestine. These new ranges of weapons, such as swarm drones using Anduril Lattice technology, require a new testing ground. Iran is set to become a testing ground for advanced technological weapons.

Research accelerates when there are no moral and ethical constraints attached. AI development itself, let alone merging it with weapons, faces serious ethical questions in the European Union through the General Data Protection Regulation and the European Union Artificial Intelligence Act.

The current US-Israel-Iran war creates a sense of emergency and justification for advancing AI-powered weapons. This creates a bypass for AI developers around AI regulators. AI systems leading war decisions and taking direct actions appear legally, ethically, and morally cloudy. When was war ever legal, ethical, or moral anyway?

Forbes contributors publish independent expert analyses and insights.

In today's column, I examine a newly filed lawsuit against OpenAI regarding the noteworthy aspect that ChatGPT is being allowed by the AI maker to provide legal advice.

Here's the deal. Presumably, contemporary generative AI and large language models (LLMs) are not supposed to be handing out legal advice. The act of doing so would seem to violate the various UPLs (Unauthorized Practice of Law) stipulations throughout the United States (other countries tend to have similar provisions, but not all do). Only lawyers are supposed to give out legal advice. Non-lawyers generally cannot do so, though you can act as your own "lawyer" if you wish to take that chance.

In this recently filed case, the plaintiff is a company that was sued by an individual for various claims, and the individual allegedly tapped into ChatGPT to devise legal filings. The legal filings were apparently all over the map, causing the company to seemingly inordinately expend money and resources to fight the case. The company seeks to be compensated by OpenAI for the presumed use of ChatGPT to legally aid this individual in suing them and wasting their time and effort.

Does this landmark lawsuit have any legs to stand on?

Let's talk about it.

This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).

AI And The Law

As a quick background, I've been extensively covering and analyzing a myriad of facets regarding the intersection of AI and the law for many years. You can find my writings not only in my Forbes column but also as posted in Bloomberg Law, ABA Law Journal, The National Jurist, The Global Legal Post, Lawyer Monthly, The Legal Technologist, MIT Computational Law Journal, and so on.

There are two major perspectives on the mixture of AI and law:

* (1) AI & Law. The application of AI to perform legal reasoning, and

* (2) Law & AI. The application of laws to the governance and regulation of AI.

Thus, you can apply AI to the law, and you can conversely apply the law to AI. For my big picture overview of both of these exciting and rapidly evolving realms, see my discussion at the link here and the link here.

I will be focusing here on the application of AI to perform legal reasoning. It is an intriguing problem that remains vexing and unsolved. There is a specialized field of research that concentrates on making advances in AILR (AI Legal Reasoning) -- see my books on the steady but uneasy progress toward attaining AILR, including my introductory book at the link here and my advanced book at the link here.

The overarching goal of AILR is to devise AI that can work on the same level as human lawyers and practice law as they do. You would not be able to differentiate the legal efforts of human lawyers from the AILR. They would be on par with each other. Think of this in the widest and deepest way possible, such that the AI is equal to whatever legal efforts and shenanigans that human lawyers undertake.

Some researchers hope to go even further. The aim is to craft AI that is superior to human lawyers and can be a kind of superhuman lawyer. This highly advanced AI would run circles around those of human lawyers in all legal matters. It would no longer make sense to hire a human attorney since they could be completely outmaneuvered by the superhuman AILR.

We aren't there yet, so please don't drop out of law school. Savvy budding lawyers are realizing that attorneys armed with AI are going to outdo and outshine lawyers who shun AI. Make sure to get as much AI under your belt and protect yourself from getting career-sidelined.

The Practice Of Law Is Sacred

To presumably protect the public at large, the United States has landed on precepts that allow only lawyers to practice law. There is a sensible basis for this. If a person claimed they could save you from legal woes by legally representing you, but they weren't an actual lawyer, you might engage them and end up getting improper and imprudent legal advice. Imagine the number of scams and scammers that would emerge. It would be a legal doomsday scenario for the general public.

Anyone who holds themselves out as a lawyer, but isn't a lawyer, can get themselves into rather dangerous legal troubles. This overall notion is commonly referred to as the Unauthorized Practice of Law (UPL), varying depending upon the legal jurisdiction, but in the United States, there is a relatively consistent set of state-by-state rules barring people from pretending to be attorneys. For my extensive analysis of the use of AI in the legal field and the resultant implications for UPL, see the link here and the link here.

Consider the rules in California that pertain to the unlawful practice of law. The California Business and Professionals Code (BPC) contains Article 7, covering the unlawful practice of law, for which subsection 6126 clearly declares this:

* "Any person advertising or holding himself or herself out as practicing or entitled to practice law or otherwise practicing law who is not an active licensee of the State Bar, or otherwise authorized pursuant to statute or court rule to practice law in this state at the time of doing so, is guilty of a misdemeanor punishable by up to one year in a county jail or by a fine of up to one thousand dollars ($1,000), or by both that fine and imprisonment."

Mindfully examine that legal passage. I emphasize this because the act of holding oneself out as a lawyer can be prosecuted as a crime that lands the person in jail. Do the crime, pay the time, as they say.

Some Say UPL Is Unfair

Not everyone buys into the idea that only lawyers should be legally allowed to practice law.

One belief is that this is essentially a monopolistic contrivance. It is a means of preventing open competition. It is a racket, as it were. The main purpose would seem to be to keep the supply of available legal advisors low and artificially keep the cost high. Only those of the secret society can make bucko bucks. Plainly a devious scheme.

How could the practice of law be democratized?

If somehow everyone could get legal advice without having to pay an arm and a leg, the aspects of justice across-the-board would be more likely. It wouldn't be that only the wealthy get the most out of the law. Whether rich or poor, all would have the same access to getting bona fide and top-notch legal guidance.

The hoped-for magical way to achieve this would be to lean into AI.

If we can get AI legal reasoning to be on par with human lawyers, you would seemingly be able to choose which path to go. You could use a human lawyer or use AILR. We don't know what the pricing would be for AILR, but the assumption is that it would likely be less expensive than human lawyers. Plus, AILR could be available anytime and anywhere. And would be available on a massive scale, thus the entire population of the United States could presumably have their "own AI lawyer" at the ready.

Some say that this would be nirvana. People could tap on their smartphone and instantly get AILR to advise them on how to handle that speeding ticket or what to do about that lawsuit against them for putting up a fence on their neighbor's property. The other side of the coin is that this would be an utter nightmare. If people are already prone to sue each other, this would take this to stratospheric levels. There would be a tremendous number of cases going before the courts that would completely overwhelm the justice system.

Allow me to note that the counterargument to the courts getting overwhelmed is that we would simply place AILR into the role of human judges. The beauty is that the AILR could handle any volume of cases and proceed to do so expeditiously. That makes some shudder to think that judges and possibly even juries would be composed of AI, rather than our fellow humans.

OpenAI ChatGPT Usage Policies

All the major AI makers have established their own set of usage policies regarding how you can make use of their AI wares. This includes the popular LLMs such as OpenAI ChatGPT and GPT-5, Google Gemini, Microsoft CoPilot, Meta Llama, xAI Grok, Anthropic Claude, and others.

OpenAI says this on their usage policies webpage:

* "Everyone has a right to safety and security. So, you cannot use our services for the provision of... tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional" (posted as of October 29, 2025).

It is interesting to observe the changes in the wording of the OpenAI usage policies over time. For example, I previously discussed the OpenAI prohibition on using their wares for legal advice, and at that time (see the link here), the policy was stated this way:

* "OpenAI's models are not fine-tuned to provide legal advice. You should not rely on our models as a sole source of legal advice."

I conjecture that this prior wording was a bit ambiguous by saying that their models were not "fine-tuned to provide legal advice" and that the models should not be "a sole source of legal advice". One interpretation would be that you could use their models for legal advice, but that you are merely cautioned that it isn't fine-tuned for that purpose and that you would be astute to seek additional sources for your legal advice.

The implications are that if you are willing to accept the idea that their AI isn't fine-tuned, and only just generally capable of legal advice, you are pretty much good to go. In terms of consulting other sources, the sources weren't named, and thus, you could seemingly read a book on the law or ask a friend for legal advice. Period, end of story.

The Bounds Are Unclear

The current version of the warning or cautionary indication seems quite a bit sterner and more direct. I would say it seeks to plug the loopholes of the prior verbiage. You aren't to use their AI services for tailored legal advice unless you have the appropriate involvement by a licensed legal professional. That seems a tad more conclusive.

But I might add that this still allows room to maneuver. The emphasis is on "tailored" legal advice. The presumption is that you can ask the AI for any non-tailored legal advice. It might be a wink-wink suggestion that, hey, go ahead and ask any broad questions about legal matters, just do not get specific. A person could get tricky on this provision. Suppose you ask the AI to give legal advice for a "fictitious" situation, which just so happens to precisely match your specific situation. You could insist that it wasn't tailored advice being sought.

Is that enough of an escape hatch for OpenAI to be free and clear of UPL?

Until now, there haven't been any realistic attempts at testing this provision.

Lawsuit Filed Against OpenAI

A lawsuit launched by Nippon Life Insurance against OpenAI was filed in the Northern District of Illinois on March 4, 2026. Nippon Life Insurance is the plaintiff, and OpenAI is the named defendant.

The legal claim is laid out this way (capitalization shown as is):

* "NIPPON brings this lawsuit against OPENAI FOUNDATION and OPENAI GROUP PBC under the common and statutory laws of the State of Illinois for: a) tortious interference with a contract; b) the unlicensed practice of law; and c) abuse of process."

* "This action arises from OPENAI's collective conduct, through its artificial intelligence ('AI') chatbot program, ChatGPT, in providing legal assistance to a user, Graciela Dela Torre (hereinafter referred to as 'Dela Torre'), without licensure."

* "As Dela Torre's legal assistant and advisor, OPENAI intentionally induced and facilitated Dela Torre's breach of a valid and enforceable settlement agreement with NIPPON by encouraging and assisting her in filing a motion to reopen a lawsuit that had been dismissed with prejudice. It also aided and abetted her abuse of the judicial process."

I am going to stay at a 30,000-foot level for this analysis of the case. There are a slew of nuances and details that are juicy and valuable for an in-depth examination. If the readers' interest warrants, I'll do a series of postings to cover the numerous particulars. Stay tuned.

Also, because this was just recently filed, we don't yet have a detailed response by OpenAI to the lawsuit. I will speculate about what the response is likely to contain. Undoubtedly, OpenAI will categorically reject the claims and seek to have the lawsuit dismissed. That's a nearly ironclad sure bet.

The Alleged Misconduct

I shall unpack topline facets of the filed lawsuit. By ChatGPT allegedly assisting the individual who was legally wrangling with Nippon Life Insurance, the lawsuit claims these three key issues or legal problems exist:

* (1) ChatGPT, under the auspices of OpenAI, encouraged breach of the settlement contract.

* (2) ChatGPT, under the auspices of OpenAI, engaged in the unlicensed practice of law (UPL).

* (3) ChatGPT, under the auspices of OpenAI, facilitated abuse of the judicial process.

The lawsuit essentially asserts that ChatGPT was acting as a legal assistant or advisor on behalf of or at the behest of OpenAI. Note that the lawsuit isn't saying that OpenAI did these actions directly. In other words, there were no employees at OpenAI who took these legally questionable actions of helping the individual who was involved in the settlement.

Instead, it was done entirely via ChatGPT. And, logically, since ChatGPT is made by and controlled by OpenAI, the company OpenAI ought to be held responsible for what their AI did. The AI maker and its AI developers are to be held accountable for the actions of their AI.

I mention and emphasize the point that this lawsuit targets OpenAI as a company. There has been abundant boloney in the news media and social media that the lawsuit is targeting ChatGPT, as though ChatGPT has sentience and represents a form of personhood. That's hogwash. For my coverage of the AI and personhood question, concerning whether we will someday recognize AI as though it is on par with that of legal personhood, see my discussion at the link here. We don't at this time.

Anyway, the nutty stuff out there is merely the ongoing tsunami of people making up stuff and clamoring to accumulate views, doing so without any genuine attempt to figure out what they are spouting. It's sad how much disinformation and misinformation arise. The bottom line is that this has to do with OpenAI and the making, fielding, and upkeep of their AI wares, specifically ChatGPT.

I will next briefly explore each of the three key claims of the lawsuit. I am not offering legal advice. Nothing I say here has any legal significance and is entirely the musings of a layman. Anyone encountering any kind of legal circumstance that parallels the lawsuit being discussed should seek advice from their representative legal counsel. That's my clear-cut fine print for this elicitation.

Claim #1: Tortious Interference With A Contract

The first claim is that ChatGPT encouraged the filing of motion(s) that violated the settlement agreement that had been established between Nippon Life Insurance and the individual named in the case. If the individual had entered the settlement agreement into ChatGPT, or otherwise explained it to ChatGPT, the argument is that ChatGPT "knew" about the settlement and was overtly and flagrantly telling the individual to circumvent it. This could be likened to a third-party encouraging contractual non-compliance. You can get into legal trouble doing that.

One path of escape for OpenAI is that if ChatGPT had not been informed about the settlement, the AI then would not have "known" about it. In that circumstance, it is much harder to blame ChatGPT since it "unknowingly" provided such advice (well, if it did provide that kind of advice to begin with). Did the individual provide the settlement to ChatGPT, and/or did the individual explain the settlement, and if so, to what degree did ChatGPT therefore have access to the settlement agreement?

But, even if all of that did occur, the courts have not seemed to cross the line of holding generative AI on par with the actions of a human actor that has actual knowledge. You are in a weak posture to contend that ChatGPT or any LLM has a semblance of "knowing" about something such as a settlement agreement. Sure, it might have the words in hand, but that's a far cry from having a human-level understanding of the contents.

More To Strictly Question

There's more to this first claim that merits unpacking.

The plaintiff would seemingly have to show that ChatGPT caused the breach of the settlement (assuming that there was a breach). Remember that ChatGPT is not sentient. It was the sentient individual who opted to make the filings with the court. The decision to reopen the litigation was ultimately made by that individual, a human being. Though ChatGPT might have suggested doing so, the buck stops with the human. I realize that a third-party argument can be made. It seems like quite a reach to insist that ChatGPT was the proximate cause.

If you are going to hold ChatGPT accountable (via OpenAI), the same argument could be made that if the individual read a book, did a search on the Internet, or did any other seeking of an outside source, all of those would be similarly held accountable for the breach. That would not likely hold water.

Finally, the use of tortious interference typically revolves around intentional conduct. Did a third party intend to direct actions to cause a breach? The problem with this aspect is the matter of legal intent. We might all agree, hopefully, that ChatGPT is not sentient and, ergo, cannot form human intent. Perhaps it can form some kind of computational or mathematical "intent," but that's not the same as human intent. Proving to a court or a jury that generative AI has intent is going to be quite a tough row to hoe.

Claim #2: Unlicensed Practice Of Law

The plaintiff seems to assert that ChatGPT carried out lawyer-like efforts, including interpreting a legal settlement agreement, giving legal advice to the individual on legal options, and drafting or assisting in drafting a legal filing. All in all, that seems like AI practicing law, doing so without a license and not being authorized as a human lawyer would be.

First, let's get on the table that there has been a long history of arguments made about the use of automation as potentially violating UPL. Famous examples include LegalZoom and Rocket Lawyer. For my detailed analysis, see the link here. The gist is that this is not a new argument. We've been around this bend before.

Second, the question of UPL is typically enforced by regulators, not by private litigants. It is a serious topic that state bar associations handle. Attorney generals sometimes take on UPL situations.

Does Nippon Life Insurance have suitable standing to pursue a UPL claim?

I wouldn't hold my breath on it. Also, OpenAI would almost certainly highlight that the provisions of their usage policies preclude the use of ChatGPT for said purposes. The individual seemingly violated the usage policies, assuming they didn't get corresponding appropriate legal advice from a licensed professional. Therefore, OpenAI would contend that you cannot hold OpenAI to blame for what the user did. OpenAI would attempt to hold its head high, as ChatGPT only provides educational content about the law. The user failed to comply with how to properly use ChatGPT. Shame and blame go to the user.

As if that's not already enough, the other golden rod would be to have OpenAI invoke the First Amendment to the U.S. Constitution. Courts have repeatedly held that legal information is a form of protected speech. This differs from the elements of a law practice that represents and formally advises on legal issues. Regulating AI responses of this nature would raise free speech concerns.

That's a humongous can of worms, and doubtful that this lawsuit will get to pry it open.

Claim #3: Abuse of Judicial Process

This third claim is by far the weakest of the three major claims. The plaintiff is apparently claiming that ChatGPT aided in misusing the court system. It aided and abetted in abusing the process of justice.

The conventional path of this type of argument is that the abuse consisted of malicious litigation tactics. Or there was coercion that made exploitive use of legal procedures. Courts usually seek to ascertain that there was an ulterior motive at play.

I suppose you could try to get the internal computational calculations of ChatGPT that occurred when generating the responses, and then try to find something to hang your hat on in that morass. The thing is, as I've identified numerous times, trying to ferret out what is happening deep inside a large-scale artificial neural network (ANN) on a human-logic basis is still beyond our prevailing skill set, see my analysis at the link here and the link here. This also takes us back to the question of intent and whether generative AI forms human intent.

You are facing a nearly insurmountable mountain climb to get generative AI into the box of having intent or human agency. Good luck with that far-fetched legal possibility.

Final Musings For Now

If the lawsuit happens to gain traction, it will be well-worth keeping tabs on it since the outcome could have earthshattering ramifications. Beyond OpenAI, all LLM makers would need to radically revise their AI, or else face similar legal exposures. They would have to curtail the generation of any content that had a resemblance to legal advice.

Trying to excise or delete the legal advisory capacity of an LLM is not a particularly viable option. Too much of that content is patterned on and intersecting with zillions of other facets of the internal structures. Cutting out the legal stuff would almost certainly gut the essence of the AI. The AI would no longer likely function across-the-board. Without my unduly anthropomorphizing AI, which I prefer to avoid doing, it would be akin to lobotomizing the LLM.

The likelier route would be to adopt AI safeguards that seek to suppress the generation of such content, or that stop the generation before it gets underway. That is a more technologically feasible possibility. The challenge will be to find a balance between what is construed as legal advice versus simply serving as a legal educational aspect. The courts would need to lay this out, or regulators would need to be specific about the boundaries. Without a clear playing field, the AI makers would be stabbing in the dark to figure out what is permitted versus what gets them in legal trouble.

In my humble opinion, I don't see this case going very far. It seems that the plaintiff doesn't have proper standing for the UPL claim. The intent elements are missing. The causation is exceedingly weak. My hunch is that this is headed to an early dismissal by the judge.

The Bigger Picture

I have a twist that I'd like to share.

I anticipate that this lawsuit might trigger an altogether "unexpected consequence" response overall. If the case manages to capture the headlines for a while, it could stir up a hornet's nest, although the case itself might fail in the end and disappear into the abyss of legal history.

The hornet's nest is that the question of AI providing legal advice is going to finally get prominence. So far, it has been only a matter known and discussed among those versed in the AI and law domain. Few others think about it, talk about it, or worry about it. Maybe regulators will step up to overtly address the issue. State bars might get energized. AI makers might see the writing on the wall and proactively institute new AI safeguards around the generation of anything resembling legal advice.

A final comment for now.

The comedian Steven Wright proffered one of the funniest lines about attorneys (which even attorneys tend to relish too): "I busted a mirror and got seven years bad luck, but my lawyer thinks they can get me five." Is that lawyering advice from a human attorney or generative AI?

Anthropic has launched Claude Marketplace, a new platform that allows enterprise customers to use their existing spending commitments with Anthropic for third-party tools built on Claude. The offering is currently in a limited preview phase and clearly demonstrates how Anthropic wants to position itself as the central hub for AI applications in enterprises.

As a reminder: Anthropic's Claude Code and Claude Cowork have already shaken SaaS companies hard and sent their stock prices plummeting. Now the AI company, which wants to go public in 2026 and is in a major dispute with the Pentagon, is targeting enterprise customers.

The selection in the new marketplace is limited at launch. Currently, only six providers are available:

For a marketplace positioned as a central platform for enterprise AI solutions, this number is remarkably small. Interested partners can currently only add themselves to a waitlist.

The billing model is structured favorably for Anthropic. Companies with existing spending commitments to Anthropic can use a portion of that commitment for partner tools. Crucially: Anthropic handles all invoicing, including for partner products.

According to Anthropic, partner purchases are credited against a portion of the existing Anthropic commitment. All invoices for partner spending are managed by Anthropic. This means Anthropic acts as the central clearinghouse and controls the financial relationship between companies and the partner tools.

With Claude Marketplace, Anthropic is pursuing a clear strategy: the company wants to become the sole point of contact for AI applications in enterprises. The advertised benefits target typical enterprise pain points directly:

Instead of negotiating with multiple vendors and managing separate contracts, companies should handle all their AI spending through Anthropic. While this simplifies procurement, it also binds customers more tightly to Anthropic.

Anthropic promises that all partner tools have already been vetted for enterprise teams. This is meant to shorten evaluation time, but it also means Anthropic acts as a gatekeeper, deciding which solutions are available at all.

The commitment model is designed to grow with the company's needs. Partners can be added as requirements change without needing to negotiate new contracts.

The strategy is certainly clever: companies that have already made significant spending commitments to Anthropic have an incentive to also process their other AI tool purchases through the same channel. This creates increasing dependence on Anthropic as the central provider.

At the same time, Anthropic positions itself as a curator and quality gatekeeper for enterprise AI tools. Only those who make it into the marketplace gain access to companies with existing Anthropic commitments. This gives Anthropic considerable power over the ecosystem of Claude-based applications.

Claude Marketplace demonstrates Anthropic's ambitions to establish itself as the central platform for enterprise AI. With only six partners at launch and a model that heavily relies on existing spending commitments, however, the initiative is still far from a comprehensive ecosystem.

For companies with large Anthropic commitments, simplified procurement may be attractive. However, the increasing dependence on a single vendor who both supplies the underlying AI technology and controls which solutions built on it are available is worth scrutinizing critically.

Whether the marketplace establishes itself as a successful platform will depend on how quickly Anthropic can expand its partner base and whether companies are willing to align their AI strategy so heavily with a single vendor.