AI moves fast. We only write what’s worth reading.

Claim Check archive

One circulating AI claim, checked against its primary source. Every check, newest first.

32 checks published · newest first

Claim Check #032 · Models

Alibaba's Circulating 52% Download Claim Has No Traced Calculation. A Separate Top-300 Count Puts Qwen at 50.9%.

“Alibaba models account for 52% of downloads among the most popular chat models on Hugging Face.”

A 52% download share for Alibaba is circulating today, attributed to Yahoo Finance data through aggregators. ALFW could not trace it to an underlying calculation.

Verdict: The circulating 52% claim could not be traced to its underlying calculation, so it is not independently verified. It is, however, broadly consistent with separate, narrower counts: VIDRAFT's defined Top-300 table puts the Qwen family at 50.9%, credited by architecture-based classification, on a counter that cannot…

Checked against: Strong Tower AI

Claim Check #031 · Policy

AI Was the Most-Cited Risk in the Fed's Survey. The Survey Has 23 People in It.

“AI is now the top risk to the US financial system.”

In the Federal Reserve's November Financial Stability Report, AI was the risk its market contacts cited most often, at 87%, up from 50% in the spring.

Verdict: The 87% headline is sentiment from 23 people, gathered in September and October, and the Fed itself says the box is not its view. Quote it as a mood reading. The durable part of this report is Box 2.2: more than $680 billion issued, more than $1.1 trillion outstanding, about 31% of nonfinancial investment-grade…

Checked against: Federal Reserve Board

Claim Check #030 · Tools

Claude Meetings Is Now in Beta. It Joins One Kind of Call.

“Claude stays out of calls with outside guests by default.”

Anthropic published "Introducing Claude Meetings, now in beta" on November 12, 2026. Claude joins a video call as a visible participant, takes the notes, answers questions in writing and builds docs and whiteboards…

Verdict: The useful part of this launch is that the work starts during the call instead of after it. The limiting part is how much has to line up first: the right plan, the right meeting app, the right retention terms, and a guest setting that Anthropic's own launch post and Help Center describe differently for Team plans.

Checked against: Anthropic

Claim Check #029 · Chips

Jensen Huang Has Resigned. Nvidia's Announcement Gives No Reason.

“Nvidia fired Jensen Huang over the investigation.”

Nvidia's board accepted Jensen Huang's resignation as president, chief executive and director on November 11, effective immediately, the company announced on November 12. CFO Colette Kress serves as interim CEO.

Verdict: The documents confirm a leadership change and explain none of it. A board that announced an investigation on Sunday accepted its chief executive's resignation on Wednesday and thanked him on Thursday, without saying whether the inquiry continues or what it found.

Checked against: Nvidia Corporation

Claim Check #028 · Models

DeepMind Reports 96% Success on One Mercedes Parts-Kitting Task

“Gemini Robotics 2.5 resists 99.5% of jailbreaks.”

DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, reporting a 96% success rate for an Apollo 2 parts-kitting proof of concept on active Mercedes-Benz production lines in Hungary and Germany.

Verdict: The factory pilot is concrete evidence of performance on a defined task type, not a field-reliability figure for general factory work. Its scoring denominator and failure handling remain undisclosed. Keep the broader task results attributed to DeepMind and the jailbreak score attached to Argon.

Checked against: Google DeepMind

Claim Check #027 · Models

GR00T 2: NVIDIA Reports 95% After Five Demos, on Tasks Selected for Five-Demo Learnability

“NVIDIA's GR00T 2 achieves 95% success on 50 new robot tasks after just five demonstrations.”

NVIDIA introduced GR00T 2 at the Conference on Robot Learning in Austin on November 9: a world-action model that predicts future video and robot actions together, pretrained on 402,730 hours of human egocentric video…

Verdict: The reported transfer and adaptation results merit further evaluation. The 95% figure describes a deliberately selected task set, a limitation the paper itself acknowledges, while the highest assembly rates evaluate an agent-supervised system. Neither establishes field reliability for an unsupervised robot worker.

Checked against: NVIDIA Blog

Claim Check #026 · Chips

Nvidia's Board Opens a Probe Into Reports Its CEO Was Targeted

“Nvidia is investigating whether its CEO was spied on.”

Nvidia's board announced on Sunday, 8 November, that it is opening an independent investigation into recent reporting that founder and CEO Jensen Huang "has been the target of a foreign intelligence effort".

Verdict: Read the board's six sentences for what they are: a governance response to reporting, not a finding about espionage. The statement casts Huang as a target, names no alleged operative, no customer and no country, and states the board's current position — no reason to believe national security or Nvidia secrets were…

Checked against: NVIDIA Board of Directors

Claim Check #025 · Models

Open DeepSeek Scores 72.8% on Cyber Test. Gated GPT Hits 80.8%.

“AI hacking ability sits safely behind vetted access.”

On Vals AI's two simulated corporate networks, restricted GPT-6 Sol leads open-weight DeepSeek V4.1 Flash by eight percentage points.

Verdict: Vals' CyberOffense results show how close a downloadable open-weight model can come to a restricted frontier model on two simulated intrusion ranges. DeepSeek V4.1 Flash scored 72.8% against GPT-6 Sol Daybreak's 80.8%, with similar three-run averages on Nova but a larger deficit on Lighthouse.

Checked against: Vals AI

Claim Check #024 · Models

Claude's 'Dreaming' Repairs Memory. Its Wiki Alone Scores 48%.

“Claude can dream now.”

Anthropic's agent-memory experiments show selective notes and offline repair can help Claude. But its largest wiki test still exposes accuracy limits and unquantified maintenance costs.

Verdict: “Dreaming” is a sleep metaphor for memory maintenance: rereading transcripts, correcting notes and rebuilding representations between sessions. Anthropic shows that this can work, but its evidence is narrower than a general solution to agent memory.

Checked against: Anthropic

Claim Check #023 · Models

Same Model, Different Scores: 1 in 4 Endpoints Below Reference

“The same model scores the same on every API.”

Artificial Analysis tested more than 100 hosted endpoints against self-hosted reference deployments of the same open weights.

Verdict: A model name is an incomplete API specification. About one in four hosted endpoints scored below their self-hosted reference, according to AlphaSignal's summary of the index.

Checked against: Artificial Analysis

Claim Check #022 · Headlines

METR's Anthropic Audit Finds Acceleration—and Redactions

“AI is improving itself twice as fast.”

METR estimates Anthropic's capability growth at 1.9x its previous rate (95% CI: 1.5–2.8x; alternative modeling: 1.1–2.9x). Its broader assessment puts the acceleration around 2x, but allows for 1.3x or 3x.

Verdict: The doubling estimate may well be right, and METR has done the work to make it credible: named methods, a confidence interval, and a list of ways it could be wrong.

Checked against: METR

Claim Check #021 · Policy

Ohio Voters Restricted Data Centers in 14 Municipalities. The Line Is 25 MW.

“Ohio voters rejected data centers.”

Ohio voters decided 21 data center measures across 18 municipalities on November 3, 2026. Unofficial results show at least 14 municipalities approved bans or restrictions, many targeting facilities above 25 megawatts.

Verdict: The headline count, 14 municipalities, understates what happened and overstates it at the same time. Understates it, because the 25 MW line plus Grove City's consent model gives opponents two tested instruments, one of which a developer cannot negotiate around at all.

Checked against: Reason

Claim Check #020 · Policy

US Businesses With 250+ Employees Report 40.7% AI Use. The Question Changed.

“AI adoption at large U.S. firms surged from 12% to 40.7% in a year.”

Simon Willison's analysis of US Census Bureau data puts reported AI use among businesses with 250 or more employees at 40.7% in October 2026, up from roughly 12% in September 2025.

Verdict: The 40.7% figure is real as a calculation from Census data, but it describes businesses with 250 or more employees under a broader question than the one used in 2025. Census changed that question after testing showed that the original wording missed important forms of business AI use.

Checked against: Simon Willison

Claim Check #019 · Headlines

42AI's 98% Accuracy Measures a Range, Not a Prediction

“42AI's system predicts trading with 98% accuracy.”

Israeli startup 42AI came out of stealth on November 2, 2026, announcing $62.5 million in seed funding and a behavioral world model designed to predict individual decisions.

Verdict: The most useful sentence in 42AI's launch is the CEO's own correction: 98% is a range the trades landed in, with the range's width undisclosed, and the directional score is an F1 of 0.8 to 0.85 that no outside party has tested.

Checked against: TechCrunch

Claim Check #018 · Tools

OpenAI Documents ChatGPT Card and Device Limits. The Caps Aren't Published.

“ChatGPT now caps paid accounts per device and subscriptions per card.”

OpenAI's Help Center now documents separate payment-card and device restrictions for ChatGPT subscriptions. The page lists the plans counted toward the device limit, explains the consequences of exceeding it, and…

Verdict: The most consequential part of this policy is not the blocked checkout. It is the possibility of losing access to a paid account on a device after the subscription has already been purchased. OpenAI documents the restrictions but leaves both numerical thresholds and the effective date unpublished.

Checked against: OpenAI Help Center

Claim Check #017 · Headlines

Meta Says Millions Use Muse. Merchants See Little Agent Business.

“Meta's Muse is becoming a major force in shopping, with millions making purchases averaging over $200.”

WIRED examined what executives at roughly 50 companies have said about AI shopping agents. Most described little immediate business, despite Meta's claim of more than 10 million Muse activations.

Verdict: Meta has demonstrated interest in Muse, not yet the volume of completed shopping transactions that would establish it as a major commerce channel. WIRED's reporting offers a useful check from the merchant side: DoorDash has seen only several hundred orders from third-party agents combined, while executives across…

Checked against: WIRED

Claim Check #016 · Policy

WIRED: An Accenture Contractor Ran a Personal Claude Agent on a Pentagon Network for Two Months

“An AI agent breached the Pentagon.”

An Accenture contractor reportedly gave a personal Claude Code agent their own login to a secure Defense Department system, allowing it to operate with administrative access for about two months, WIRED reports.

Verdict: Nothing in WIRED's reporting required the agent to defeat a security system. According to the sources, a person with legitimate access handed it administrative credentials, allowing a personal AI agent to operate inside a production environment.

Checked against: WIRED

Claim Check #015 · Policy

Dartmouth's Provost Resigned After an AI Review. The Logs Mattered More Than Detectors.

“A detector rated the Dartmouth provost's writing up to 96% AI-generated.”

Dartmouth Provost Santiago Schnell announced his resignation on Wednesday, October 28, 2026, and is leaving the university.

Verdict: The important fact is not simply that a provost used AI in his writing. Dartmouth's committee found two clear publication-policy violations using dated publisher rules, communications and the university's enterprise Claude logs, without relying on a detector.

Checked against: Dartmouth College

Claim Check #014 · Chips

Malaysia Seized $12.1M in AI Servers. The Chips and Route Remain Unnamed.

“Malaysia stopped a chip-smuggling route at KLIA.”

Malaysian Customs disclosed on October 28 that it had seized 30 servers containing advanced AI chips, worth nearly RM49.5 million, at Kuala Lumpur International Airport on October 15.

Verdict: Malaysia's seizure is evidence of active strategic-trade enforcement, not proof that a particular chipmaker or destination country was involved. Customs says it seized 30 servers containing advanced AI chips, found missing authorization and a false declaration, and linked the shipment to four other cases this year.

Checked against: Reuters

Claim Check #013 · Models

Baidu's Unlimited OCR 2 Parses 100-Page Documents. Its 77.1 Score Is Still Only Baidu's.

“Baidu's new OCR model scores 77.1 on hundred-page documents — and its own predecessor scored 0.0.”

Baidu released Unlimited OCR 2 on October 27, an open-weight model that parses documents of a hundred pages or more in a single request instead of page by page.

Verdict: The engineering is plausible, and Baidu did the thing most vendors skip: it published the benchmark's data and evaluation protocol the day before the model, so the headline claim can be checked by anyone willing to run it.

Checked against: Baidu

Claim Check #012 · Headlines

AI Tenants Now Want $5 Million for Every Late Day. In Spring It Was $1 Million.

“AI data center tenants now collect $5 million a day in late fees.”

Tenants negotiating new AI data center leases are demanding late-delivery penalties of more than $5 million a day, The Information reported on Sunday, up from about $1 million this spring.

Verdict: The $5 million figure is a negotiating demand, not a verified market rate or a documented payment. It nevertheless signals how valuable on-time delivery has become to AI infrastructure tenants.

Checked against: The Information

Claim Check #011 · Chips

Musk Says SpaceX Closed $40 Billion in GPU Debt. The Terms Were Not Published.

“SpaceX has closed $40 billion in GPU debt.”

On October 25, in a reply on X rather than a corporate announcement, Elon Musk said SpaceX had closed $40 billion in debt financing for GPU purchases.

Verdict: "Closed" is the consequential word in Musk's announcement, and its meaning remains unverified outside SpaceX. His October 25 replies move the public account from reported $40 billion financing talks to a chief executive's claim that the debt has been secured.

Checked against: Elon Musk

Claim Check #010 · Policy

Google's Gmail Ads Promise Now Comes With an Exception

“Google doesn't use your Gmail to target ads.”

Google's privacy policy, effective October 16, 2026, keeps its sentence promising no personalized ads based on your Drive, Gmail or Photos content, and adds a condition: unless you choose to make that data available to…

Verdict: The promise was the product. Since 2017, "we don't use your email content for ads" was the sentence Google pointed to whenever the Gmail question came up. The new wording does not break that promise for people who never connect their apps, but it redefines the consent: a setting sold as "better answers" now doubles as…

Checked against: Google

Claim Check #009 · Headlines

Microsoft Says Its Diagnostic AI Is Nearly Four Times as Accurate as Doctors. The Comparison Needs Context.

“Microsoft's AI diagnoses up to four times more accurately than human doctors.”

A peer-reviewed study in Science reports that Microsoft's diagnostic system MAI-DxO reached the correct diagnosis in 83.2% of 304 hard NEJM cases.

Verdict: The Science paper is peer-reviewed, its method is published, and its benchmark code is open under the MIT License, so the 83.2% is inspectable in a way vendor claims usually are not. But the headline comparison still needs a matched denominator: the system's score covers 304 cases, the physicians' averages cover 56.

Checked against: Nori et al.

Claim Check #008 · Headlines

Meta's AI Spending Just Outran Its Cash Flow

“Meta's free cash flow went negative because of its $9.9 billion legal settlement.”

Meta reported third-quarter results on Wednesday 21 October 2026 with revenue up 26%. In the same release, capital spending of $34.91 billion exceeded the $32.79 billion of cash its operations generated, and free cash…

Verdict: Meta's negative free cash flow in Q3 is a capital-spending story, not evidence that its core business stopped generating cash. Operating cash flow rose 9% year over year, but capital expenditures rose about 80% and exceeded operating cash generation by $2.12 billion.

Checked against: Meta Platforms

Claim Check #007 · Chips

NVIDIA Says an AI Factory Can Deploy in 14 Days. The Clock Starts After Site Prep.

“AI factories now deploy in 14 days, down from six months.”

At GTC Berlin on Wednesday, 21 October 2026, NVIDIA introduced DSX POD, a reference design for complete Vera Rubin AI factories built by partners inside modular containers.

Verdict: DSX POD targets a real source of delay: integrating and validating AI infrastructure after it reaches the deployment site. NVIDIA says factory-built Vera Rubin systems can cut onsite deployment from its six-month baseline to as little as 14 days, but only where power, cooling, networking and supporting infrastructure…

Checked against: NVIDIA press release

Claim Check #006 · Policy

WhatsApp's Help Pages Told Its AI Not to Mention the Regulator

“WhatsApp tells users they can complain to a data protection regulator, and its support help will point them there.”

Redact says it found 102 hidden instruction blocks across 80 WhatsApp help pages, invisible to readers. Two of them reportedly told WhatsApp's support AI not to recommend a data protection regulator, and not to cite…

Verdict: A support bot declining to give legal advice is ordinary. Instructions telling it not to name a regulator, or not to cite privacy law after a user mentions legal action, are a different matter. Redact says it found those instructions in WhatsApp's help-page HTML, where they were not visible to ordinary readers.

Checked against: Redact

Claim Check #005 · Headlines

One Hacker Used AI Agents to Breach 100+ Company Websites. What Was Taken Is Unclear.

“AI agents hacked 100 companies for $50 each.”

A single operator set AI agents loose against more than 1,200 e-commerce websites and breached more than 100 of them, Reuters reported on Monday, in an investigation built on the attacker's own exposed control panels.

Verdict: The reported campaign is a serious illustration of what one operator can sustain with agents: probing across a list of more than 1,200 targets, with more than 100 reported compromises.

Checked against: Reuters

Claim Check #004 · Tools

OpenCode Says Provider Keys May Be Exposed. 1.4% of Console Users Affected.

“OpenCode's breach was a “large scale distillation attack” that hit 1.4% of Console users.”

OpenCode disclosed unauthorized access to a legacy system on October 17 and warned that third-party API keys stored in some Console workspaces may have been exposed.

Verdict: OpenCode has established an incident serious enough to warrant immediate credential rotation, but it has not publicly established that keys were extracted, used, or involved in model distillation. Its large-scale-distillation description remains a company claim while the investigation continues.

Checked against: OfficeChai

Claim Check #003 · Headlines

Anthropic's Internal Model Says $44 Trillion by 2030. Clark Called It Outlandish Himself.

“AI could expand US economic output to $44 trillion by 2030.”

Anthropic co-founder Jack Clark told IMF and World Bank officials in Bangkok on October 17 that one of the company's internal economic models projects US output reaching $44 trillion by 2030, from just over $30 trillion…

Verdict: The $44 trillion figure is a conditional scenario from one of Anthropic's internal models, not an independently established GDP forecast — and its own author called it outlandish in the same lecture.

Checked against: Reuters

Claim Check #002 · Policy

Claude's Saved Instructions Can Now Be Read by Enterprise Admins Across Organisations

“Your saved Claude instructions are private to you.”

A new beta endpoint in Anthropic's Compliance API returns the exact text each member saved as their personal instructions for Claude.

Verdict: The endpoint is documented enterprise compliance tooling, and the criticism is a design choice, not a secret: a personal, account-level setting now sits inside organisation-level audit access, replicated across every Enterprise organisation a member belongs to, with audit events for administrators but no member…

Checked against: Anthropic

Claim Check #001 · Models

Gemini 4 Argon Is Finally Open. The Low Price Expires in Seven Weeks.

“Gemini 4 Argon costs $2/$10 per million tokens.”

Thirteen days after Google unveiled Gemini 4 Argon and handed it only to approved cybersecurity defenders, the model opened on 13 October to paying API developers, AI Studio, Antigravity and Gemini app subscribers.

Verdict: Argon is a credible frontier option on professional work, first on one independent index and mid-pack on another, and Google's own engineering stories are unusually concrete. None of that changes the commercial fact: the launch price is a seven-week trial window dressed as a price list.

Checked against: Google