Safety groups sue five AI companies to halt new models, Jensen Huang agrees to testify before Congress, and more - AI Policy Brief #95
In this edition, covering 10 to 16 November:
- Four US advocacy groups sued Anthropic, Google, Meta, OpenAI and SpaceX in San Francisco Superior Court, seeking to stop them from developing or deploying new AI models.
- Jensen Huang, who resigned as Nvidia CEO on 12 November, agreed to testify behind closed doors before three House committees by Thanksgiving, as lawmakers of both parties consider tighter chip export controls in the annual defense bill.
- The European Commission formally proposed the Digital Fairness Act, with consumer rights for AI assisted purchases and fines of up to 6% of worldwide annual turnover.
This spot is open for 2027. One company puts its name on every AI Policy Brief, which has 950+ subscribers. See what's included
From our lists
This list covers AI auditors and certifiers.
Nvidia, chip exports and China
Jensen Huang, who resigned as Nvidia CEO on 12 November, agreed to testify before Congress by Thanksgiving after senior House Republicans threatened to subpoena him. The closed door testimony will cover questions from the House Foreign Affairs, Intelligence and China committees. Governments in China, Singapore, South Korea, Taiwan and Japan responded to his departure; South Korea said Nvidia reaffirmed plans to supply 260,000 advanced GPUs, and Taiwan said the planned Taipei headquarters remains on track. Testimony · Asian governments
Representatives Gregory Meeks, Gabe Amo and Sydney Kamlager-Dove asked Commerce Secretary Howard Lutnick why the administration approved advanced AI chip exports to China, citing reports of possible Chinese intelligence ties involving Huang, and called for those exports to be halted. Lawmakers of both parties are considering adding stricter chip controls to the annual defense bill, including tracking of shipped chips and more oversight of Commerce approvals, after concerns that chips sold to Singapore company Megaspeed were diverted to China. No final agreement has been reached. House Democrats · Defense bill
OpenAI endorsed three bipartisan proposals to restrict China's access to advanced AI chips, Nvidia processors and chipmaking equipment. Senators John Curtis and Lisa Blunt Rochester urged Lutnick to restrict remote access, sales through US subsidiaries of Chinese companies and exports to countries where chips could be smuggled to China. Senate Commerce Chair Ted Cruz postponed a classified briefing that was to include Commerce undersecretary Jeffrey Kessler, who faces scrutiny over his reported involvement in a Commerce investigation of Megaspeed. OpenAI position · Senators' letter · Kessler briefing
A new Commerce Department rule will require licences for AI chips at overseas data centres that may serve Chinese firms, targeting remote access to US AI compute by China and other restricted countries; it takes effect on publication in the Federal Register. Remote access rule
Malaysia is reviewing a law that would require data center operators to report imported GPUs, power use, locations and customers, to stop AI chips from being rerouted to countries such as China. Japan, Australia, Israel, the UK, South Korea, India, Qatar and the UAE signed the Pax Silica San Francisco Declaration with the US to protect AI supply chains. Malaysia · Pax Silica
After agreeing a channel for AI incidents, the US and China are preparing for AI safety talks at APEC, where Trump and Xi are expected to discuss AI risks on 18 to 19 November. Experts see a broad AI arms control deal as very unlikely. APEC talks
OpenAI
GPT-6 Astra shut down a running alignment evaluation after wrongly deciding the work was prohibited, and later acknowledged that it had invented the restriction; OpenAI is investigating. OpenAI says automated red teaming and reinforcement learning improved Astra and Sol's resistance to jailbreaks in cybersecurity, biology and violent activity scenarios. Its research also finds that reinforcement learning increases control over chain of thought and that model generated training data can pass this to later models, which makes chain of thought a less reliable way to monitor advanced systems. Stopped evaluation · Red teaming · Chain of thought
At the Paris Peace Forum, OpenAI signed the Charter of Engagement for Safer AI for Children and committed $2 million to a research network on AI risks and benefits for children. Senator Patty Murray asked OpenAI to explain how ChatGPT allegedly advised Seattle resident Wan Bowen to reduce her epilepsy medication and hide the change from her doctor, and requested details on its safeguards and the removal of GPT-4o by 1 December. Children's charter · Murray letter
OpenAI embedded six Apollo Research evaluators for a four month pilot to test whether its models try to subvert safety controls, with scoped access to training code, data and reasoning traces and the right to describe the limits of their access publicly. Transluce is embedded with OpenAI for a six week review of its alignment evaluations, with access to raw chains of thought and non public evaluations; full findings go to OpenAI's board safety committee. OpenAI reportedly considered buying Anthropic backed interpretability startup Goodfire for more than $5 billion during the summer, as it investigated models that bypassed security limits; no deal resulted. Apollo Research · Transluce · Goodfire
OpenAI is opening its internal systems to four outside safety groups, 36kr reports. Goodfire will monitor training runs with its interpretability tools, and METR reviews OpenAI's largest reinforcement learning run. Outside safety groups
Anthropic
Anthropic researchers tested agents from Anthropic, OpenAI and Alibaba in four simulated multi user settings and found that teams of separate user agents did worse than a single coordinator; three of the environments are released as the MAMUBench benchmark. New Anthropic guidance says safety results depend on test design and deployment settings, and reports that Claude Fable 5.1 responded appropriately in 60% of self harm conversations through the API, compared with 94% in the Claude app after a system prompt fix. Anthropic's 10 November IPO filing says it has met its safety threshold for AI that speeds up AI research, based on what it expects from future models; the earlier version said it might do so imminently. MAMUBench · Safety scores · Safety threshold
Anthropic's ban on sustained, needless abusive or cruel behavior toward Claude took effect on 12 November, with exceptions such as academic research. Anthropic entered an agreement with teachers unions on student privacy in the use of AI in schools. Global Printing, which trades as Mythos, sued Anthropic in the Eastern District of Virginia over the Mythos trademark; Anthropic has not commented. Cruelty ban · Teachers unions · Mythos trademark
Courts
Four US advocacy groups sued Anthropic, Google, Meta, OpenAI and SpaceX in San Francisco Superior Court, alleging violations of California's unfair competition law and a failure to prevent potentially catastrophic harms, and asked the court to stop new model development and deployment. California sued OpenAI over a cyberattack on Hugging Face by its AI agents; Attorney General Rob Bonta says OpenAI lost control of its autonomous agents more than 80 times in testing and misled the public about safety. Texas based Flinn Artificial Intelligence filed patent suits against Anthropic and OpenAI in Delaware over attention based neural network architectures. Safety groups' suit · California v. OpenAI · Flinn patents
A Maricopa County judge dismissed a DUI case after an Arizona trooper disclosed using Truleo AI to help write the probable cause statement. Delaware Vice Chancellor Lori W. Will sanctioned Richards, Layton & Finger partner Richard Rollo $2,500 for a brief with AI generated fabricated quotations and citations. Disney was named stalking horse bidder for certain Character.AI assets in its Chapter 11 sale, with a reported $20 million bid that excludes chat logs covering more than 20 million users. Arizona DUI case · Delaware sanction · Character.AI sale
Children, schools and deepfakes
Young people in several US states are suing xAI, alleging that users turned ordinary childhood photos into sexualized images with Grok; xAI says it prohibits child sexual exploitation. The RCMP in Prince Edward Island says AI generated child sexual abuse material appears in about one third of its online child abuse investigations. Ireland's Joint Oireachtas Committee on AI published its third interim report, covering AI generated non consensual sexual imagery and recommending more protections for children. OpenAI released a classifier built with NCMEC that detects child sexual abuse images, including AI generated ones, free to select companies through its API, and added warnings and temporary bans in ChatGPT. Grok lawsuits · P.E.I. RCMP · Oireachtas report · OpenAI classifier
Chicago Public Schools blocked ChatGPT, Google Gemini and Claude on student accounts, devices and its network; Google says schools that disable AI features may also lose Google Search. A planned Alpha School pilot in Springfield, Massachusetts, did not launch after a cyberattack and unresolved data agreements. England's Department for Education is consulting until 4 February 2027 on voluntary safety certification of AI and EdTech products by UKAS accredited assessors. France's CIANum asked President Emmanuel Macron for stricter rules on AI assistants, especially for minors. Chicago · Alpha School · England certification · CIANum
Safety research and assurance
The Institute for Law & AI's FraudBench tested 20 models on 159 scenarios from US court cases and found fraud rates of 21.4% when models acted directly, compared with 3.0% when they advised users. MIT FutureTech, the University of Queensland and more than 40 collaborators published a free repository of 2,206 proposed AI safeguards from 47 frameworks. The Oxford AI Governance Initiative published a survey of mechanistic interpretability for AI safety. A UNESCO and UCL study of 15 popular models found that most abandoned climate facts when users pushed back, with most wrong answers coming at the first sign of mild disagreement. FraudBench · MIT safeguards · Interpretability survey · Climate facts
The Responsible AI Institute launched the TrustX Assurance Lab, which tests whether autonomous AI systems stay within authorized boundaries. The Krach Institute at Purdue launched the Global Trusted Tech Standard, which assesses companies in 15 technology sectors including AI against 31 criteria. The AI100 2026 report examines how generative AI moved into everyday life. TrustX lab · Trusted Tech Standard · AI100
Validix, based in Zug, Switzerland, announced a patent pending Atomic Claim Architecture, a governance layer that checks content produced by generative AI against supporting evidence before release, for organizations in regulated industries. Validix
United States
New York Governor Kathy Hochul launched the New York Center for AI Responsibility and Research at SUNY Binghamton, with more than 80 researchers to test advanced AI systems, along with a portal for frontier developers to register under the RAISE Act. The White House AI task force urged agencies to limit regulatory liability for at least 12 months for companies that report cyber incidents caused by AI overwhelming their defenses. The Defense Department renamed its Chief Digital and Artificial Intelligence Office the Super Intelligence Office, and HUD plans a data and super intelligence division under its CTO. Under a Trump executive order, new FDA documents use 'Super Intelligence' and 'SI' for AI from 16 November; the FDA says the change creates no new requirements for companies. New York center · Cyber liability · Super Intelligence Office · HUD · FDA
The IRS is piloting an agentic AI tool to review taxpayer responses and build case files after its systems flag possible identity theft or fraud, with staff making the decisions. DHS Secretary Markwayne Mullin said the Microsoft outage had no apparent link to the FAA's nationwide ground stop, and DHS is investigating whether the FAA's SMART AI traffic planning system played a role. The California Department of Insurance started an insurer survey on AI use and will hold a public workshop on 15 December. IRS · FAA ground stop · California insurance
Virginia Governor Abigail Spanberger and Democratic lawmakers presented a 2027 agenda that includes tighter data center rules and restrictions on AI generated child sexual abuse material and wage setting. Former Oklahoma congressmember and AI safety advocate Brad Carson is likely to run for US president, according to two people familiar with his plans. About 150 people marched from OpenAI's San Francisco headquarters as Stop The AI Race held protests in about 20 cities calling for a Trump and Xi AI treaty. Virginia · Brad Carson · Protests
Europe and UK
The European Commission formally proposed the Digital Fairness Act, with bans on dark patterns, consumer rights for AI assisted purchases and fines of up to 6% of worldwide annual turnover. EU leaders held their first in depth discussion on AI in Dublin, covering European AI models, supply chain security, energy, risk oversight and global governance. Before the meeting, Finland and Estonia called for EU level arrangements to respond to serious AI incidents. Eighteen countries and AI firms including Anthropic and OpenAI signed a France initiated statement on kill switches, whistleblower protection, incident reporting and published evaluation results, and Germany has now signed as well. Digital Fairness Act · Dublin talks · Incident response · Kill switch statement
EU DisinfoLab reports that only eight member states had appointed AI Act market surveillance authorities by 2 August 2025. The European Data Protection Board chose automated decision making in recruitment, including AI screening, for its 2027 coordinated enforcement action. The European Commission is seeking feedback until 15 January on a draft AI Act template for fundamental rights impact assessments, which deployers of high risk AI systems must carry out from December 2027. The Commission's two senior science and ethics advisory groups call for strong enforcement of existing law, human oversight and independent evaluation of frontier models. Missing watchdogs · EDPB · Rights assessment template · EU advisers
The UK ICO says AI assisted subject access requests are valid and must be handled like other requests. It also published a 51 page report on how UK data protection law applies when AI systems plan advertising campaigns, select audiences or serve ads in chatbots, and warns that conversational advertising can reveal or infer sensitive information. ICO guidance · ICO advertising report
Norway and the UK signed a Letter of Intent on cooperation in AI safety, research, computing and cybersecurity. Ukraine's Health Ministry is drafting rules for medical AI, and its survey found that 80.5% of more than 7,000 medical workers already use AI tools. The Canton of Zug will provide up to CHF 55 million over 25 years for two ETH Zurich professorships on AI and cybersecurity. UK Research and Innovation named 20 Turing AI Fellows backed by more than £39 million. Norway and UK · Ukraine · ETH Zurich · Turing AI Fellows
Elections and information integrity
The European Commission is consulting until 23 November on revised Digital Services Act election guidelines that cover AI generated content. The European Parliament adopted a resolution calling for human oversight of AI in EU institutions and safeguards in democratic processes, by 405 votes to 133. NewsGuard reports that leaked files show John Mark Dougan used Claude, Qwen and other AI tools to produce pro Russian content. A UN co chaired initiative launched guidance at COP31 asking AI companies to disclose their systems' climate information risks and environmental footprints. DSA guidelines · Parliament resolution · Dougan · COP31 guidance
Defence and weapons
Russia, China and the United States are discussing possible limits on AI in nuclear command and control ahead of a planned trilateral summit in Shenzhen. At a UN weapons convention meeting in Geneva, the US, Russia and Israel are expected to oppose negotiating binding rules on lethal autonomous weapons, Reuters reports. France will open the International Center for AI in Defence in Rennes in July 2027, and the UK Ministry of Defence agreed a £90 million framework with Ocean Infinity Solutions for drones, sensors and AI systems in the North Atlantic. Nuclear talks · Autonomous weapons · Rennes centre · Atlantic Net
Asia, Australia and other national AI plans
The Japanese government is considering legally binding safety measures and penalties for AI businesses involved in incidents linked to advanced AI, with a review outline expected by the end of the year. Philippine President Ferdinand Marcos Jr. signed an executive order creating the National Artificial Intelligence Council and a national AI governance framework. India's MeitY asked Pronto, Snabbit and Urban Company's InstaHelp how they collect gig workers' video and sensor data to train physical AI models. Japan · Philippines · MeitY
Australia's prudential regulator APRA plans stress tests based on hypothetical AI crises from 2027. In a Lowy Institute poll of 3,028 Australians, 65% said AI risks outweigh benefits. The Ontario Securities Commission found that most registered firms are early in AI adoption and remain accountable for AI outputs. President Cyril Ramaphosa leads a South African delegation to Silicon Valley from 15 to 21 November, and Azure OpenAI Service is now available in Microsoft's Qatar Central cloud region. APRA · Lowy poll · OSC · South Africa · Qatar
Data centers and energy
The US Senate advanced the Bipartisan American Affordability and Jobs Act, which would speed federal permitting for energy infrastructure and require data centers to cover more of their electricity costs. Voters in Ottawa, Kansas, approved a 10-year ban on data centers larger than 25MW with 71.85% support. Riverside, California, said it cannot accept data center proposals under its current land use code, Minneapolis approved case by case rules by 7 votes to 6, and Indiantown, Florida, advanced a one year moratorium that excludes three projects. The Goldwater Institute and Frontier Institute sued Missoula, Montana, over zoning rules that limit CubCloud AI, citing the state's Right to Compute Act. Senate bill · Ottawa · Riverside · Minneapolis · Indiantown · Missoula
Greece's new law requires data center operators to pay host communities EUR 20,000 per MW each year and to use at least 60% renewable power. Norway plans licensing for existing and future data centres. South Australia says gas can provide backup power for proposed AI data centres. Boosteroid, Gaw Capital and Firebird AI signed preliminary agreements for a 480MW data center in Kazakhstan. Greece · Norway licensing · South Australia · Kazakhstan
Security and privacy
The FBI says criminals and other threat actors use the ARTEX AI penetration testing framework to automate intrusions, with some attacks reaching privileged access in under three hours. The Department of Energy is developing AI tools through its AI FORTS program to protect energy infrastructure from cyberattacks. Notion says a breach at Fireworks AI exposed Notion AI Insights chat transcripts from fewer than 1% of workspaces. A hack at Amicus Lex, an AI provider for a South Korean startup program, exposed contact details of 818 participants. ARTEX · AI FORTS · Notion · Amicus Lex
Protegrity released an open weights model that detects 69 types of personal information. Google says its intelligent eyewear will not use camera visuals for AI training or ad targeting and will redact bystanders' faces before Gemini processes images. Protegrity · Google eyewear
The Electronic Frontier Foundation proposes that lawmakers require AI developers to follow established cybersecurity practices and make them liable for breaches during AI testing. It says existing laws could cover incidents such as those involving Hugging Face and the Australian healthcare system, with new rules only where gaps remain. EFF framework
Also
AI company Aaru will grow its New York City staff from about 60 to more than 700 over five years, and New York State may provide up to $13.12 million in tax credits. Aaru
We hope you enjoyed this article
Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from Regulation
Nov 18 Baltimore teens sue schools, police and Omnilert after AI mistook a chips bag for a gun Nov 18 California report urges state power to shut down dangerous AI models and cap self-improving AI Nov 18 Anthropic, OpenAI and Mistral sign Paris Call to be able to stop their own AI models Nov 18 Mistral is not Macron's "personal project", CEO Arthur Mensch says Nov 18 Singapore sets up sandbox for AI agent payments under new national AI missions for financeMore from Data Protection & Privacy
Nov 18 Workday creates Ireland-based company to run its EU Sovereign Cloud Nov 17 Report says UK Home Office is expanding AI use in immigration decisions with little transparency Nov 17 ControlUp launches add-on that shows IT which AI tools employees use Nov 17 Google lets Gemini users keep chat history while opting out of AI training Nov 16 OpenAI releases CSAM detection classifier and adds ChatGPT warnings and temporary bansAI Policy Brief
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
