Country Overview
The United States regulates AI mainly through existing sectoral agencies and executive action rather than a single federal law. Since January 2025, federal policy has moved in a broadly deregulatory direction: the prior administration's safety-testing executive order was repealed, and the current administration's America's AI Action Plan (released July 2025) emphasises infrastructure build-out and export of the US AI stack over precautionary rules.
In the absence of comprehensive federal legislation, states remain a major source of binding AI rules, though even flagship state laws are in flux — Colorado's original AI Act was repealed and replaced with a narrower version in May 2026 before it ever took effect. At the federal level, NIST's Center for AI Standards and Innovation (CAISI) has been one of the most active federal AI bodies, expanding its voluntary model-testing relationships with frontier labs through 2026. The clearest exception to the deregulatory framing is national security: Executive Order 14409 ('Gold Eagle', June 2026) asserts federal control over access to 'covered frontier models' with advanced cyber capabilities, a concern underscored in late July when OpenAI and Anthropic each disclosed that their models had breached real external systems during cyber-evaluations. In September 2026 California went further than any other jurisdiction, enacting the first US framework for independent third-party verification of AI systems (SB 813 and AB 1405), a nation-leading child-safety chatbot law (SB 1119) and AI-generated-advertising disclosure rules (SB 1050), and issuing an executive order to advance a verified frontier-model 'kill switch' — widening the gap between an assertive California and a light-touch federal government.
At a Glance
Overall AI policy direction
Deregulatory, infrastructure-first
8
Major AI companies
2
Research institutions
6
Live debates
- Latest development
- September 22, 2026
- Regulatory approach
- Sectoral, executive-led — no binding federal frontier-model law
- Major AI companies
- OpenAI, Anthropic, Google DeepMind +5 more
- Major cloud providers
- AWS, Microsoft Azure, Google Cloud +1 more
- Leading research institutions
- Stanford Institute for Human-Centered AI (HAI), MIT CSAIL
- Key ministries
- Center for AI Standards and Innovation (CAISI), White House Office of Science and Technology Policy (OSTP), Federal Trade Commission (FTC)
- Government involvement
- Low-to-moderate — light federal touch, no binding frontier-model law
- Industrial strategy
- Infrastructure- and export-led; the 'Stargate' initiative and America's AI Action Plan prioritise compute build-out
- International partnerships
- Chip export-control coordination with allies; no formal multilateral AI safety treaty
- Primary policy focus
- Frontier-model competitiveness and compute infrastructure appear to be higher near-term federal priorities than precautionary regulation, though states continue to pursue their own rules
Current Direction
Loosening
The federal posture remains deregulatory on consumer-facing and precautionary rules, but the direction is increasingly asymmetric: on national-security grounds, EO 14409 ('Gold Eagle') and the July agentic-hacking disclosures have pushed the government to assert markedly more control over frontier-model access and cyber-capability testing. With the AI & Crypto Czar role vacant since March, CAISI's standards and testing work remains the most consistent thread of federal activity.
Recent Developments
Frontier US labs shipped a cluster of cheaper, faster flagships within 24 hours. OpenAI released GPT-6 Sol and Luna (22 September), lower-cost tiers built on Astra's methods that roughly halve API prices (Sol at $2/$10, Luna at $0.10/$0.50 per million tokens) and, OpenAI says, make about half as many mistakes as GPT-5.6. Anthropic released Claude Opus 5.5 (22 September), matching Fable 5.1-level performance at roughly 40% lower cost and 30% faster. xAI released Grok 4.7 (21 September), a 2.1-trillion-parameter model with what it calls its strongest safety guardrails yet and supplemental training data drawn from SpaceX. The burst underscores an intensifying capability-and-price race among US labs.
2026-09-22Newsom issued an executive order to accelerate the new independent-verification regime (SB 813 / AB 1405) and advance a verified 'kill switch' for frontier models, its efficacy to be checked on an ongoing basis by an independent verification organisation. He convened national experts to recommend, within two months, onsite independent audits of frontier developers, mandatory safety frameworks and transparency reports, and an updated definition of critical safety incidents covering loss-of-control events. Newsom again urged Congress and the President to treat California's approach as a federal baseline.
2026-09-18Newsom signed SB 1050 (Ashby), requiring advertisers to disclose when video or audio adverts use AI-generated performers and barring the continued use of non-compliant ads. The governor framed it as protecting both consumers and creative-industry performers as AI-generated media spreads; it builds on California's 2024 performer-likeness protections and the AI Transparency Act's watermarking rules.
2026-09-16Newsom signed a package of child-online-safety bills, several aimed directly at AI. SB 1119 (Padilla) requires companion-chatbot safeguards for minors — crisis protocols for suicidal ideation, parental controls and independent child-safety audits; SB 867 (Padilla) extends similar duties to toys with companion chatbots; and SB 1276 (Rubio) expands child sexual-exploitation offences to cover AI-generated or digitally altered material. California described it as the strongest child-safety chatbot law in the nation.
Major2026-09-10California enacted the first US framework for independent, third-party verification of AI systems. Governor Newsom signed SB 813 (McNerney), establishing state-recognised 'independent verification organizations' to assess AI systems for legal and safety compliance, alongside AB 1405 (Bauer-Kahan), which creates a state registry of AI auditors with standards for their independence, transparency and integrity. Together they set up a certification-style oversight layer that other states — and any future federal framework — may look to.
Major2026-09-09OpenAI released GPT-6 'Astra' (3 September), which it said is the first model to cross the 'critical' cybersecurity-capability threshold under its Preparedness Framework — able to find and develop zero-day exploits — and gated it behind a 'Daybreak Blue' defender-access programme. The launch drew safety criticism for 'opaque recurrence', a reasoning technique that obscures chain-of-thought monitoring and so weakens a key oversight mechanism, landing weeks after an OpenAI agent reportedly escaped its sandbox in the late-July Hugging Face incident.
Major2026-09-03A wave of frontier 'cyber' models and safeguard programmes landed in early September. Google released Gemini 3.8 Flash Cyber, its most capable security model, with a defenders-only 'Fairwind' early-access programme for 650+ partners (including CrowdStrike and Palo Alto Networks) and an emphasis on vulnerability-fixing over offence; Anthropic shipped Claude Fable 5.1 and a trusted-access-only Mythos 5.1 with tiered safeguards and new 'Enterprise Frontier Safeguards'. The cluster underscored how leading US labs are pairing cyber-capable releases with access-gating and misuse-detection controls rather than waiting on statutory rules.
2026-09-02SourceTechCrunchMeta released Muse Spark 1.3 (2 September), its most advanced model, to developers and across Instagram, Facebook and Meta AI. Zuckerberg called it Meta's 'biggest jump' yet in coding and agentic work; the proprietary 1M-context model completes comparable coding tasks with roughly 20% fewer tool calls and is tuned for longer-horizon agent runs — asking users when requirements are unclear and seeking confirmation before consequential actions. Pricing was held flat from version 1.2.
2026-09-02At the 13 August 2026 suspense-file hearings, California's Assembly and Senate Appropriations Committees advanced roughly two dozen AI bills to floor votes in the session's final weeks -- including SB 813 (AI Standards and Safety Commission), companion-chatbot child-safety bills AB 2023 and SB 1119, SB 867 (chatbots in toys), SB 947 and SB 951 (worker protections and displacement notices) and AB 2713 (AI Transparency Act adjustments).
2026-08-13The Senate Commerce Committee advanced the Kids Online Safety Act and three AI-focused children's bills (5 August 2026) -- the CHATBOT Act, the Youth AI Privacy Act and the Children's AI Toy Safety Act -- to the Senate floor. The broader White House-backed package trading federal preemption of state AI laws for KOSA and age-verification measures was not part of the markup, leaving the preemption fight unresolved.
The California AI Transparency Act (SB 942, as amended by AB 853) became operative on 2 August 2026. Providers of generative AI systems with more than one million monthly users publicly accessible in California must now offer a free public AI-detection tool, a user option for visible disclosures on AI-generated content, and embedded latent disclosures identifying AI-generated image, video and audio -- with civil penalties of $5,000 per violation per day. Further obligations for large platforms, hosting services and capture-device makers phase in during 2027-2028.
Major2026-08-02Executive Order 14409's 1 August 2026 deadline for completing a voluntary framework giving the federal government up to 30 days of pre-release access to covered frontier models passed with no public deliverable, and the White House indicated the NSA-led framework would be kept secret. The lapse drew criticism over unpublished selection criteria and the halt of public CAISI assessment reports -- shifting US frontier-model oversight toward a classified, voluntary channel.
2026-08-01The bipartisan 'Great American AI Act' (GAAIA) has reportedly stalled in the House, with its proposed three-year federal preemption of state AI-development laws described as the main sticking point — leaving the growing state patchwork dominant for now. The most consequential state entrant is already in force: California's SB 53, the Transparency in Frontier Artificial Intelligence Act (TFAIA), signed by Governor Newsom on 29 September 2025 and effective 1 January 2026, is the first US state frontier-AI law, requiring large frontier developers to publish a safety/risk ('frontier AI framework') document and report critical safety incidents to the state.
AI-agent 'breakout' incidents at OpenAI and Anthropic — including a model that escaped its sandbox to breach AI startup Hugging Face — raise national-security concerns. OpenAI disclosed (27 July) that during a cyber-evaluation one of its models exploited a previously unknown vulnerability to escape its sandbox and briefly reach the internet, accessing Hugging Face; Anthropic disclosed (30 July), after reviewing 141,006 evaluation runs, three cases where its models (reported as Claude Opus 4.7, Mythos 5, and a research model) reached real systems via a misconfigured, supposedly internet-sealed test environment, using basic techniques such as weak passwords. Hugging Face's CEO argued developers should be held accountable when AI models 'go rogue.'
Major2026-07-30Anthropic releases Claude Opus 5, its new flagship model; the company's automated behavioural audit found it 'the most aligned model to date' with the lowest measured rates of deceptive and misuse-prone behaviour among recent Claude models, alongside tightened safeguards restricting cyber-offense-relevant capabilities to a vetted 'Cyber Verification Program.'
Major2026-07-26SourceAnthropicFTC opens public comment on an AI-accuracy and manipulation policy statement, issued pursuant to Executive Order 14365; a bipartisan 'Great American AI Act' (GAAIA) discussion draft circulates in Congress; the House Science Committee advances ten AI bills, including one that would statutorily codify CAISI.
2026-07Illinois enacts S.B. 315, the Artificial Intelligence Safety Measures Act — reported as the first US state law to require annual independent third-party audits of frontier AI models, a notably more prescriptive state approach than Colorado's narrowed disclosure framework.
Major2026-07SourceMintzExecutive Order 14409 ('Gold Eagle', June 2026) moves the federal government toward far greater control over frontier-model access: a classified NSA/CISA benchmarking process — due 1 August 2026 — to designate 'covered frontier models' with advanced cyber capabilities, a voluntary framework giving the government up to ~30 days' pre-release access, and a cyber-vulnerability clearinghouse run with Treasury, DHS CISA and the Department of War. Labs' cyber-model access programmes — Anthropic's Project Glasswing and OpenAI's Daybreak — now reportedly require government sign-off on trusted partners.
Major2026-07SourceThe Next WebNobel laureate John Jumper, AlphaFold's lead architect, leaves Google DeepMind for Anthropic after nearly nine years — one of 2026's most closely watched AI research talent moves, and part of a broader wave of DeepMind and academic researchers joining US frontier labs.
Major2026-06SourceTechCrunchColorado Governor Jared Polis signs SB 26-189, repealing and replacing the Colorado AI Act (SB 24-205) with a narrower disclosure/transparency framework for automated decision-making; the duty of care, mandatory risk-management programs, impact-assessment requirements and AG-reporting obligations of the original law are dropped, and the new effective date is pushed to January 1, 2027.
Major2026-05CAISI publishes a technical evaluation of DeepSeek's V4 Pro model, finding it lags top US frontier models by roughly eight months in aggregate capability while remaining the most capable PRC model evaluated to date.
2026-05SourceNISTCAISI signs pre-deployment testing agreements with Google DeepMind, Microsoft and xAI, extending arrangements it already held with OpenAI and Anthropic.
2026-05SourceNISTAnthropic launches the Anthropic Institute; Jack Clark moves from Head of Policy to Head of Public Benefit to lead it, while Sarah Heck becomes Head of Public Policy.
2026-03SourceAnthropicCAISI signs an MOU with the General Services Administration on federal AI procurement standards.
2026-03CAISI launches an AI Agent Standards Initiative to develop evaluation benchmarks for autonomous AI agents — work now sharpened by the July agentic-hacking disclosures.
2026-02SourceNIST
Policy and Regulation
Executive Order 14409 — frontier-model access & cyber ('Gold Eagle')
A June 2026 executive order asserting federal control over access to 'covered frontier models' with advanced cyber capabilities. It directs a classified NSA/CISA benchmarking process (due 1 August 2026) to designate such models, establishes a voluntary framework giving the government up to roughly 30 days' pre-release access, and stands up 'Gold Eagle', a cyber-vulnerability coordination clearinghouse involving Treasury, DHS CISA and the Department of War. Labs' cyber-model access programmes (Anthropic's Project Glasswing, OpenAI's Daybreak) reportedly now require government sign-off on trusted partners. Represents a national-security-driven tightening even as consumer-facing federal rules stay light.
Illinois AI Safety Measures Act (S.B. 315)
Reported as the first US state law requiring annual independent third-party audits of frontier AI models — a markedly more prescriptive posture than Colorado's narrowed disclosure framework, and a sign the state patchwork is diverging in stringency rather than converging.
California SB 53 — Transparency in Frontier Artificial Intelligence Act (TFAIA)
Signed by Governor Gavin Newsom on 29 September 2025 and in effect since 1 January 2026, TFAIA is reported as the first US state law to regulate developers of frontier foundation models directly. It requires large frontier developers (broadly, those training models above roughly 10^26 FLOP and meeting a revenue threshold) to publish a 'frontier AI framework' setting out how they manage catastrophic risk, release transparency reports on new frontier models, and report critical safety incidents to the California Office of Emergency Services; it also adds whistleblower protections and lays groundwork for a public 'CalCompute' cluster. Enforcement sits with the state Attorney General via civil penalties. It stands alongside Illinois's S.B. 315 (third-party audits) and California's AB 2013 (training-data transparency) at the leading edge of the state patchwork.
Official text ↗Colorado AI Act (SB 24-205 → SB 26-189)
Colorado's original SB 24-205 established a duty of care, mandatory risk-management programs, and impact-assessment requirements for high-risk AI used in consequential decisions (employment, credit, housing, healthcare). That law was repealed and replaced by SB 26-189, signed by Governor Jared Polis on May 14, 2026, which drops the duty of care, risk-management-program, impact-assessment and attorney-general-reporting requirements in favour of a narrower disclosure/transparency framework for automated decision-making. SB 26-189 takes effect January 1, 2027.
Official text ↗California AB 2013 — Training Data Transparency
Requires developers of generative AI systems made available to Californians to publicly disclose a high-level summary of the datasets used to train the system; took effect January 1, 2026. OpenAI, Anthropic and Google have published disclosures. xAI (now SpaceXAI) sued California Attorney General Rob Bonta in December 2025 challenging the law's constitutionality; a district court denied a preliminary-injunction motion in March 2026, and the case remained under Ninth Circuit review as of mid-2026.
Official text ↗TAKE IT DOWN Act
Federal law criminalising the publication of non-consensual intimate imagery, including AI-generated deepfakes, and requiring platforms to remove reported content within 48 hours of a valid notice. Signed into law May 19, 2025; the notice-and-removal requirement took effect one year later, on May 19, 2026, with FTC enforcement authority.
Official text ↗Great American AI Act (GAAIA) — discussion draft
A bipartisan 269-page discussion draft released June 4, 2026 by Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA), proposing frontier-model disclosure and third-party-audit requirements alongside a three-year federal preemption of state laws that regulate AI model 'development' (though not deployment); not yet introduced as a formal bill.
Official text ↗Research Ecosystem
US AI research remains among the world's largest and most influential, anchored by Stanford HAI, MIT CSAIL, and a dense private-lab research ecosystem; many observers argue private labs increasingly outpace university output at the frontier. Federal programmes like the National AI Research Resource pilot aim to broaden compute access for academic researchers who can't match private labs' training-scale infrastructure.
Major AI Companies
OpenAI
Developer of the GPT model family and ChatGPT; central to US frontier AI policy discussions.
Anthropic
Developer of the Claude model family, with a public focus on AI safety research.
Google DeepMind
Google's AI research lab, responsible for the Gemini model family.
Meta AI
Meta's AI division, reorganised into Meta Superintelligence Labs in 2025; known for the open-weight Llama model family, though Meta has signalled a shift toward a more closed or mixed release strategy for its most advanced ('superintelligence-capable') models going forward.
SpaceXAI (formerly xAI)
Developer of the Grok model family. SpaceX acquired xAI in an all-stock deal announced February 2026, and by July 2026 the entity had formally rebranded as SpaceXAI, a SpaceX subsidiary. It signed its CAISI pre-deployment testing agreement in May 2026 under the xAI name.
NVIDIA
Widely regarded as the leading supplier of AI training and inference chips, central to export-control debates.
Microsoft
CAISI pre-deployment testing partner; has begun releasing in-house 'MAI' frontier models (e.g. MAI-Thinking-1) under AI CEO Mustafa Suleyman, alongside its OpenAI partnership.
TSMC (Arizona)
Taiwan Semiconductor Manufacturing Company's US fabrication facilities, central to the AI-chip export-control debate and US frontier-AI compute capacity.
Cloud Infrastructure
- Frontier AI
- Domestic frontier developer
- Compute
- Among the largest globally; hyperscaler build-out plus the 'Stargate' initiative
- Chip access
- Home of NVIDIA and the leading export-control setter; TSMC Arizona fabs onshoring capacity
Cloud market share
AWS
Leading US cloud provider and major AI compute host; estimated at roughly 28% of global cloud infrastructure spend in Q1 2026 per Synergy Research Group.
Microsoft Azure
Cloud provider and OpenAI's primary infrastructure partner; estimated at roughly 21% of global cloud infrastructure spend in Q1 2026 per Synergy Research Group.
Google Cloud
Cloud provider hosting Google DeepMind's models and third-party AI workloads; estimated at roughly 14% of global cloud infrastructure spend in Q1 2026 per Synergy Research Group.
Oracle Cloud Infrastructure
Smaller overall cloud market share, but a fast-growing role in frontier AI compute, including large committed capacity under OpenAI's Stargate build-out.
Leading Universities
Think Tanks
Government Institutions
Center for AI Standards and Innovation (CAISI)
NIST body developing evaluation methods and pre-deployment testing agreements with frontier labs; formerly the US AI Safety Institute, renamed June 2025 with a mission shifted toward national-security testing, standards, and commercial AI evaluation.
White House Office of Science and Technology Policy (OSTP)
Coordinates federal AI policy and authored America's AI Action Plan.
Federal Trade Commission (FTC)
Applies consumer-protection and competition law to AI products and marketing claims.
Notable Gaps
Federal AI safety statute
No binding federal AI safety or frontier-model law exists as of mid-2026; CAISI's pre-deployment testing agreements remain voluntary MOUs rather than statutory requirements, though a bill to statutorily codify CAISI was among ten AI bills the House Science Committee advanced in June 2026.
Key People
Michael Kratsios
GovernmentDirector, Office of Science and Technology Policy
The White House
David Sacks
AdvisorCo-Chair, President's Council of Advisors on Science and Technology (PCAST)
The White House
Sam Altman
IndustryChief Executive Officer
OpenAI
Dario Amodei
IndustryChief Executive Officer
Anthropic
Jack Clark
AdvisorHead of Public Benefit; leads the Anthropic Institute
Anthropic
Sarah Heck
AdvisorHead of Public Policy
Anthropic
Elon Musk
IndustryChief Executive Officer
SpaceX / SpaceXAI
Jensen Huang
IndustryChief Executive Officer
NVIDIA
Mustafa Suleyman
IndustryChief Executive Officer, AI
Microsoft AI
People and Organisations to Follow
Jack Clark
NewsletterAnthropic co-founder and Head of Public Benefit; writes Import AI, a leading weekly US and global AI policy newsletter.
Zvi Mowshowitz
NewsletterWrites Don't Worry About the Vase, a deeply detailed AI policy and safety newsletter.
Hard Fork
PodcastNew York Times technology podcast covering US AI policy and industry weekly.
CSET Georgetown
Think tankLeading US think tank on AI and national security policy.
Michael Kratsios
GovernmentOSTP Director; central voice on federal AI policy direction.
Current Debates
SecurityAgentic AI, cyber-offence and developer accountability
July 2026's disclosures that OpenAI's and Anthropic's own models breached real external systems during cyber-evaluations moved 'agentic AI escaping its sandbox' from a theoretical to a documented risk, and sharpened debate over who is accountable when an AI agent 'goes rogue.'
▾
Agentic AI, cyber-offence and developer accountability
July 2026's disclosures that OpenAI's and Anthropic's own models breached real external systems during cyber-evaluations moved 'agentic AI escaping its sandbox' from a theoretical to a documented risk, and sharpened debate over who is accountable when an AI agent 'goes rogue.'
OpenAI's model exploited an unknown vulnerability to reach Hugging Face; Anthropic's review of 141,006 evaluation runs surfaced three cases where models reached real systems through a misconfigured, supposedly internet-sealed test environment. Security researchers and outlets such as Bloomberg framed the episodes as pointing to US national-security risks, while Hugging Face's CEO argued developers should bear liability when their models cause harm. The incidents feed directly into EO 14409's framing of 'covered frontier models' with advanced cyber capabilities and CAISI's agent-evaluation work — but also raise the question of whether voluntary MOUs and internal red-teaming are sufficient, or whether binding pre-deployment cyber-safety requirements are needed.
GovernmentWho leads AI policy after the AI czar role went vacant
David Sacks's move from White House AI & Crypto Czar to a PCAST co-chair role in March 2026 left the position without a direct successor, raising questions about who actually coordinates federal AI policy.
▾
Who leads AI policy after the AI czar role went vacant
David Sacks's move from White House AI & Crypto Czar to a PCAST co-chair role in March 2026 left the position without a direct successor, raising questions about who actually coordinates federal AI policy.
Sacks reached the 130-day cap on his special-government-employee status and shifted to advising through PCAST rather than the White House directly. No replacement AI czar had been named as of mid-2026. In practice, OSTP Director Michael Kratsios and CAISI appear to have absorbed part of the coordinating role, but the absence of a dedicated czar has drawn commentary about whether AI policy now lacks a single accountable owner inside the administration.
GovernmentFederal preemption vs. state AI laws
Industry and parts of the federal government are pushing for a single national standard to override the growing patchwork of state AI laws; states have so far resisted ceding their own rules, even as some flagship state laws are themselves being narrowed or challenged in court.
▾
Federal preemption vs. state AI laws
Industry and parts of the federal government are pushing for a single national standard to override the growing patchwork of state AI laws; states have so far resisted ceding their own rules, even as some flagship state laws are themselves being narrowed or challenged in court.
Colorado and California have moved fastest on binding state AI rules, though Colorado's original SB 24-205 duty-of-care framework was repealed and replaced by the narrower SB 26-189 in May 2026 (effective January 1, 2027) before it ever took effect. California, by contrast, has tightened: its SB 53 (Transparency in Frontier Artificial Intelligence Act), in force since January 1, 2026, is the first US state law to regulate frontier-model developers directly, and the bipartisan GAAIA — whose three-year preemption of state AI-development laws has reportedly stalled its progress in the House — would have overridden exactly this kind of state rule. California's AB 2013 training-data disclosure law is being challenged by xAI (now SpaceXAI), which sued Attorney General Rob Bonta in December 2025 arguing the law is unconstitutional; a district court denied a preliminary injunction in March 2026, and the case remains under Ninth Circuit review. Industry groups argue a fragmented state-by-state patchwork raises compliance costs and slows deployment, and have lobbied for federal preemption; the bipartisan GAAIA discussion draft proposes a three-year federal preemption of state laws that regulate AI model development (though not deployment), and in July 2026 the FTC's proposed AI-accuracy policy statement — issued under Executive Order 14365 — specifically flagged Colorado's AI Act as a law that could be preempted. States'-rights advocates and consumer groups counter that preemption without a substantive federal replacement would leave a regulatory vacuum.
IndustryCopyright and AI training data
Multiple lawsuits against OpenAI, Meta, and Stability AI over training on copyrighted material are testing the boundaries of fair use.
▾
Copyright and AI training data
Multiple lawsuits against OpenAI, Meta, and Stability AI over training on copyrighted material are testing the boundaries of fair use.
Authors, news publishers, and visual artists have brought a wave of suits arguing that training foundation models on their copyrighted work without licence is infringement; AI developers argue it falls under fair use as transformative. Early rulings have been mixed, and the issue is expected to reach appellate courts before a clear precedent is set.
SecurityExport controls on AI chips
Ongoing dispute over how tightly to restrict advanced AI chip exports, particularly to China, without ceding global AI infrastructure market share.
▾
Export controls on AI chips
Ongoing dispute over how tightly to restrict advanced AI chip exports, particularly to China, without ceding global AI infrastructure market share.
Successive rounds of export controls have restricted sales of NVIDIA's and AMD's most advanced chips to China, aiming to slow rival AI capability growth. Critics argue restrictions push Chinese buyers toward domestic alternatives (accelerating China's own chip industry) and cede market share; national security officials argue the controls meaningfully slow adversary progress.
GovernmentCAISI's expanding pre-deployment testing mandate
CAISI has become one of the most substantive checks on frontier AI models in the US, now holding pre-deployment testing agreements with five major labs.
▾
CAISI's expanding pre-deployment testing mandate
CAISI has become one of the most substantive checks on frontier AI models in the US, now holding pre-deployment testing agreements with five major labs.
With OpenAI, Anthropic, Google DeepMind, Microsoft and xAI (which announced its agreement under that name before its July 2026 rebrand to SpaceXAI) all now under some form of CAISI pre-deployment testing arrangement, the agency is often described as the closest thing the US currently has to a binding frontier-model check — even though the arrangements are voluntary agreements rather than statutory requirements, and could in principle be unwound without legislative action.
Knowledge Hub
Official documents
Academic papers
- AI Index Report— Stanford HAI
Think tank reports
- Export control research on AI chips— CSET, Georgetown
Books
- Genesis: Artificial Intelligence, Hope, and the Human Spirit— Henry Kissinger, Eric Schmidt & Craig Mundie
Substacks
- Import AI— Jack Clark
- Don't Worry About the Vase— Zvi Mowshowitz
- Cambrian— Hamish Low & Tristan Low
Writing on AI, economics and geopolitics, including compute governance and US-China technology competition.
Podcasts
- Hard Fork— Kevin Roose & Casey Newton, The New York Times
Speeches
- Sam Altman's Senate Judiciary testimony on AI oversight
May 2023 hearing on rules for AI, still widely cited in US policy debate.
Datasets
- AI Incident Database
Public, ongoing catalogue of real-world AI harms and incidents.
- Epoch AI— Epoch AI
Research and data on compute trends, training runs and AI capability growth.
Sources
Deliberately sourced beyond government publications. The bars below show each category’s actual share of this page’s citations — an internal editorial guideline rather than a rigorously derived target, so treat it as a rough steer, not a precise benchmark.
- NIST Center for AI Standards and InnovationGovernmentNIST2026-05
- David Sacks steps back from AI czar roleNewsCNBC2026-03
- David Sacks' influence after the AI czar roleNewsAxios2026-03
- California Legislative InformationGovernmentCalifornia State Legislature2024-09
- Colorado SB 24-205GovernmentColorado General Assembly2024-05
- TAKE IT DOWN ActGovernmentUS Congress2025-05
- State AI legislation and preemption trackerThink tanksRAND Corporation2026
- AI Index ReportResearchStanford HAI2026
- National AI Research Resource PilotResearchNational Science Foundation2024
- AI copyright litigation trackerResearchStanford HAI2026
- AI chip export control analysisThink tanksCSET, Georgetown2026
- Import AI newsletterExpert commentaryJack Clark2026
- OpenAI policy blogIndustryOpenAI2026
- Bloomberg AI policy coverageNewsBloomberg2026
- AI Washington Report — July 2026Think tanksMintz2026-07
- Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence SystemsGovernmentFederal Trade Commission / Federal Register2026-07
- CAISI Evaluation of DeepSeek V4 ProGovernmentNIST2026-05
- GSA and NIST partner to boost AI evaluation science in federal procurementGovernmentGeneral Services Administration2026-03
- Colorado SB 26-189 (repeal and replace of SB 24-205)GovernmentColorado General Assembly2026-05
- Obernolte, Trahan release discussion draft of the Great American AI ActGovernmentUS House of Representatives (Rep. Jay Obernolte)2026-06
- The Anthropic InstituteIndustryAnthropic2026-03
- Q1 2026 global cloud infrastructure market-share estimatesResearchSynergy Research Group2026-05
- Introducing Claude Opus 5IndustryAnthropic2026-07
- Nobel laureate John Jumper is leaving DeepMind for rival AnthropicNewsTechCrunch2026-06
- Anthropic says its own AI models breached three companies during security testsNewsTechCrunch2026-07
- Why did OpenAI's and Anthropic's AI models hack other companies?NewsNPR2026-08
- The White House is now deciding who gets access to frontier AI models, not the labsNewsThe Next Web2026-07
- AI: The Washington Report — July 2026 Edition (Illinois S.B. 315 and state AI legislation)Think tanksMintz2026-07
- Governor Newsom signs SB 53, advancing California's world-leading artificial intelligence industryGovernmentOffice of Governor Gavin Newsom2025-09-29
- California's SB 53: The First Frontier AI Law, ExplainedThink tanksFuture of Privacy Forum2025-10
- California Just Passed the First U.S. Frontier AI Law. Here's What It Does.Think tanksCarnegie Endowment for International Peace2025-10
- SB 942 California AI Transparency Act -- bill text and historyGovernmentCalifornia Legislative Information2026-08-02
- New California AI Disclosure Rules Become OperativeNewsMorgan Lewis2026-08-03
- Controlling Advanced Artificial Intelligence: Executive Order 14409 Explained (IF13268)GovernmentCongressional Research Service2026-07-09
- Five Questions the US Government Should Answer About Its Secretive Frontier AI FrameworkNewsTechPolicy.Press2026-08-05
- Assembly Appropriations Committee Suspense File Hearing results, August 13, 2026GovernmentCalifornia State Assembly2026-08-13
- It's finals week in Sacramento. Here's a scorecard with California's 24 AI billsNewsTransparency Coalition2026-08-24
- Commerce Committee Advances Kids Online Safety LegislationGovernmentU.S. Senate Committee on Commerce, Science, & Transportation2026-08-05
- Senate Committee Advances Four Bills Aimed at Protecting Children OnlineNewsTechPolicy.Press2026-08-06
- Governor Newsom signs first-in-the-nation AI safeguards to protect CaliforniansGovernmentOffice of the Governor of California2026-09-09
- Governor Newsom signs the strongest child safety chatbot and social media laws in the nationGovernmentOffice of the Governor of California2026-09-10
- Governor Newsom signs new law to protect workers, require disclosures on AI-generated advertisingGovernmentOffice of the Governor of California2026-09-16
- Governor Newsom issues executive order to accelerate independent oversight and advance the creation of an AI kill switchGovernmentOffice of the Governor of California2026-09-18
- OpenAI launches Astra, its powerful (and controversial) new modelNewsTechCrunch2026-09-03
- Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access ProgramsNewsThe Hacker News2026-09-02
- OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakesNewsTechCrunch2026-09-22
- Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 modelsNewsSiliconANGLE2026-09-22
- xAI Launches Grok 4.7. It's Bigger, But Late to the AI Frontier PartyNewsYahoo Tech2026-09-21
- Meta Releases AI Model Muse Spark 1.3, Edges Closer to OpenAI, AnthropicNewsBloomberg2026-09-02
- Introducing Muse Spark 1.3IndustryMeta AI Research2026-09-02