Which laws regulate frontier AI models?

Which laws actually regulate frontier AI models today?

As of October 2026, five jurisdictions have binding rules aimed at frontier models or the services built on them: the European Union, California, New York, China, and South Korea. The United Kingdom and the U.S. federal government have no binding frontier-model statute.

That is a small set. Most AI laws regulate uses, such as hiring tools, credit decisions, deepfakes, and chatbots for children. The laws in this guide regulate the models themselves, based on how large or capable they are, or in China’s case on whether a service reaches the public.

None of these laws requires a government license before training a frontier model. They require notice, documentation, published safety plans and incident reports. For the difference, see AI licensing explained.

What does the EU AI Act require of the largest models?

The EU AI Act, Regulation (EU) 2024/1689, requires every general-purpose model provider to document its model and summarize its training content, and adds testing, risk-mitigation, incident-reporting, and cybersecurity duties for models presumed to carry systemic risk. Those are its two tiers.

All general-purpose AI models

Under Article 53, every provider must keep technical documentation, give information to downstream developers, adopt a copyright policy that respects EU text-and-data-mining opt-outs, and publish a summary of training content using the AI Office template, released July 24, 2025.

Open-source models get a partial exception from the documentation duties. It does not cover the copyright policy or the training summary, and it does not apply to systemic-risk models.

Models with systemic risk

Article 51 presumes a model has “high impact capabilities” when its cumulative training compute is greater than 10²⁵ floating-point operations. The Commission can also designate a model on other criteria, and can change the threshold by delegated act. We found no delegated act changing it as of October 2026.

A provider must notify the Commission “without delay and in any event within two weeks” after the threshold is met or it becomes known that it will be met (Article 52).

Article 55 then adds four duties: evaluate the model with adversarial testing, assess and mitigate systemic risks, track and report serious incidents to the AI Office, and protect the model and its infrastructure with adequate cybersecurity.

Dates and fines

• August 2, 2025: obligations for general-purpose AI model providers applied.
• August 2, 2026: the Commission’s enforcement powers, including fines, began.
• August 2, 2027: deadline for models placed on the market before August 2, 2025.

Under Article 101, fines for general-purpose model providers can reach 3% of worldwide annual turnover or €15 million, whichever is higher.

Did the 2026 “digital omnibus” delay these rules?

No. The AI omnibus, Regulation (EU) 2026/1744, entered into force on July 27, 2026, according to a Mondaq summary. It pushed back high-risk system deadlines to December 2, 2027 and August 2, 2028. It left the general-purpose model obligations on their original schedule.

What is the Code of Practice?

The General-Purpose AI Code of Practice is a voluntary rulebook the Commission published on July 10, 2025, with chapters on transparency, copyright, and safety and security. Signing it is one way to show compliance. The Commission’s signatory list includes Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. xAI signed only the safety and security chapter. Meta declined to sign in July 2025.

What does California’s SB 53 require?

SB 53 requires frontier developers to publish safety documents and report critical safety incidents to the state. It does not require approval. Governor Gavin Newsom signed the bill, the Transparency in Frontier Artificial Intelligence Act, on September 29, 2025. Its core duties took effect January 1, 2026.

• Frontier model: a foundation model trained with more than 10²⁶ operations, counting the original run plus later fine-tuning and material modifications.
• Frontier developer: anyone who trained or started training a frontier model.
• Large frontier developer: a frontier developer with more than $500 million in annual gross revenue, combined with affiliates, in the prior year.

Large frontier developers must publish a frontier AI framework explaining how they assess and manage catastrophic risk. Every frontier developer must publish a transparency report before or when it deploys a new or substantially modified frontier model.

Critical safety incidents must be reported to the California Office of Emergency Services within 15 days of discovery. An incident that poses an imminent risk of death or serious physical injury must be disclosed within 24 hours to an appropriate authority, such as a law enforcement or public safety agency.

“Catastrophic risk” means a foreseeable, material risk that a model contributes to the death or serious injury of more than 50 people or more than $1 billion in damage, through weapons help, autonomous cyberattacks or crimes, or evading its developer’s control.

Only the Attorney General can sue for the law’s civil penalties: a large frontier developer that fails to publish required documents, report an incident, or follow its own framework faces up to $1 million per violation. The law also protects whistleblowers and creates a consortium to design a public computing cluster called CalCompute, once the state funds it. Starting by January 1, 2027, the Department of Technology must assess each year whether to update the definitions of “frontier model,” “frontier developer” and “large frontier developer,” and recommend changes to the Legislature.

What does New York’s RAISE Act require?

New York’s RAISE Act follows California’s SB 53 closely, but sets a 72-hour incident deadline, adds a disclosure statement and a fee to fund the regulator, and creates an enforcement office at the Department of Financial Services.

Governor Kathy Hochul signed the Responsible AI Safety and Education (RAISE) Act on December 19, 2025, on the condition that lawmakers pass amendments. She signed that chapter amendment on March 27, 2026. The final law takes effect January 1, 2027.

The final version uses the same 10²⁶ compute definition and the same $500 million revenue tier, according to Cooley. It dropped the original bill’s $100 million training-cost test, its third-party audit requirement and its deployment ban.

The differences that matter:

• Critical safety incidents must be reported within 72 hours, not 15 days. Imminent-risk incidents still have 24 hours.
• Large frontier developers may not develop, deploy or operate a frontier model in New York without a current disclosure statement on file with the regulator. The statement must be renewed at least every two years, and developers must pay an assessment, a fee that covers the regulator’s operating expenses.
• A new office inside the Department of Financial Services enforces the law and can write rules.
• Penalties are up to $1 million for a first violation and $3 million for later ones.

Because a large frontier developer may not develop, deploy or operate a frontier model in New York without a current statement on file, the requirement works as a registration duty. Cooley’s summary describes no power for the regulator to approve or refuse a model.

How does China regulate generative AI models?

China requires public-facing generative AI services to carry out a security assessment and complete a filing with the Cyberspace Administration of China (CAC).

The Interim Measures for the Management of Generative Artificial Intelligence Services took effect August 15, 2023. Under Article 17, services with “public opinion attributes or social mobilization capabilities” must carry out a security assessment and complete an algorithm filing. Under Article 2, the measures do not apply to organizations that develop or use generative AI without offering services to the public in China.

The CAC publishes the results in batches. Its September 14, 2026 notice reported that 1,112 generative AI services had completed filing and 731 apps or features that call an already-filed model through an interface had registered with local cyberspace offices, as of August 31, 2026.

China has added other rules on top. Labeling measures for AI-generated content took effect September 1, 2025. An amendment to the Cybersecurity Law, passed in October 2025 and effective January 1, 2026, added the law’s first article on AI.

China’s system is not keyed to compute. It is keyed to whether a service reaches the public.

What about South Korea and the United Kingdom?

South Korea has binding duties for the largest models; the United Kingdom does not.

South Korea’s AI Basic Act took effect January 22, 2026. The Ministry of Science and ICT’s legislative notice for the Act’s enforcement decree designates systems trained with at least 10²⁶ floating-point operations of cumulative compute as “high-performance AI” with added safety duties, according to Cooley. Large foreign providers without a Korean office must appoint a local representative. The ministry has signaled a one-year grace period before fines.

The United Kingdom has no AI bill. The King’s Speech in May 2026 contained “no dedicated AI Bill,” according to Lewis Silkin. The UK’s AI Security Institute, renamed from the AI Safety Institute in February 2025, tests models under voluntary arrangements with developers. It has no statutory power to approve or block a model.

Is the U.S. federal government overriding state laws?

Not yet. The executive branch is pressing against state AI laws, but as of October 2026 no federal statute preempts them.

On December 11, 2025, President Trump signed Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence.” It calls for a “minimally burdensome national standard” and directs:

• an AI Litigation Task Force at the Justice Department to challenge state AI laws inconsistent with the order’s policy;
• a Commerce Department review of “onerous” state laws;
• limits on remaining broadband (BEAD) grant funds for states with such laws;
• an FCC proceeding on a federal reporting standard and an FTC policy statement;
• a legislative recommendation for a federal framework that would preempt conflicting state laws. Under section 8(b), that recommendation is not to propose preempting state laws on child safety, data center infrastructure or state procurement.

Attorney General Pam Bondi set up the task force on January 9, 2026, according to Baker Botts. In April 2026 the Justice Department moved to intervene in xAI’s lawsuit against Colorado’s AI Act, Axios reported. A federal magistrate judge stayed enforcement of the Colorado law on April 27, 2026, and Colorado replaced it on May 14, 2026 with SB 26-189, which takes effect January 1, 2027. We found no federal suit against SB 53 or the RAISE Act as of October 2026.

An executive order cannot by itself void a state statute. State laws stay enforceable unless a court strikes them down or Congress preempts them.

Congress has not preempted them. On July 1, 2025, the Senate voted 99–1 to strip a 10-year moratorium on state AI regulation from the budget reconciliation bill. On July 23, 2026, Representatives Jay Obernolte and Lori Trahan introduced the FRONTIER Act (H.R. 9925). It would set federal transparency, incident-reporting and independent-audit duties for large frontier developers, and it would bar states from imposing new substantive obligations on AI developers in the areas it covers, with exceptions including laws protecting minors and state procurement rules. On September 23, 2026, 26 state attorneys general wrote to Congress asking for federal frontier AI rules that do not preempt state law.

How do the frontier AI laws compare?

Jurisdiction	Status (Oct 2026)	Trigger	Core duty	Pre-approval?	Max penalty
European Union (AI Act)	In force; enforcement from Aug 2, 2026	>1025 FLOP (presumption) or designation	Notify within 2 weeks; evaluations, risk mitigation, incident reports, cybersecurity	No	3% of global turnover or €15M
California (SB 53)	In force Jan 1, 2026	>1026 operations; $500M revenue tier	Publish framework and transparency reports; incidents within 15 days	No	$1M per violation
New York (RAISE Act)	Takes effect Jan 1, 2027	>1026 operations; $500M revenue tier	Framework, reports, disclosure statement and regulator fee; incidents within 72 hours	No	$1M first, $3M later
China	In force since Aug 15, 2023	Public-facing generative AI service with public-opinion attributes	Security assessment and filing with the CAC	Filing and security assessment before public launch; no formal license	Varies by statute
South Korea (AI Basic Act)	In force Jan 22, 2026	≥1026 FLOP for safety duties	Lifecycle safety measures; local representative	No	KRW 30M administrative fines
United Kingdom	No AI bill	None	Voluntary testing with the AI Security Institute	No	None
United States (federal)	No statute; EO 14365	None	Litigation and funding pressure on state laws	No	None

What do these laws have in common?

Apart from China’s pre-launch filing, they regulate disclosure more than permission. Four of the five binding regimes, all but China, use a compute threshold. The reasons and the criticism are in compute thresholds.

Dates and thresholds here are stated as of October 2026 and some will change. This guide is general information, not legal advice.