AI safety regulation in 2026: which rules actually bind a frontier lab

Two pieces of law currently create binding obligations for a company training a frontier model. The EU AI Act, whose general-purpose model rules became enforceable on 2 August 2026, and California's Transparency in Frontier Artificial Intelligence Act, in force since 1 January 2026. Everything else that gets described as AI safety regulation is either a voluntary commitment, a testing arrangement, or a bill.

What does the EU AI Act actually require of a frontier lab?

The Act separates two things that get conflated: rules about high-risk uses of AI systems, and rules about the general-purpose models underneath them. A frontier lab is caught by the second set.

Obligations for providers of general-purpose AI models have applied since 2 August 2025 for models placed on the market on or after that date. Models already on the market before then generally have until 2 August 2027.

The Act presumes a model carries systemic risk when its cumulative training compute exceeds 10 to the 25th floating-point operations, which brings additional obligations around evaluation, incident reporting and cybersecurity. We have been through how the compute thresholds work and who sets them separately.

Why does August 2026 matter if the rules started in 2025?

Because for a year the obligations existed and nobody could enforce them.

The substantive duties applied from August 2025, but the European Commission, acting through the AI Office, only became entitled to investigate and enforce them, and the prohibited-practices rules, on 2 August 2026. Penalties for general-purpose model providers reach the higher of 15 million euros or 3 per cent of worldwide annual turnover.

This gap between duty and enforcement is the single most useful thing to understand about reading AI law. A compliance deadline tells you when a company is supposed to have done something. An enforcement date tells you when a regulator can do anything about it not having been done.

What did the Digital Omnibus change?

It moved the high-risk deadlines and left the transparency ones alone.

The Digital Omnibus on AI, in force from late July 2026, shifted the Annex III stand-alone high-risk obligations from 2 August 2026 to 2 December 2027. High-risk AI embedded in products already regulated under Annex I, such as medical devices, toys and machinery, moved to 2 August 2028. Certain public-authority systems moved later still.

What did not move: the Article 50 transparency duties, which have applied since 2 August 2026, and the general-purpose model regime. The Omnibus also added two new prohibited practices phasing in from 2 December 2026, and gave systems already on the market before August 2026 until 2 December 2026 to comply with the marking of AI-generated content.

So the headline that the EU delayed the AI Act is half right. It delayed the part about high-risk applications. It did not delay the part that applies to the model builders.

What does California's law require?

California's Transparency in Frontier Artificial Intelligence Act was signed on 29 September 2025 and took effect on 1 January 2026. It is the first US statute written specifically for frontier model developers.

It uses two thresholds. A frontier model is one trained using more than 10 to the 26th computing operations. A large frontier developer is a frontier developer with annual revenue of $500 million or more, which is the tier that carries the heavier duties.

The obligations are disclosure and reporting rather than licensing. Large frontier developers must publish a framework describing how they assess and manage catastrophic risk. All frontier developers must publish a transparency report before deploying a frontier model. Critical safety incidents must be reported to the California Office of Emergency Services within fifteen days, or within twenty four hours where there is an imminent threat. Civil penalties reach $1 million per violation.

New York passed its own frontier AI law with a different structure, and we have compared the two statutes clause by clause.

Is there any binding US federal rule?

Not one that sets safety obligations for frontier training runs. Federal activity has run through executive action, procurement conditions, export controls on compute, and agency enforcement under existing consumer-protection and civil-rights law.

That means the operative US rules for a frontier developer in 2026 are state rules, which is unusual for a technology this concentrated and is the reason preemption keeps appearing in Washington debate.

What about the United Kingdom, China and South Korea?

The United Kingdom has no binding frontier AI statute. Its AI Safety Institute was renamed the AI Security Institute on 14 February 2025, announced by the technology secretary Peter Kyle at the Munich Security Conference, and it conducts pre-deployment testing by agreement with labs rather than by legal requirement. That distinction matters: a lab can decline.

China regulates by use and by service rather than by capability threshold, with measures covering generative AI services, algorithmic recommendation and the labelling of synthetic content. Obligations attach to deploying a service to the Chinese public, not to the size of a training run.

South Korea's AI Framework Act was promulgated on 21 January 2025 and took effect on 22 January 2026 after a one-year transition. It imposes obligations on high-impact AI in sectors such as healthcare, energy and public services, and labelling requirements for some generative applications, and it reaches AI developed outside Korea where it affects Korean users.

What is voluntary, and does that matter?

Most of what the public hears about lab safety practice is voluntary, and the frameworks are the clearest example. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework are company documents. We have gone through DeepMind's capability levels and alert thresholds in detail.

Voluntary does not mean meaningless. California's statute requires a large frontier developer to publish a framework, which turns a voluntary document into a disclosure obligation without the state specifying its contents. The EU's general-purpose model rules work partly through a code of practice. In both cases the law requires you to have a policy and say what it is, rather than telling you what the policy must contain.

It also means the substance can change without anyone's approval. A framework is revised by the company that wrote it.

Does any of this cover the risks people actually worry about?

Partly, and the gap is worth naming.

Both binding regimes are built around disclosure and process. They require a developer to assess risk, write down how, publish some of it, and report incidents. Neither tells a developer that a capability is off limits, and neither requires permission before a training run.

That is a deliberate design choice rather than an oversight. Legislators writing in 2024 and 2025 could describe a process without having to define, in statute, what an unacceptable capability is. The cost is that compliance and safety can come apart: a developer that publishes a thorough framework, follows it, and ships a dangerous model has complied.

The EU does draw one hard line, in its list of prohibited practices, but those prohibitions are about uses such as social scoring and certain biometric categorization rather than about model capability. Two further prohibitions phase in from 2 December 2026.

So the honest summary of the current regime is that it regulates what developers must tell people, and not what they may build.

How should a reader judge a claim about AI regulation?

Four questions, in order.

• Which instrument is it: a regulation, a statute, an executive order, a code of practice, or a press release?
• What triggers it: a compute threshold, a revenue threshold, a use case, or a deployment to the public?
• When does the obligation apply, and separately, when can anyone enforce it?
• What is the penalty, and has it ever been imposed?

Run those four questions over most AI regulation coverage and a lot of it resolves into something considerably narrower than the headline.

What to watch next

Whether any enforcement action is actually brought under the EU's general-purpose model regime, since the powers arrived in August 2026 and a power unused is not yet a constraint. Whether California publishes or acts on the incident reports it now receives. And whether the Omnibus delay to the high-risk regime turns out to be the last one.