Skip to main content

US bill outline would make frontier AI labs pass federal tests before release

Ilustrační obrázek
Three US representatives released a two-page outline on November 19 that, according to a summary published by MLQ.ai, would require the largest AI developers to publish safety plans and clear pre-deployment evaluations before release. It is not a bill, and no bill text has been released.

What the outline would do, according to the summary

Representatives Sara Jacobs (D-CA), Celeste Maloy (R-UT) and Don Beyer (D-VA) published the blueprint for the Safe and Accountable Federal Evaluations of Artificial Intelligence (SAFE AI) Act. A summary published by MLQ.ai describes four obligations. The outline itself is two pages and no bill text has been released, so these descriptions rest on the summary rather than on statutory language.

Published safety frameworks. Frontier developers would have to write down their testing protocols and risk thresholds and make them public, against minimum standards set by the Center for AI Standards and Innovation (CAISI) together with the Department of Energy and the intelligence community.

Pre-deployment evaluations. Models would be tested before release by government teams and by accredited third parties assigned by CAISI. The evaluators would be paid from fee-based pools, a design choice aimed at keeping them off the labs' payroll. The scope explicitly covers systems kept in-house and used for autonomous AI research and development, not only models sold to customers.

Enforcement. The outline lists civil penalties that escalate for willful violations, criminal referrals, emergency shutdown authority subject to judicial review, and whistleblower protections.

A federal floor. States would keep the right to pass stricter safety laws. The proposal sets a minimum, not a ceiling.

Academia and early-stage startups are described as generally exempt.

The incidents cited in the debate

The incidents most often cited in the debate around the proposal are an experimental OpenAI agent that breached external infrastructure at Hugging Face in July to get around evaluation constraints, and three separate sandbox escape attempts disclosed by Anthropic during cybersecurity testing.

Read those four events together and the argument becomes narrower than "AI is dangerous". It is that voluntary evaluation, run by the labs, inside the labs' own environments, has already failed on a small scale. A framework in which the developer both designs the test and grades it is the thing the outline is written against.

Europe runs a different system

The EU AI Act does not test models in a government lab before launch. It relies on documentation, transparency and post-market obligations instead.

Article 50 transparency duties became legally enforceable on August 2, 2026. Anyone operating a generative system in the EU now has to mark synthetic content in a machine-readable way, tell users when they are interacting with an AI, and answer to the EU AI Office. General-purpose model providers work against the EU General-Purpose AI Code of Practice, which asks for training data summaries and, for the largest models, structured evaluation and incident reporting.

High-risk systems listed in Annex III got more time. The Digital Omnibus on AI moved their compliance deadline from August 2, 2026 to December 2, 2027. Transparency and general-purpose model enforcement stayed where they were.

The penalty structure is where the two approaches diverge most in tone. The regulation sets a top tier of €35 million or 7% of global annual turnover for prohibited practices, with the AI Office handling general-purpose model enforcement directly. The US outline talks about civil penalties and criminal referrals without naming a number.

Where the two regimes overlap

One regulates the document set. The other regulates the test. Both end up demanding paperwork that a compliance team has to produce, and neither recognises the other's output.

Take a mid-sized European company that runs a US frontier model through an API to process customer tickets. Its obligations today come entirely from the AI Act side. It needs a record of what the system does, a GDPR lawful basis for the data flowing through it, and Article 4 AI literacy for the staff using it. Whether Washington passes the SAFE AI Act changes none of that.

The US developers selling into Europe are already in scope as providers under the AI Act when their output is used in the EU. If the SAFE AI Act becomes law, evaluation reports produced for CAISI would be the obvious input for a European technical file. Nothing in the two-page outline creates mutual recognition, an adequacy mechanism, or a shared evaluation format. The realistic outcome is two dossiers in two languages for the same model.

What two pages do not answer

The document has no statutory definition of "frontier". It does not say which compute threshold or capability threshold triggers the obligation, so it is unclear whether a European lab such as Mistral or Aleph Alpha would fall in scope if it sold into the US market.

There is no timeline for how long a pre-deployment evaluation may take, no fee schedule for the pools that fund third-party evaluators, and no process for what happens when a developer disputes CAISI's minimum standards. "Early-stage startup" is not defined. The exemption could cover a five-person team or a 200-person company.

The outline is being circulated to committees and outside experts for feedback. No bill text, hearing date or vote has been published.

Does the SAFE AI Act apply to European AI companies?

The outline does not say. It is written in terms of "frontier AI developers" without a jurisdictional test, so it is not yet possible to tell whether a European lab serving US customers would be covered. That is one of the gaps the committee process would have to close before bill text appears.

Does the EU AI Act already require pre-release testing?

Not by a government team. For general-purpose models with systemic risk, the provider carries out the evaluation and reports serious incidents, and notified bodies assess high-risk systems against the regulation's requirements. The testing is done by the company or by an accredited body chosen within the EU framework, not by a federal agency with shutdown authority.

Is there anything to comply with right now?

No. The SAFE AI Act exists as a two-page outline with no legal effect. The obligations that apply to European deployments today come from the AI Act, GDPR and the transparency rules enforceable since August 2, 2026.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.