Three members of the US House of Representatives have released an outline of forthcoming legislation they call the Safe and Accountable Federal Evaluations of Artificial Intelligence Act, or the SAFE AI Act. In a statement dated 19 November 2026, Representatives Sara Jacobs, a Democrat from California, Celeste Maloy, a Republican from Utah, and Don Beyer, a Democrat from Virginia, described it as “forthcoming bipartisan legislation that would create federal safety requirements for frontier artificial intelligence.”
It is an outline, not a bill that has been filed, and it is not law. Beyer, in the same remarks, said the sponsors “are continuing to seek input on our forthcoming legislative text from experts and safety leaders, in coordination with committees of jurisdiction,” and that they were sharing “this blueprint today to show the direction our work is taking.”
The SAFE AI Act outline would set a federal floor for frontier developers
The outline itself, posted by Beyer’s office, starts from a claim about the status quo: companies building frontier AI systems “create their own safety rules with no clear guidance from the federal government,” and “no one outside the company can reliably verify whether the rules are actually being followed.” The proposal would require frontier developers to “publish enforceable safety frameworks” that set out what each company is testing for, what thresholds would trigger mitigations, and what those mitigations involve. Those frameworks “cannot be unilaterally loosened.”
Minimum content for the frameworks would not be left entirely to the firms. The outline says the Center for AI Standards and Innovation (CAISI), “in collaboration with interagency partners including the intelligence community and the DOE, [would] write minimum requirements for safety frameworks informed by technical experts, standards, public input, and best practices.” States, it adds, “can shape regulation above the federal floor.” Jacobs put the same idea more sharply: “Our proposal is a floor – not a ceiling, empowering states to adopt stronger protections for their communities.”
Coverage is deliberately limited. To “avoid stifling innovation,” the outline says, “startups and academics are not generally covered.” Frontier developers would have to “disclose substantially autonomous AI R&D.” Any person could “confidentially report a safety incident to CAISI,” and developers would have to report safety incidents; evaluators and developers would also have to report an “imminent risk of severe harm.”
Independent evaluators, incident reporting and a narrow emergency power
The second block of the outline is about who checks the frameworks. Government teams would examine them for compliance. Third-party verifiers would be “accredited by the government,” assigned by CAISI “with mandatory rotations,” and, under the outline, developers would not pay evaluators directly. Conflict-of-interest rules would restrict “revolving-door talent between developers and accredited independent evaluators.” Evaluations would take place “before commercial release or broad internal use, including automated AI R&D.”
That design is a pointed contrast with company-run pilots in which a lab chooses and pays its own reviewers. OpenAI’s evaluator pilots with Apollo Research and METR are the closest recent example of that voluntary model; the outline would put assignment, accreditation and payment on a different footing.
Enforcement, in the outline, includes “narrow emergency authority with judicial review”: where safeguards are inadequate and a system “poses imminent risks of severe harm, the government may order mitigation measures subject to prompt judicial review.” There would be civil penalties, “enhanced penalties for knowing or willful violations,” procedures for criminal referral, a clarification of “liability of developers for reasonably foreseeable injuries to third parties,” and “whistleblower protection.” Jacobs said the measure would include “civil penalties for violations and targeted emergency authority to stop dangerous models when they pose a serious and immediate danger.”
Maloy, whose remarks appear in Jacobs’s release, said: “The companies building the most powerful AI systems ought to be able to show they’ve done their homework before those systems go into wide use.” She added that the proposal “leaves room for Utah and other states to set safeguards of their own.”
An outline is not a bill, and the hard fights are still ahead
The honest way to read 19 November is as a political and technical preview, not as a change in the law. No statutory text has been introduced. No committee has marked anything up. Beyer’s own comment is that the sponsors are still collecting views on the legislative language. Until that text exists, the SAFE AI Act outline’s most important sentences are the ones that sketch a fight rather than settle it.
One fight already visible in the SAFE AI Act outline is federalism. A “floor, not a ceiling” is an answer to the argument that Washington should pre-empt state AI rules. It is also an invitation to the states that have already moved. New York’s attorney general has just stood up an ASSET unit to enforce the state’s RAISE Act, which already requires frontier safety frameworks. A California Government Operations Agency report has proposed shutdown powers and caps on recursive self-improvement. The outline would have CAISI set a national minimum and then leave those capitals free to go further. Whether a later draft still says that, once industry lobbying and committee jurisdiction enter the room, is the first thing to watch.
A second fight is independence. Direct payment from developers to evaluators is ruled out, and “mandatory rotations” are an attempt to fix a conflict that voluntary lab-sponsored evaluations cannot quite escape. The outline still has to define who counts as a frontier developer, what “substantially autonomous AI R&D” means, and how a “narrow” emergency power with judicial review would work in hours rather than months. Those are the clauses that turn a blueprint into something a floor manager can actually bring up. Until they are written, the SAFE AI Act is a direction of travel, and no more.
Sources
- Office of Representative Sara Jacobs, “Jacobs, Maloy, Beyer Release Outline For Sweeping, Bipartisan AI Safety Legislation,” 19 November 2026
- Office of Representative Don Beyer, press release on the SAFE AI Act outline, Document ID 9310
- SAFE AI Act outline PDF, offices of Representatives Jacobs, Maloy and Beyer
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- New York AG Creates ASSET Unit to Enforce the RAISE Act is the closest live TSN account of a US state already enforcing frontier-model safety frameworks, and the right next stop for the “floor, not a ceiling” claim in this outline.
- California Report Proposes Shutdown Powers and Caps on RSI
- OpenAI Evaluator Pilots Start With Apollo and METR
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X


