US federal frontier-AI regulation
How far will binding federal regulation of frontier developers go?
Chance by Q3 2031
- Transparency
- 75%
- Independent audits
- 57%
- Pre-deployment evaluation
- 45%
- Halt authority
- 39%
Cumulative probability
- ≥ L1 · Transparency
- ≥ L2 · Independent audits
- ≥ L3 · Pre-deployment evaluation
- ≥ L4 · Halt authority
| Quarter | L1 | L2 | L3 | L4 |
|---|---|---|---|---|
| Q3 2026 | 1% | <1% | <1% | <1% |
| Q4 2026 | 13% | 8% | 6% | 5% |
| Q1 2027 | 17% | 10% | 7% | 6% |
| Q2 2027 | 22% | 13% | 10% | 8% |
| Q3 2027 | 26% | 17% | 12% | 10% |
| Q4 2027 | 32% | 21% | 15% | 13% |
| Q1 2028 | 34% | 23% | 17% | 14% |
| Q2 2028 | 37% | 25% | 19% | 16% |
| Q3 2028 | 40% | 27% | 20% | 17% |
| Q4 2028 | 45% | 31% | 23% | 19% |
| Q1 2029 | 49% | 34% | 26% | 21% |
| Q2 2029 | 53% | 37% | 28% | 24% |
| Q3 2029 | 57% | 39% | 31% | 26% |
| Q4 2029 | 60% | 43% | 33% | 28% |
| Q1 2030 | 63% | 45% | 35% | 29% |
| Q2 2030 | 65% | 47% | 37% | 31% |
| Q3 2030 | 67% | 49% | 39% | 33% |
| Q4 2030 | 70% | 52% | 41% | 35% |
| Q1 2031 | 71% | 53% | 42% | 36% |
| Q2 2031 | 73% | 55% | 44% | 37% |
| Q3 2031 | 75% | 57% | 45% | 39% |
Model reasoning
The question asks how far the US federal government has actually gone in binding frontier AI developers — not what's been proposed, but what's been signed into law, finalized as a rule, or ordered in a way that's enforceable against companies like OpenAI, Anthropic, and Google DeepMind. Today the answer is nothing: level zero. We put the chance of mandatory transparency with real penalties arriving by the end of 2026 at 13%, rising to 75% by September 2031. The strongest rung — a federal body empowered to order a halt — sits at 5% by end-2026 and 39% by 2031.
Nothing on the books qualifies. The June 2026 executive order on frontier models creates a pre-release review process, but it's voluntary and says on its face that it doesn't authorize licensing or preclearance. Biden's 2023 order was revoked, and its reporting requirements never became a final rule anyway. The export-control actions that briefly forced Anthropic to suspend two models were company-specific letters under existing authority, not a new instrument covering all the major labs. What exists instead is an unusually crowded pipeline: the FRONTIER Act in the House, which would require published safety frameworks, incident reporting with penalties, recurring independent verification, and emergency suspension authority; and a Senate negotiation involving the Majority Leader and the Commerce chair around a duty of care and power to block unsafe releases.
Two forces push the number up. Frontier labs are now actively asking for federal rules, largely because they want state laws preempted — and preemption is the bargaining chip that could buy mandatory obligations. And a run of 2026 incidents, plus high-profile warnings from inside the labs, has made this politically live in a way it wasn't before.
Holding it down: Congress's record on regulating digital technology is close to zero over thirty years. The President dismisses the risk and frames AI as a race with China; the Speaker says developers should handle safety themselves; the filibuster and the preemption fight both cut against a deal. The calendar is brutal — very little floor time before the midterms.
Notice how close L4 tracks L1. That's deliberate: the leading bills bundle halt authority together with transparency, so if anything passes, it likely passes with teeth. The biggest open question is whether the eventual compromise keeps penalties at all, or dissolves into voluntary certification plus preemption — which would satisfy none of these levels.
Watch for published Senate text with a scheduled markup, and for any shift in the President's posture. A confirmed mass-casualty incident would move everything at once.
Resolution criteria
Question. For each quarter, what is the highest level L such that an instrument at level ≥ L has been enacted on or before the last day of the quarter?
Instrument. A federal statute signed into law, or a final agency rule with force of law. An executive order counts only if it imposes obligations enforceable against private developers other than as a condition of federal procurement or participation in a voluntary program. Introduced bills, discussion drafts, committee passage, and passage by one chamber do not count. Resolve on enactment date; annotate the effective date. Repeal does not un-resolve.
Frontier developer. Defined by the instrument's own threshold, provided that threshold captures at least OpenAI, Anthropic, and Google DeepMind as of enactment.
Levels (cumulative).
- L0 — None.
- L1 — Mandatory transparency: publish a safety framework and/or report critical incidents to a federal body, with penalties for non-compliance. Codification of a voluntary framework without penalties does not count.
- L2 — Mandatory recurring independent audit or third-party verification of compliance or capabilities.
- L3 — Government or certified third-party capability evaluation required before deployment, with deployment conditioned on it.
- L4 — A federal body has authority to order a halt or pause of development or deployment based on capability or risk findings.
Resolving source. congress.gov (enrolled bill / public law), federalregister.gov (final rule), whitehouse.gov (executive order).
Forecast horizons. Every quarter-end from 2026-09-30 through 2031-09-30.