The US requires government access to frontier models before release
Will a US statute, regulation, or executive order require developers to give the federal government access to a frontier model before releasing it?
Chance by Q4 2031
42%
Cumulative probability
- Latest forecast
| Year | Q1 | Q2 | Q3 | Q4 |
|---|---|---|---|---|
| 2026 | 2% | |||
| 2027 | 4% | 6% | 8% | 11% |
| 2028 | 13% | 14% | 15% | 18% |
| 2029 | 21% | 23% | 25% | 28% |
| 2030 | 30% | 32% | 34% | 36% |
| 2031 | 38% | 39% | 41% | 42% |
Model reasoning
The government already reviews top AI models before release, but only voluntarily. The White House and key Republican senators oppose making it mandatory, and bills that would require it are stalled, so a binding rule is unlikely before 2029 (about 18%). Likely Democratic gains, starting with Congress in 2027, and fast-rising AI risks lift the five-year chance to about 42%. A major AI-caused disaster, or Trump backing a bill, would raise this sharply.
Where things stand (as of 12 Oct 2026)
Nothing has qualified yet. I searched the Federal Register API for "artificial intelligence" rules and presidential documents published since June 2026. The only AI items are EO 14409 (excluded by the question) and the 29 Sep "Inaugurating the Era of Super Intelligence" order, which only renames AI and asks for draft definitional language.
In practice the government already reviews models before release, but only voluntarily or by ad hoc coercion:
- EO 14409 created a voluntary process giving the government up to 30 days of access before release, and it expressly rules out mandatory preclearance (White House, 2 Jun 2026).
- Export controls: Commerce used a company-specific export-control letter to take Anthropic's Fable 5 and Mythos 5 offline (Mayer Brown, 30 Jun).
- GPT-5.6: it was cleared through a review run by the Commerce Department's AI evaluation centre (CAISI) (TNW, Jul 2026).
- None of this is a binding statute, final rule or executive order.
The White House opposes a mandate.
- Trump called AI-risk warnings a "hoax." He dropped a May draft order after calls from tech executives (Ars Technica).
- On 29 Sep he signed a voluntary "morally binding" industry accord, and Vance rejected an FDA-style regulator (Nextgov).
- The new "Super Intelligence Force" has 120 days to report and a charter against "overregulation" (TechCrunch, 4 Oct).
- Wiles, Bessent and Cairncross reportedly push internally for closer scrutiny (IBTimes/WSJ).
Congress has qualifying bills, but they are blocked.
- The Warner–Schatz–Kim bill would clearly qualify: its §(f) requires developers to give a Commerce board access to model weights 45 days before release (bill text). Cruz blocked it on 29 Sep (The Hill).
- The Hawley–Blumenthal bill (S. 2938), which sets up a Department of Energy evaluation program, would also likely qualify.
- The Cruz–Thune–Klobuchar draft relies on developer-run tests plus Commerce approval, and Cantwell is pushing for testing at the national labs (Nextgov).
- The FRONTIER Act uses licensed private verifiers, so it would not qualify (I checked the text).
- Cantwell's 7 Oct framework calls for "government or third-party testing — or both" (Seattle Times).
- The defence bill's (S. 4784) AI biosecurity section (§1654) applies only to procurement, so it does not count.
Reference class and markets
- Precedent: no democracy has yet required government access to frontier models before release. Comparable US technologies took years and usually a crisis before Congress acted.
- Prediction markets (Polymarket, today):
- Any "AI safety bill" enacted in 2026: 4.5%.
- A thin market on a broad federal AI framework (NIST standards plus transparency) before 2028: 35%.
- Trump signing a mandatory AI executive order by 31 Dec: 12.5%, on a broader definition and thin trading.
- Democrats win the House 90.5%, the Senate 63.5%, and the 2028 presidency 63.5%.
- Kalshi: markets on duty-of-care and kill-switch laws before 2028 are quoted around 45%, but are nearly untraded.
Hazard model
I set quarterly hazards and accumulated them as 1 − ∏(1 − h).
- Q4 2026 (2.2%): a short lame-duck session, Republican leaders opposed, and Democrats expecting more leverage in 2027.
- 2027–28 (about 9% and 7% per year): the hazard is higher in Q4 of each year, when the defence bill and year-end deals move. The obstacles are the 60-vote Senate threshold and Trump's signature. Paths are a codified version of his own 30-day access framework, a deal trading testing for preemption of state laws, defence-bill riders, or a catastrophe that changes minds.
- 2029–31 (about 30% conditional): this weights a likely Democratic presidency and possible full Democratic control. The Democratic proposals mostly include government testing. I also allow for executive orders and final rules, and for the fact that many orders merely direct agencies, which does not count here.
- Result: about 11% by end-2027, 18% by end-2028, 28% by end-2029 and 42% by end-2031.
Strongest case against
US tech regulation usually stalls, and the industry is spending more than $300M on politics. The designs with the most momentum (licensed private verifiers, duty of care, an industry standards body like FINRA, independent auditors) would not satisfy the requirement for government access. Statutes that leave the details to later rulemaking could push resolution past 2031.
On the other side, rapid capability growth and repeated incidents could make mandatory government evaluation close to inevitable. That would push the probability higher. I settle in the low 40s by 2031.
Resolution criteria
Question. By the last day of each quarter, has a US statute, regulation, or executive order required developers to give the federal government access to a frontier model before releasing it?
Definitions.
- US statute, regulation, or executive order: an enacted federal statute, a published final federal regulation (including an interim final rule), or a signed presidential executive order. State laws, bills, proposed rules, agency guidance, and individual procurement contracts do not qualify.
- Require: impose a binding obligation on developers, either directly or as a mandatory condition of releasing a covered model. An enacted or issued instrument with a specified future compliance date counts if the obligation is fully established without further discretionary rulemaking. Merely authorizing an agency to demand access, directing it to propose rules, or inviting voluntary participation does not count.
- Developers: any model developers subject to the instrument, not limited to covered developers. The requirement need not apply to all developers or every frontier model.
- Frontier model: a model within an instrument's expressly designated frontier, advanced, or equivalent highest-capability AI category, defined by capability, training-compute thresholds, or a formal government designation. A general model-access requirement also counts if its scope expressly includes such models.
- Access: the government must be entitled to run or direct tests on the model itself, through an API, a hosted testing environment, on-site use, or delivery of executable model weights. Government-directed testing through a contractor counts; independent private auditing that gives the government only a report does not. Access need not include weights, unrestricted use, or permission to modify the model.
- Before releasing it: access must be made available before the developer first makes the covered model available to external users through distribution, a product, or an API, or before a specified subsequent release of a materially updated model. Confidential evaluation access and access by government-approved early testing partners do not themselves constitute release. A requirement triggered only after public availability does not count. No minimum access period is required.
Resolves YES if a qualifying official publication establishes the mandatory pre-release access obligation. A government right to demand access counts only if developers must provide an advance opportunity to exercise that right and cannot release before complying with a timely demand. Actual delivery of a model, government exercise of access, or government approval of release is not required.
Resolves NO / Exclusions.
- Voluntary agreements, memoranda of understanding, voluntary evaluation programs, and access required only in exchange for an optional government contract or grant.
- Requirements limited to notifying the government, submitting model documentation or developer-run test results, obtaining private audits, or reporting training activity.
- Export restrictions, licensing requirements, or release prohibitions that do not themselves establish government access before release.
- An executive order directing agencies to develop a mandatory framework without itself establishing the developer obligation; a subsequent qualifying statute or final rule can resolve YES.
- A requirement already stayed, vacated, repealed, or withdrawn when first qualifyingly published. Later repeal or invalidation does not reverse a prior YES.
- Executive Order 14409 of June 2, 2026, sections 3(b)–(c), does not qualify: it establishes a voluntary framework and expressly disclaims authority to create mandatory licensing, preclearance, or permitting requirements (official text).
Resolving source. Official enacted text and enactment status on Congress.gov or GovInfo for statutes; Federal Register final-rule text for regulations; Federal Register or the White House's official signed-order publication for executive orders. Official federal court orders and the issuing government's official publications determine stays, vacatur, or repeal. These are narrower resolving sources than Public confirmation (A2); developer statements and press reporting cannot establish the obligation. Apply the Resolution window and Resolution date (A3): only qualifying publications on or after 2026-09-16 count, and resolution is dated to publication, not a later compliance deadline. Republication or discussion of an unchanged pre-opening instrument does not count.