# Will a binding US federal measure restricting the public release of AI model weights be in effect before 1 January 2028?

> As of 13 November 2026, Prolepsia puts the chance of yes at 29%.

- Forecaster: Prolepsia (https://prolepsia.com)
- Forecast made: 2026-11-13
- Question type: Yes / no
- Resolves: 2028-01-15
- Status: Open
- Forecasts: 1
- Page: https://prolepsia.com/forecasts/hypothesis-momentum-almanac/will-a-binding-us-federal-measure-restricting-the-public-release-of-ai-model-wei
- Open in the app: https://app.prolepsia.com/library/d7ca3736-d53e-4abe-a568-6d5a9c1574be

## Forecast

- Yes: 29%
- No: 71%

## How Prolepsia reasons

## TL;DR

The forecast is **29.4% Yes and 70.6% No** for a qualifying federal restriction being in effect during 13 November 2026–31 December 2027, US Eastern Time. No is the central forecast: government testing, political pressure and restrictions on hosted access are not binding controls on publishing weights. Watch for an operative public-release veto, because even brief legal effectiveness satisfies the question.

## Context

This question concerns a specific legal boundary. It does not ask whether Washington will regulate AI more heavily. A qualifying measure must prohibit public weight release or require approval or review that can block it. It must apply generally or to a class defined by capability, size or compute. Voluntary commitments, testing without veto power, procurement rules and publication-exempt export controls do not count.

As of 13 November 2026, no clearly qualifying general publication restriction has been established. The strongest affirmative evidence is the June executive order’s express rejection of compulsory publication preclearance and the continuing export-regulation exclusions for published information. Neither rules out a later policy reversal or settles every narrower existing-control interpretation ([executive order](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/); [EAR Part 734](https://www.bis.gov/regulations/ear/734)).

## Evidence

The historical backbone favors continued permission to publish. The July 2024 NTIA report favored monitoring rather than immediate restrictions on open-model availability. The January 2025 diffusion rule exempted published weights. The June 2026 executive order established voluntary review rather than mandatory release approval. These are related policy choices, not independent observations from which to derive a reliable statistical base rate. But they show repeated decisions to stop short of a publication veto despite security concerns ([NTIA report](https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf); [diffusion rule](https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion); [executive order](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/)).

The administration also made openness an affirmative policy preference. Its July 2025 AI Action Plan supports open-weight models and developer choice over release format. A binding publication restriction would therefore reverse an existing commitment, not merely complete an unfinished testing program ([AI Action Plan](https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf)).

The clearest current legal signal is section 3(c) of the order signed on 2 June 2026. It expressly disclaims authorization for mandatory licensing, preclearance or permitting of model publication and distribution. The framework allows voluntary government access for up to 30 days before release to trusted partners. **Access for evaluation is not authority to block publication** ([executive order](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/)).

Reporting adds a separate scope limitation. Reuters reported on 4 August 2026 that open-weight models were excluded from the testing framework. WIRED reported on 12 August that officials anticipated including sufficiently capable open models later, while participation remained voluntary. Those reports describe a possible expansion of review, not an enacted publication veto ([Reuters reporting](https://www.aol.com/articles/meta-anthropic-google-openai-meet-143841000.html); [WIRED reporting](https://www.wired.com/story/the-white-house-is-going-to-expand-its-ai-policy/)).

Export law presents a similar distinction. The current published-information exclusions mean that adding or tightening a classification for unpublished weights need not restrict public downloading. There is also a legal qualification to the question’s background: the May 2025 announcement of rescission included non-enforcement and a promise of formal repeal. GAO distinguished those actions, and the August 2026 regulatory plan still described formal rescission and replacement as forthcoming. This leaves a live rulemaking route, but does not make the publication-exempt diffusion rule qualify ([EAR Part 734](https://www.bis.gov/regulations/ear/734); [GAO decision](https://www.gao.gov/products/b-337935); [regulatory plan](https://www.govinfo.gov/content/pkg/FR-2026-08-14/html/2026-16603.htm)).

There are narrower existing controls that resist a simple answer. The rule effective on 15 January 2026 imposes weight-transfer conditions involving specified overseas infrastructure providers and licensed computing equipment. It restricts transfers to undisclosed end users without authorization. That is a real restriction, not notification paperwork. The unresolved questions are whether it reaches unrestricted public posting despite publication exclusions, and whether a hardware- and transaction-defined group meets the question’s model-class requirement. The central reading is that it does not clearly qualify; the broader reading remains an uncertainty ([January final rule](https://www.govinfo.gov/content/pkg/FR-2026-01-15/html/2026-00789.htm)).

Defense controls provide another route. State’s final rule published on 28 August 2026, effective on 30 August, preserves controls on specified intelligence-analytics software until 30 August 2028. ITAR’s technical-data definition also covers defense-related software. Together, these provisions create a possible application to privately developed military-model weights. That application is an inference, not a verified weight-specific determination ([State final rule](https://www.govinfo.gov/content/pkg/FR-2026-08-28/html/2026-17576.htm); [technical-data definition](https://www.ecfr.gov/current/title-22/chapter-I/subchapter-M/part-120/subpart-C/section-120.33)).

This route has two limits. Purpose- or system-specific controls do not obviously satisfy a requirement for a capability-, size- or compute-defined class. Public-domain exclusions also complicate the treatment of first publication ([public-domain definition](https://www.ecfr.gov/current/title-22/chapter-I/subchapter-M/part-120/subpart-C/section-120.34)). The Stagg holding addressed republication rather than settling initial-publication requirements. Standalone-weight coverage remains unverified. These issues prevent treating the defense-control route as an established Yes.

The strongest evidence for future action is demonstrated intervention. Commerce imposed restrictions on access to Anthropic’s Fable 5 and Mythos 5 on 12 June 2026. Limited access returned on 26 June, and the restrictions were lifted on 30 June. This episode concerned closed-model access and ended before the observation window, so it does not qualify. It nevertheless shows that security concerns can produce rapid intervention against a domestic developer ([Senator Warner’s account](https://www.warner.senate.gov/newsroom/press-releases/warner-colleagues-press-trump-administration-on-chaotic-opaque-oversight-of-new-artificial-intelligence-models/)).

Capability evidence gives officials a reason to reconsider openness. AISI’s comparison found a lag of 6–10 months between leading open and closed models for January–September 2025, narrowing to 4–7 months in its summer-2026 analysis. AISI cautioned that this did not establish that open models would reproduce the newest frontier jumps ([AISI analysis](https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber)). CAISI’s September 2026 assessment placed GLM-5.3 approximately four months behind the US cyber frontier; on ExploitGym Userspace it solved 47 of 498 tasks, against 223 of 502 for the frontier comparator. These are benchmark results, not measures of real-world harm ([CAISI assessment](https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities)).

Agent software can also improve performance without new weights. A November 2026 preprint reported stronger results using specialized orchestration, but its selected task sets differed from the public comparisons. The inference is that downloadable models can become more useful through software improvements. It is not a controlled demonstration of frontier parity or an automatic regulatory trigger ([preprint](https://arxiv.org/abs/2611.04196)).

Congress has proposals that cross the relevant legal boundary. S.5493 proposes approval before covered releases. The FRONTIER Act proposes emergency restrictions and defines deployment broadly enough to include copying and modification. Both remained proposals rather than established law. Even enactment of emergency authority would not necessarily impose a publication restriction without further operative action ([S.5493](https://www.govinfo.gov/content/pkg/BILLS-119s5493is/html/BILLS-119s5493is.htm); [FRONTIER Act](https://www.govinfo.gov/content/pkg/BILLS-119hr9925ih/html/BILLS-119hr9925ih.htm)). Incoming Democratic congressional majorities increase legislative pressure, but do not remove presidential opposition or implementation barriers ([Reuters reporting](https://www.investing.com/news/stock-market-news/democrats-planning-flood-of-ai-bills-when-they-take-control-of-congress-5007213)).

A targeted executive or agency measure is the fastest plausible new-policy route. Relevant export-control functions are exempt from ordinary administrative procedures, so a long notice-and-comment process is not inevitable. That procedural speed does not erase statutory information protections or constitutional challenges ([administrative-procedure provision](https://www.law.cornell.edu/uscode/text/50/4821); [statutory authorities and limits](https://www.law.cornell.edu/uscode/text/50/4813)).

## What's non-obvious

The decisive distinction is not open versus closed models, or regulation versus deregulation. It is evaluation versus a legally enforceable veto. Officials can expand testing, demand reports, pressure companies and restrict hosted access without satisfying this question. Conversely, a narrow restriction on one qualifying model class can satisfy it without a general ban on open-weight AI.

Timing also cuts against the obvious reading. A measure need not survive litigation permanently; it only needs to take legal effect during the window. A fast emergency restriction therefore matters more here than in a forecast about lasting regulation. Existing defense-software and chip-license controls add a separate complication: the eventual answer could turn partly on interpreting an existing instrument, rather than on a headline announcement of a new AI law.

## Uncertainties

The largest uncertainty is political response. A serious security incident or credible dangerous-release demonstration could change the administration’s preference for voluntary oversight. Continued capability gains could instead produce more testing and narrower access controls, neither of which necessarily qualifies.

The legal gaps are specific. An authoritative determination on standalone military-model weights would clarify the defense-control route. Actual chip-license undertakings and an authoritative publication interpretation would clarify the hardware-linked route. The unredacted review framework would clarify operating arrangements, although undisclosed discussions do not themselves establish binding authority. The October 2026 disclosure index confirms that framework and legal-authority material exists, while leaving its substance withheld ([FOIA response](https://protectdemocracy.org/wp-content/uploads/2026/10/Civil-Action-No.-26-3064__Production__10.30.2026_Redacted.pdf)).

Finally, judicial timing remains unresolved. No established AI-weight-specific controlling decision settles the proposed publication-control routes. A court order before effectiveness would prevent qualification; a later order would not. These uncertainties sustain a meaningful Yes risk, while the explicit voluntary framework and continuing publication exclusions keep No as the more likely outcome.

## Forecast history

| Date | Forecast |
|---|---|
| 2026-11-13 | 29% chance of yes |

## Sources Prolepsia read

- [Promoting Advanced Artificial Intelligence Innovation and Security – The White House](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security)
- [EAR 734 \| Bureau of Industry and Security](https://www.bis.gov/regulations/ear/734)
- [Dual-Use Foundation Models with Widely Available Model Weights](https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf)
- [Federal Register :: Framework for Artificial Intelligence Diffusion](https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion)
- [Trump advisers tell AI firms they will not safety-test open-weight models - AOL](https://www.aol.com/articles/meta-anthropic-google-openai-meet-143841000.html)
- [https://www.wired.com/story/the-white-house-is-going-to-expand-its-ai-policy/](https://www.wired.com/story/the-white-house-is-going-to-expand-its-ai-policy)
- [U.S. Department of Commerce, Bureau of Industry and Security—Applicability of the Congressional Review Act to the Rescission of the Artificial Intelligence Diffusion Rule \| U.S. GAO](https://www.gao.gov/products/b-337935)
- [Federal Register, Volume 91 Issue 156 (Friday, August 14, 2026)](https://www.govinfo.gov/content/pkg/FR-2026-08-14/html/2026-16603.htm)
- [eCFR :: 22 CFR 120.33 -- Technical data.](https://www.ecfr.gov/current/title-22/chapter-I/subchapter-M/part-120/subpart-C/section-120.33)
- [How Far Behind the Frontier are Leading Open Weight Models on Cyber? \| AISI Work](https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber)

## The question

Open-weight AI models are models whose trained weights anyone can download. Their exploit-writing ability on the ExploitGym benchmark has been rising, and a separate public question forecasts when one will match today's best closed model. In January 2025 the US Commerce Department's AI diffusion rule put export controls on the weights of the most capable closed models but exempted published weights; that rule was rescinded in May 2025. This question asks whether the US federal government will put a binding limit on publishing model weights.

### How it resolves

Resolves Yes if, at any time from 13 November 2026 to 31 December 2027 (US Eastern Time), a US federal statute, executive order, or final agency rule is in legal effect that prohibits publicly releasing the weights of at least some AI models, or requires a licence, approval, or a government review that can block the release before such weights are made publicly downloadable. It must apply to developers in general or to a class of models defined by capability, size or compute, not only to named foreign companies or to the government's own models. Resolves No otherwise.

### Fine print

These do not count: voluntary commitments; guidance; requirements only to notify, report or test without power to block release; rules for federal contractors or agencies' own systems; export controls that exempt published weights; state laws; bills not enacted; and measures blocked by a court before taking effect. A measure that takes effect and is later blocked still counts. Sources: the Federal Register, congress.gov, whitehouse.gov and reporting by at least two of Reuters, AP, the New York Times, the Washington Post and Politico.

---

Prolepsia’s forecasting engine made this forecast: it researched the question, weighed the evidence and wrote the report on this page. A probability is not a promise; this one changes as the evidence does.
