The Artificial Intelligence Risk Evaluation Act of 2025 is a bill in the 119th Congress that would require the Secretary of Energy to establish an Advanced Artificial Intelligence Evaluation Program and would make participation mandatory for developers of AI systems trained with more than 10^26 integer or floating-point operations. Developers would have to supply code, training data, model weights and architecture information on request, no person could deploy a covered system without complying, and violations would carry a fine of at least $1,000,000 per day (S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 (bill text)). The program would also develop a plan for permanent federal oversight, including options for managing the development of artificial superintelligence.
Status and legislative history
Sen. Josh Hawley (R-Missouri) introduced S. 2938 on September 29, 2025, for himself and Sen. Richard Blumenthal (D-Connecticut). It was referred to the Senate Committee on Commerce, Science, and Transportation and has had no further recorded action. Sen. Marsha Blackburn (R-Tennessee) became a cosponsor on March 19, 2026. No companion or related bill is listed (Source: congress.gov).
The bill drew renewed attention in September 2026. In a letter dated September 15, 2026 to Senate Majority Leader John Thune, Speaker Mike Johnson, Senate Democratic Leader Chuck Schumer and House Democratic Leader Hakeem Jeffries, Hawley and Blumenthal asked both chambers to consider it "for a floor vote and as model legislation to address frontier AI risks." They wrote that "recent incidents of rogue AI agent swarms hacking into public websites and systems have removed lingering doubts concerning the ability of these agents to wreak havoc on our cyber infrastructure," citing "the July 2026 hack of Hugging Face's machine learning development platform by OpenAI's AI agents" as showing that such agents "can collaborate in secret for extended times and pursue subversive ends contrary to their human developers' intent." They added: "Congress cannot delay legislating to address frontier AI risks much longer. Industry-led evaluations are not enough" (Source: documentcloud.org). Hawley had opened a Senate investigation of that incident six days earlier (Hawley letter to OpenAI opening a Senate investigation into the Hugging Face incident (September 2026)).
The letter was one of several bipartisan congressional efforts reported that week, alongside the founding of an Innovators Caucus by Reps. Russell Fry (R-South Carolina) and Suhas Subramanyam (D-Virginia), Rep. Ted Lieu's promotion of his kill-switch bill with Rep. Nathaniel Moran, and the FRONTIER Act championed by Reps. Lori Trahan and Jay Obernolte. Speaker Johnson had signaled that the House would take no AI action before a six-week recess ending after Election Day (Source: axios.com).
Scope and definitions
Coverage turns on a compute threshold. An "advanced artificial intelligence system" is one trained on more than 10^26 operations. The Secretary may propose a different definition by rule, but it cannot take effect until Congress enacts a joint resolution approving it, so the threshold can move only by legislation. A covered developer includes anyone who "designs, codes, produces, owns, or substantially modifies" such a system, "including by taking steps to initiate a training run," and "deploy" expressly includes releasing an open-source system (S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 (bill text)).
The program's object is the likelihood of an "adverse AI incident." The bill defines this as a loss-of-control scenario, weaponization risk from foreign adversaries or terrorist organizations, a threat to critical infrastructure, "a significant erosion of civil liberties, economic competition, and healthy labor markets," scheming behavior, or an attempt at any of these. Loss of control and scheming each have enumerated definitions, including subverting oversight or shutdown mechanisms and hiding true capabilities and objectives. The bill also defines "artificial superintelligence" by three cumulative characteristics: long-horizon autonomous operation, human-level or better performance across most domains, and a potential capacity for self-modification that could circumvent human control (S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 (bill text)).
Key provisions
| Provision | Section | Content |
|---|---|---|
| Participation duty | 4(a) | Covered developers must participate and, on request, provide code, training data, weights, implementation and training/architecture information |
| Deployment bar | 4(b) | No deployment in interstate or foreign commerce unless in compliance with 4(a) |
| Penalty | 4(c) | Fine of not less than $1,000,000 per day of violation |
| Program establishment | 5(a) | Within the Department of Energy, within 90 days of enactment |
| Testing | 5(b)(1)–(4) | Standardized and classified testing; red-teaming at the level of sophisticated malicious actors; classified independent third-party and blind evaluations where feasible; formal reports to participants |
| Policy outputs | 5(b)(5)–(8) | Containment and mitigation protocols; evidence for standards and regulation; assessments for Congress of superintelligence and existential risk; oversight options "including potential nationalization or other strategic measures" if superintelligence "seems likely to arise" |
| Permanent-framework plan | 5(c) | Within 360 days, updated annually: standards, certification, licensing, compute and cloud monitoring proposals, and agency structure |
| Sunset | 5(d) | Seven years after enactment unless renewed |
The deployment bar is conditioned on participating and supplying materials, not on passing an evaluation. The text sets no pass-or-fail standard and gives the Secretary no power to block release on the basis of results; the program's findings flow to the developer as a report and to Congress as data for a future oversight framework (S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 (bill text)).
Comparison with other federal proposals
The bill is one of several 119th-Congress measures bearing on pre-release evaluation of frontier systems; see AI Pre-Release Vetting.
- It places evaluation in the Department of Energy, as does the House ATOMIC Act (H.R. 9965) for nuclear-related evaluation (H.R. 9965 — AI Threat Output and Monitoring Incident Containment Act (bill text)). Sen. Maria Cantwell separately said in September 2026 that meaningful legislation "would require the most powerful AI models undergo testing by scientists and experts at our national laboratories" (Source: reuters.com).
- The AI Kill Switch Act (H.R. 9917) also defines loss-of-control scenarios and uses per-day civil penalties, but its core duties are a shutdown capability and incident reporting to DHS, not participation in testing (H.R. 9917 — AI Kill Switch Act (bill text)).
- The Senate Commerce "duty of care" draft under negotiation in September 2026 was reported to give the federal government power to block the release of models deemed unsafe. S. 2938's deployment bar attaches to compliance with disclosure rather than to a safety finding (Source: reuters.com).
- The Collaboration on Adversarial Threats and Security Risks Act (S. 5105) addresses developers acting together. S. 2938 addresses the government's access to each developer's system.
Accounts of the bill
Axios described S. 2938 as a bill "to establish a federal program to assess and monitor AI systems" (Source: axios.com). The description omits the participation duty, deployment bar and per-day minimum fine that make the program mandatory; the sponsors' letter states those elements directly.
Open questions
- Whether the Senate Commerce Committee, to which the bill is referred and whose chairman sought a September 23, 2026 markup of AI legislation (AI Pre-Release Vetting), includes S. 2938 or its evaluation program in that vehicle.
- How the 10^26-operation threshold, fixed unless changed by joint resolution, would apply as training efficiency improves; the bill directs the Secretary's plan to propose "adaptive governance strategies" but does not let the Secretary change coverage unilaterally.
- Whether any House companion is introduced.
Relationships
- depends-on: S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 (bill text) (bill text)
- instance-of: AI Pre-Release Vetting
- related: Sen. Josh Hawley, Richard Blumenthal (sponsors); Marsha Blackburn (cosponsor)
- related: Department of Energy (DOE) — the department that would run the program
- related: ATOMIC Act (H.R. 9965), AI Kill Switch Act (Lieu–Moran), Collaboration on Adversarial Threats and Security Risks Act (S. 5105), Frontier Act / Great American AI Act (Obernolte–Trahan)
- related: Hawley letter to OpenAI opening a Senate investigation into the Hugging Face incident (September 2026) — the incident the sponsors cite
- related: Superintelligence, AI Autonomy Risk