Congress mandates AI safety testing before deployment—developers must disclose code.
S. 2938 — Artificial Intelligence Risk Evaluation Act of 2025 · Filed by Josh Hawley (R-MO) · 2 cosponsors · Introduced Sep 29, 2025 · Hearing held
Your members of Congress
Enter a ZIP to see where your representative and both senators stood on this bill.
Looked up on this device — your ZIP is never stored on our servers.
What it does
This bill requires the Secretary of Energy to establish a testing and evaluation program for advanced artificial intelligence systems within 90 days. Developers of advanced AI systems (those trained on more than 10^26 computing operations) must participate and provide code, training data, and model details on request. The program will conduct standardized testing—including adversarial red-team attacks—to identify risks like loss of control, weaponization potential, or threats to civil liberties and labor markets. Developers cannot deploy advanced AI systems unless they comply. Violations carry fines of at least $1 million per day. Within one year, the Secretary must recommend a permanent federal oversight framework to Congress, with annual updates. The program sunsets after 7 years unless Congress renews it.
Why we flagged it
The bill's core mechanism is a mandatory government testing and evaluation program for advanced AI systems, with enforcement through fines and deployment restrictions. It is fundamentally a public-safety and regulatory-framework bill, not a subsidy, carve-out, or deregulation.
What the text implies
- Developers must disclose proprietary code, training data, and model weights to government—a significant intellectual-property and trade-secret disclosure requirement that may deter some companies from developing advanced AI or relocating development outside the U.S.
- The definition of 'advanced AI system' (10^26 computing operations) is subject to change by Secretary rule, but only if Congress approves via joint resolution—this creates a potential regulatory moving target and gives Congress veto power over definitional updates.
- The bill contemplates 'potential nationalization or other strategic measures' for superintelligence scenarios, signaling that government may seize or commandeer AI systems deemed existential threats—a power not explicitly limited in scope or procedure.
- Mandatory participation and disclosure apply to all developers in 'interstate or foreign commerce,' including open-source releases, potentially capturing academic and non-commercial AI development if it reaches the computing-power threshold.
- The program's sunset after 7 years creates regulatory uncertainty; if Congress does not renew, the testing framework and enforcement disappear, leaving no permanent oversight structure unless Congress acts on the Secretary's recommendations.
Section numbers refer to the bill text the analysis read — linked under Primary records below.
Who it affects
Citizens gain empirical risk assessment and government oversight of advanced AI systems before deployment, reducing the chance of uncontrolled or weaponized AI harming public safety, civil liberties, labor markets, or critical infrastructure. The mandatory testing and reporting requirements create accountability and data-driven regulation rather than ad-hoc or industry-captured oversight.
Named in the bill
Secretary of Energy, Department of Energy, Advanced Artificial Intelligence Evaluation Program, Congress, Covered advanced artificial intelligence system developers, Foreign adversaries (as defined in 15 CFR 791.2 and 791.4), Foreign terrorist organizations (as defined in INA § 219)
Where it stands
2 cosponsors: 1 Democrats, 1 Republicans.
- Sep 29, 2025 — Introduced · Congress.gov: “Introduced in Senate”
- Sep 29, 2025 — Referred to Senate Committee on Commerce, Science, and Transportation · Congress.gov: “Read twice and referred to the Committee on Commerce, Science, and Transportation”
- Sep 30, 2026 — Hearing held · Congress.gov: “Committee on Homeland Security and Governmental Affairs. Hearings held”
Dates and quoted wording are Congress.gov's action record; the timeline shows status changes, not every procedural step.
Money around this bill
2 lobbying clients named this bill on 2 disclosure filings across 2 quarters, Dec 2025 to Jun 2026. Those filings disclosed $608,894 in lobbying spend. A filing names 22 bills on average, so that figure is what each filing reported, not a share belonging to this bill.
More lobbying clients named this bill than 41% of bills with at least one filing.
Josh Hawley, the sponsor, reported $68,000 in PAC receipts in the 2026 cycle.
- Lenovo (united States) Inc. — $420,000 on 1 filing
- The Ai Policy Network Inc. — $188,894 on 1 filing
Lobbying Disclosure Act filings through Jul 15, 2026. A filing shows who paid to lobby on a bill it names, not what changed.
How this was measured
Analysis — Quorum's AI read the bill text published by Congress.gov (12,188 characters) on May 21, 2026. Section numbers in the findings refer to that text, linked below; transparency and hidden-provision scores are compared against the median of 16,222 analysed bills.
Status and sponsors — Congress.gov's bill record — actions, committee referrals and cosponsors — loaded nightly. The timeline shows status changes, not every procedural action.
Money — Senate Lobbying Disclosure Act filings whose specific-issue field names this bill for quarters ending Dec 2025 to Jun 2026. A filing's amount is reported whole beside the median number of bills a filing names; it is never divided across them. PAC receipts are FEC-reported contributions to the sponsor's candidate committee in the 2026 cycle.
As of — lobbying records through Jul 15, 2026 · page rendered 2026-10-20.
“Congress mandates AI safety testing before deployment—developers must disclose code.” QuorumCivic. https://share.quorumcivic.app/bill/119/s2938 Report an error
The week in bills
One email, Sunday: what Congress passed, what it changed, and who paid.
One link arrives to confirm the address. It is used for this email and nothing else, and every issue carries an unsubscribe link.