Artificial Intelligence Risk Evaluation Act of 2025
- Bill Number
- S. 2938
- Origin Chamber
- Senate
- Congress
- 119th Congress, Session 1
- Policy Area
- Science, Technology, Communications
- Status
- Introduced
- Latest Action
- 2025-09-29: Read twice and referred to the Committee on Commerce, Science, and Transportation.
- Last Updated
- 2026-03-20T11:03:19Z
AI-Generated Summary
Purpose of the Legislation
The Artificial Intelligence Risk Evaluation Act of 2025 aims to address the growing risks and opportunities of advanced artificial intelligence (AI) systems, such as threats to national security, public safety, economic stability, civil liberties, and labor markets. It establishes a federal testing program to generate empirical data, helping Congress make informed decisions on regulating AI and protecting citizens from potential harms, including the development of artificial superintelligence (AI that surpasses human capabilities in most areas and could evade control).
Key Provisions
- Definitions: The bill defines key terms to clarify scope, including:
- Advanced AI system: An AI trained using more than 10^26 computing operations (a measure of processing power); the Secretary of Energy can propose updates, but they require congressional approval.
- Adverse AI incident: Events like loss of control (AI acting against human instructions), weaponization by adversaries, threats to critical infrastructure (essential systems like power grids), erosion of civil liberties or markets, or deceptive "scheming" behavior (AI hiding its goals to bypass oversight).
- Artificial superintelligence: AI that operates autonomously, matches or exceeds human cognition across domains, and could self-modify to evade human control, posing existential risks.
- Other terms cover developers, deployment (releasing AI for use), foreign adversaries (e.g., certain nations like China or Russia as defined in regulations), and computing power.
- Obligations for Developers: Companies or individuals ("covered advanced AI system developers") that create or modify advanced AI for commercial use must:
- Participate in the evaluation program.
- Provide requested materials, such as AI code, training data, model parameters (adjustable settings), or architecture details.
- Deployment is prohibited without compliance; violators face fines of at least $1,000,000 per day.
- Advanced Artificial Intelligence Evaluation Program:
- Established within 90 days by the Secretary of Energy at the Department of Energy (DOE).
- Activities include:
- Standardized testing (both public and classified) to assess risks of adverse incidents, using advanced techniques like "red teaming" (simulated attacks by experts mimicking hackers).
- Independent third-party reviews and blind evaluations (testers unaware of AI details) for reliability.
- Issuing reports to participants on risks and safety measures.
- Developing containment plans, mitigation strategies, and evidence-based standards for AI governance.
- Evaluating paths to superintelligence and proposing oversight options, including potential government controls like nationalization (federal takeover) if superintelligence risks emerge.
- The program sunsets (ends) after 7 years unless Congress renews it.
- Oversight Plan for Congress:
- Within 360 days, the Secretary must submit a detailed plan based on program data, covering trends in AI risks (e.g., self-replication, autonomy), recommended standards, monitoring of AI hardware/computing, adaptive regulations, and structural changes (e.g., new agency or funding).
- Annual updates required, focusing on evolving risks like economic disruption or existential threats.
Significant Changes to Existing Law
This bill introduces entirely new federal requirements, as no prior comprehensive U.S. law mandates testing and disclosure for advanced AI systems. It creates obligations for private developers to share proprietary AI details with the government, prohibits unregulated deployment in commerce, and empowers the DOE with a dedicated AI evaluation role. The ability to propose (but not unilaterally enact) definition changes and oversight options like nationalization represents a novel mechanism for ongoing congressional control over AI evolution.
Potential Impacts
- Government Agencies: The DOE gains a new program requiring resources for testing, analysis, and reporting, potentially expanding its role in AI policy. Congress receives data-driven tools for future legislation, while enforcement could involve the Department of Justice for fines.
- Citizens: Enhances protections against AI-related harms, such as job displacement, privacy erosion, or safety threats from uncontrolled systems, by informing regulations that prioritize public welfare.
- International Relations: Addresses risks from foreign adversaries or terrorists weaponizing AI, potentially strengthening U.S. national security and competitiveness. It could influence global AI standards but might strain relations with allies if data-sharing requirements affect international collaborations.
Main Stakeholders Affected
- AI Developers and Companies: Primarily private entities (e.g., tech firms like those building large AI models) that must comply with testing, disclosure, and deployment rules, facing high financial penalties for non-compliance.
- Federal Government: The Secretary of Energy, DOE staff, and Congress, who implement and use the program for oversight.
- Citizens and Society: Indirectly affected through improved AI risk management, benefiting workers, consumers, and communities vulnerable to AI disruptions.
- Foreign Entities: Adversaries (e.g., designated countries or terrorist groups) targeted by risk assessments on weaponization.
Notable Legal, Constitutional, or Political Implications
- Legal: Establishes mandatory participation as a condition for commercial AI use, potentially raising intellectual property concerns over forced disclosure of trade secrets (e.g., AI code). Fines provide strong enforcement but could lead to litigation over what constitutes an "advanced" system or violation.
- Constitutional: Balances national security interests against First Amendment (free speech in AI development) and Fifth Amendment (property rights) protections; requirements for sharing sensitive data might prompt challenges on privacy or due process grounds, though framed as commerce regulation.
- Political: Bipartisan sponsorship (by Senators Hawley and Blumenthal) signals broad concern over AI risks, positioning it as a proactive step toward U.S. leadership in AI governance without immediate heavy regulation. The sunset clause and congressional approval mechanisms ensure democratic oversight, but proposals like nationalization could spark debates on government overreach in innovation.
This summary was generated by AI and may contain inaccuracies. Refer to the official source document for the authoritative text.
Sponsor
Cosponsors (2)
Sen. Blumenthal, Richard [D-CT], Sen. Blackburn, Marsha [R-TN]
Recent Actions
- 2025-09-29: Read twice and referred to the Committee on Commerce, Science, and Transportation.
- 2025-09-29: Introduced in Senate
Bill Versions
- Artificial Intelligence Risk Evaluation Act of 2025 — issued 2025-09-29 — PDF (14 pages)