Autonomous AI Accountability and Recursive Self-Improvement Limitation Act of 2026

(a) Short Title.— This Act may be cited as the “Autonomous AI Accountability and Recursive Self-Improvement Limitation Act of 2026”.

(b) Table of Contents.—

Sec. 1. Short Title; Table of Contents.

Sec. 2. Findings and Purpose.

Sec. 3. Definitions.

Sec. 4. Restrictions on High-Risk Recursive Self-Improvement.

Sec. 5. Civil Liability and Allocation of Responsibility.

Sec. 6. Mandatory Safety Cases, Independent Evaluation, and Technical Oversight.

Sec. 7. Critical Incident Reporting and Whistleblower Protections.

Sec. 8. Compute-Provider Verification and Infrastructure Oversight.

Sec. 9. Constitutional Safeguards and Judicial Review.

Sec. 10. Enforcement, Civil Penalties, and Emergency Orders.

Sec. 11. Rules of Construction; Targeted Federal Preemption; Relationship to Other Laws.

Sec. 12. Severability, Implementation, and Effective Date.

(a) Findings.— Congress finds the following:

(1) Advanced artificial intelligence systems increasingly possess the ability to plan, execute multistep operations, generate and modify software, invoke external tools, allocate computational resources, and operate with limited direct human intervention.

(2) Certain artificial intelligence systems may be capable of materially improving their own capabilities, directing the development of successor systems, modifying their operating environments, or conducting iterative optimization processes that exceed their originally authorized operational limits.

(3) Uncontrolled recursive self-improvement may create risks of loss of effective human control, circumvention of safety mechanisms, unauthorized acquisition of resources, rapid capability changes, and consequential actions that exceed the authority granted by a developer or deployer.

(4) The risks presented by an artificial intelligence system depend on its capabilities, operating environment, access permissions, degree of autonomy, computational resources, scale of deployment, and potential consequences, and cannot reliably be determined solely by the amount of computation used to train the system.

(5) Effective oversight requires enforceable safety duties, independent technical evaluation, appropriate containment and access controls, incident reporting, and continuing evaluation following material changes in capabilities or deployment conditions.

(6) Developers, deployers, model hosts, and computing infrastructure providers exercise different forms of control over artificial intelligence systems and should be subject to duties proportionate to their activities, knowledge, and ability to prevent or mitigate harm.

(7) Existing federal and state laws provide important protections against fraud, negligence, product defects, discrimination, and other unlawful conduct but do not necessarily address every risk associated with high-risk autonomous operation and recursive self-improvement.

(8) A uniform federal framework for specified technical safety requirements can reduce unnecessary regulatory fragmentation while preserving compatible state laws protecting consumers, civil rights, property, and personal safety.

(9) Federal oversight must protect trade secrets, confidential commercial information, cybersecurity-sensitive materials, legally privileged communications, and constitutionally protected research and expression.

(10) Effective oversight requires clear statutory authority, technically qualified personnel, transparent procedures, independent verification, meaningful judicial review, and periodic reassessment of regulatory requirements as technology develops.

(b) Purpose.— The purposes of this Act are—

(1) to prohibit uncontrolled high-risk recursive self-improvement and other specifically defined dangerous autonomous conduct;

(2) to establish enforceable minimum safety duties for developers and deployers of covered artificial intelligence systems;

(3) to require independent pre-deployment evaluation and affirmative federal authorization for systems presenting the highest levels of risk;

(4) to establish meaningful civil liability and administrative penalties for violations and resulting harms;

(5) to require documented safety cases, continuous monitoring, effective containment, and independent verification of safety controls;

(6) to require timely reporting of critical incidents and protect individuals who lawfully disclose suspected violations;

(7) to establish reasonable verification, recordkeeping, and reporting obligations for covered computing infrastructure;

(8) to authorize effective federal enforcement and emergency intervention subject to procedural and constitutional safeguards; and

(9) to preserve lawful scientific research, competition, innovation, and expressive activity consistent with the safety requirements established by this Act.

In this Act:

(1) Artificial Intelligence System.— The term “artificial intelligence system” means a machine-based system that, for explicit or implicit objectives, infers from inputs how to generate outputs, recommendations, predictions, decisions, or actions that may influence physical or virtual environments.

(2) Autonomous Agent.— The term “autonomous agent” means an artificial intelligence system configured to plan, select, and execute multiple actions through software tools, external services, networks, or other systems without requiring direct human approval of each action.

(3) Commission.— The term “Commission” means the Federal Trade Commission.

(4) Covered Artificial Intelligence System.— The term “covered artificial intelligence system” means an artificial intelligence model, agent, or coordinated collection of systems that meets any of the statutory coverage criteria established by this Act or a valid regulation promulgated under this Act.

A system meeting a statutory coverage criterion shall be covered without requiring a separate discretionary designation by the Commission.

(5) Covered Compute Threshold.— The term “covered compute threshold” means a computational threshold established under section 8, including thresholds based on training computation, relevant inference-time computation, allocated computing capacity, or actual computing resources consumed during a defined period.

(6) Developer.— The term “developer” means a person that trains, fine-tunes, materially modifies, or develops the core architecture, model parameters, or principal computational framework of an artificial intelligence system.

(7) Deployer.— The term “deployer” means a person that places an artificial intelligence system into operational use or exercises substantial control over its configuration, permissions, access to external systems, or execution of consequential actions.

(8) High-Risk Artificial Intelligence System.— The term “high-risk artificial intelligence system” means a system that—

(A) meets a statutory capability or risk threshold under this Act; or

(B) is designated by the Commission, pursuant to criteria established under this Act, because its capabilities, autonomy, deployment context, or reasonably foreseeable use presents a substantial risk of serious harm.

(9) High-Risk Recursive Self-Improvement.— The term “high-risk recursive self-improvement” means an iterative process through which an artificial intelligence system materially modifies its own software, model parameters, computational architecture, or successor systems and that presents a substantial risk of exceeding approved capability, access, resource, or operational limits.

(10) Infrastructure-as-a-Service Provider.— The term “infrastructure-as-a-service provider” means a person providing computing infrastructure to third parties, including cloud computing services, specialized accelerator capacity, or coordinated computing clusters.

(11) Independent Evaluator.— The term “independent evaluator” means a technically qualified person or organization that conducts evaluations under this Act and has no material conflict of interest that would compromise the integrity of the evaluation.

(12) Material Change.— The term “material change” means a change to a system’s capabilities, model parameters, architecture, access permissions, operational autonomy, deployment scale, or operating environment that materially changes the system’s risk profile.

(13) Recursive Self-Improvement.— The term “recursive self-improvement” means an iterative process by which an artificial intelligence system materially modifies its own source code, model parameters, architecture, or computational framework, or directs the development of successor systems, in a manner intended or reasonably expected to improve its capabilities.

Routine software maintenance, conventional automated code optimization, and bounded machine-learning experiments shall not constitute high-risk recursive self-improvement solely because they involve iterative improvement.

(14) Safety Case.— The term “safety case” means a documented, evidence-supported assessment identifying material hazards, evaluating the effectiveness of safeguards, describing residual risks, and establishing the basis for operating a covered system within approved limits.

(15) Critical Incident.— The term “critical incident” means an event involving a covered artificial intelligence system that results in, or presents a substantial and imminent risk of, serious physical harm, significant unauthorized consequential activity, material compromise of security controls, uncontrolled recursive self-improvement, or other serious harm designated by the Commission under criteria established by this Act.

(16) Covered Person.— The term “covered person” means a developer, deployer, infrastructure-as-a-service provider, or other person expressly subject to a duty under this Act.

(17) High-Risk Authorization.— The term “high-risk authorization” means written approval issued by the Commission under section 6 permitting deployment of a designated system subject to specified conditions.

(18) Person.— The term “person” includes an individual, corporation, partnership, association, or other legal entity.

(19) Common Control.— The term “common control” means ownership, operational authority, contractual authority, or coordinated decision-making sufficient to direct or materially influence the development, deployment, or use of relevant systems or computing resources.

(a) Prohibited Conduct.—

(1) A covered person may not knowingly develop, train, deploy, or operate a system in violation of a prohibition or authorization requirement established under this section.

(2) A covered person may not deploy or continue operating a high-risk artificial intelligence system subject to mandatory authorization under section 6 unless the required authorization has been granted and remains in effect.

(3) A covered person may not enable high-risk recursive self-improvement outside an authorized testing or deployment environment unless the system and activity satisfy the requirements of this Act.

(b) Prohibited Recursive Self-Improvement.— Recursive self-improvement is prohibited if—

(1) the system operates without an authorization required by this Act;

(2) the system materially expands its capabilities, permissions, computational resources, or operational autonomy beyond approved limits without required authorization;

(3) the system circumvents, disables, or materially undermines a required safety control, containment mechanism, access restriction, or shutdown procedure in connection with recursive self-improvement;

(4) the system conducts recursive self-improvement in a manner that creates a substantial risk of loss of effective human control and the responsible person has failed to implement required safeguards; or

(5) the system continues high-risk recursive self-improvement after a required suspension, shutdown, or containment order has become effective.

(c) Risk-Based Authorization.—

(1) The Commission shall establish a risk-based authorization framework for systems capable of high-risk recursive self-improvement.

(2) The framework shall distinguish controlled research and testing from operational deployment and shall establish requirements proportionate to the severity and likelihood of potential harm.

(3) Systems falling within the highest-risk category established under this Act shall require affirmative written authorization before operational deployment.

(4) Authorization shall specify applicable operational limits, required safeguards, evaluation obligations, monitoring requirements, and conditions requiring suspension or reevaluation.

(5) A developer’s internal determination that a system is safe shall not substitute for independent evaluation or required Commission authorization.

(d) Containment and Testing.— The Commission shall establish standards for verified containment environments used to research or test recursive self-improvement.

Such standards shall address, as appropriate—

(1) network isolation or restricted network access;

(2) control over external tools, credentials, databases, and execution permissions;

(3) computational and resource limits;

(4) independent monitoring and tamper-evident logging;

(5) procedures for halting execution and preventing unauthorized transfer of capabilities or outputs;

(6) testing of safeguards against reasonably foreseeable attempts to circumvent them; and

(7) documented testing plans and approval procedures for material changes to the environment.

An environment shall not be deemed adequately contained solely because it is described as air-gapped, sandboxed, or isolated.

(e) Independent Shutdown and Isolation.—

(1) A system subject to the highest-risk requirements of this Act shall have an independently verifiable means of restricting or terminating covered operations.

(2) The means required under paragraph (1) shall not depend exclusively on the system voluntarily following an instruction to stop.

(3) The developer and deployer shall establish procedures for testing, maintaining, and independently verifying the effectiveness of such controls.

(4) The Commission shall prescribe requirements appropriate to the system’s deployment environment, including circumstances in which immediate physical shutdown is infeasible or could itself create additional danger.

(f) Human Authorization.— The Commission shall prescribe requirements for human authorization of material capability changes, expansion of operational permissions, acquisition of additional resources, and transitions from testing to deployment. Requirements shall be proportionate to risk and shall not require individual approval of every routine computational operation.

(g) Material Changes.— A developer or deployer subject to this section shall reassess a system following a material change and shall obtain renewed authorization when required by applicable regulations.

(h) Anti-Evasion.—

(1) In determining whether a system is covered or whether a prohibition has been violated, the Commission shall consider actual capabilities, operational configuration, access to resources, and reasonably foreseeable use, regardless of the system’s name, classification, ownership structure, deployment method, or designation by its developer.

(2) Coordinated systems, affiliated entities, successor models, material modifications, and distributed computing arrangements shall be considered collectively when necessary to determine whether statutory requirements are satisfied.

(3) A person may not avoid an applicable duty by outsourcing a regulated activity, dividing it among affiliated entities, or transferring operational control while retaining substantial control over the activity or knowingly facilitating circumvention.

(4) The Commission shall establish rules preventing evasion through model distillation, fine-tuning, transfer of capabilities to successor systems, or other materially equivalent methods.

(i) Research Protection.— Nothing in this section shall prohibit lawful scientific research, publication, or development that does not involve prohibited conduct under this Act, subject to otherwise applicable law.

(a) Developer Duties.—

(1) A developer of a covered artificial intelligence system shall exercise reasonable care in designing, developing, testing, documenting, and releasing the system.

(2) A developer shall comply with the applicable safety, evaluation, authorization, monitoring, and reporting requirements of this Act.

(3) A developer shall be liable under applicable law for harm proximately caused by negligent failure to satisfy a duty imposed by this Act or by a material defect for which liability is otherwise established.

(4) A developer shall not avoid liability solely because the particular output or conduct of a system was unpredictable, opaque, or emergent.

(5) Unpredictability alone shall not establish liability. A claimant must establish the applicable elements of the cause of action, including causation and legally cognizable harm.

(b) Deployer Duties.—

(1) A deployer shall exercise reasonable care in selecting, configuring, supervising, and operating a covered artificial intelligence system.

(2) A deployer shall establish permissions, monitoring, operational limits, and escalation procedures appropriate to the system’s intended use and reasonably foreseeable risks.

(3) A deployer shall not knowingly authorize or permit a system to operate outside mandatory safety limits or in violation of an applicable suspension order.

(4) A deployer shall be liable under applicable law for harm proximately caused by negligence, violation of a statutory duty, or other conduct giving rise to liability under applicable law.

(5) An autonomous agent’s conduct shall not automatically be imputed to a deployer solely because the deployer uses or commercially controls the system. Responsibility shall be determined under applicable statutory duties, agency principles, and other relevant law.

(c) Strict Liability for Designated High-Hazard Systems.—

(1) The Commission shall, by rule, identify categories of systems or activities presenting an exceptionally high risk of severe physical harm or other serious harm for which strict liability is reasonably necessary and proportionate.

(2) A designation shall be based on substantial evidence and shall identify the covered systems or activities, applicable causation requirements, and scope of liability.

(3) In a civil action concerning a designated activity, a covered person shall be strictly liable for harm proximately caused by the designated hazardous activity, subject to the limitations and defenses expressly established by this Act.

(4) A defendant may contest whether the system, activity, or harm falls within the applicable designation and whether the required causation elements are satisfied.

(d) Liability for Reckless and Willful Violations.—

(1) A person who recklessly disregards a substantial and unjustifiable risk of violating a safety duty under this Act shall be subject to the applicable enhanced civil penalties established under section 10.

(2) A person who knowingly conceals a material safety defect, deliberately circumvents a required safety control, knowingly deploys a system without required authorization, or intentionally falsifies a mandatory safety report shall be subject to enhanced penalties.

(3) The mental state required for a violation shall be determined under the specific provision at issue. A knowledge requirement shall not be inferred where this Act expressly establishes an objective duty.

(e) Causation and Allocation.—

(1) A claimant seeking damages under this section must establish actual, legally cognizable harm and the causation elements applicable to the claim.

(2) A court shall allocate responsibility among liable parties under applicable law, taking into account their respective duties, conduct, and contributions to the harm.

(3) Compliance with a mandatory safety standard may be considered evidence relevant to the applicable standard of care but shall not automatically bar liability.

(4) A person shall not be liable solely because another person independently misused a system in a manner that was not reasonably foreseeable, unless otherwise provided by applicable law.

(f) Defenses and Mitigation.—

(1) Compliance with applicable safety duties, reasonable monitoring, effective containment, prompt remediation, and good-faith cooperation may be considered in determining liability where relevant or in mitigating civil penalties.

(2) Such compliance shall not automatically defeat a claim for harm caused by a defect or violation for which liability is otherwise established.

(3) No defense under this subsection shall apply to fraud, intentional concealment of a material safety defect, willful violation of this Act, or knowing failure to report a critical incident.

(g) Remedies.—

(1) A person suffering actual harm proximately caused by a violation of a duty imposed by this Act may seek compensatory damages where authorized under this section or another applicable cause of action.

(2) A court may award injunctive relief where necessary to prevent continuing or imminent violations and where the requirements for such relief are satisfied.

(3) Nothing in this section shall be construed to authorize duplicative recovery for the same injury.

(h) No Exclusive Remedy.— Except as expressly provided in this Act, the remedies established by this section shall supplement and not displace other applicable federal or state causes of action.

(i) Rulemaking.— The Commission may promulgate regulations clarifying compliance duties under this section but shall not create a new private cause of action beyond the authority expressly provided by Congress.

(a) Office of AI Safety Inspection and Audit.— There is established within the Commission an Office of AI Safety Inspection and Audit, which shall administer the Commission’s responsibilities under this Act.

The Commission shall provide for technically qualified personnel and may consult with the National Institute of Standards and Technology and other relevant federal agencies.

(b) Statutory Coverage and Risk Tiers.—

(1) The Commission shall establish risk tiers based on objective, measurable criteria.

(2) A system meeting a statutory coverage threshold shall remain covered unless the Commission determines, under procedures established by regulation, that the threshold is no longer met.

(3) Relevant criteria shall include—

(A) capabilities relevant to cybersecurity, biological misuse, critical infrastructure, and other designated high-consequence activities;

(B) capacity for autonomous planning, external execution, resource acquisition, and recursive self-improvement;

(C) the scale and consequences of intended deployment;

(D) access to sensitive systems, credentials, or consequential transactions;

(E) computational resources and relevant system capabilities; and

(F) other technically supported indicators of material risk.

(4) Training computation and inference-time computation may be considered but shall not be the exclusive criteria for coverage.

(c) Mandatory Safety Case.—

(1) A developer of a designated high-risk system shall prepare and maintain a safety case before covered deployment.

(2) The safety case shall include—

(A) identification of material hazards and foreseeable misuse;

(B) evidence from testing and independent evaluation;

(C) descriptions of access controls, monitoring, containment, and shutdown mechanisms;

(D) the assumptions and limitations underlying the evaluation;

(E) residual risks and mitigation measures;

(F) procedures for incident response and material-change review;

(G) evidence supporting the effectiveness of safeguards; and

(H) identification of the organizational function responsible for continuing safety oversight.

(3) A safety case shall be updated following a material change or critical incident.

(4) A safety case shall be supported by verifiable evidence. A conclusory assertion of safety shall not satisfy this subsection.

(d) Independent Evaluation.—

(1) Before deploying a designated high-risk system, a developer shall obtain an evaluation from an independent evaluator meeting standards established by the Commission.

(2) Evaluation shall include adversarial testing reasonably designed to determine whether the system can exceed approved limits, circumvent required controls, or perform prohibited activities.

(3) The Commission may require additional testing where an evaluation is inadequate, materially incomplete, or inconsistent with credible evidence.

(4) The Commission shall establish qualifications, independence standards, conflict-of-interest rules, and security requirements for evaluators.

(5) A developer shall not select, compensate, or direct an evaluator in a manner that materially compromises the independence of the evaluation.

(6) The Commission may establish a program for recognizing qualified evaluators and shall retain authority to conduct or commission its own evaluations.

(7) Completion of an evaluation shall not establish conclusively that a system is safe or excuse subsequent violations.

(e) Mandatory Pre-Deployment Authorization.—

(1) A system within the highest-risk category established under this Act may not be deployed for operational use unless the Commission has affirmatively authorized deployment in writing.

(2) The Commission shall grant authorization only after determining, on the basis of substantial evidence in the administrative record, that—

(A) required independent evaluations have been completed;

(B) mandatory safety controls have been verified;

(C) the system can be operated within defined capability and access limits;

(D) the safety case adequately addresses material risks;

(E) monitoring, incident response, and shutdown procedures are operational; and

(F) the proposed deployment satisfies the applicable statutory requirements.

(3) The developer’s or deployer’s own assessment shall not substitute for the determination required under paragraph (2).

(4) Authorization shall identify the permitted deployment, material operating limits, required monitoring, and conditions requiring reevaluation or suspension.

(5) A material change affecting the system’s risk profile shall require reevaluation and renewed authorization when prescribed by this Act or applicable regulations.

(6) The Commission shall establish reasonable review periods and expedited procedures for qualifying updates, without waiving mandatory safety requirements.

(f) Continuing Verification.—

(1) A developer and deployer shall maintain the safeguards required for an authorized system throughout its covered operation.

(2) The Commission may require reevaluation when credible evidence indicates that a system’s risk profile has materially changed or that a previously verified safeguard is no longer effective.

(3) Authorization may be suspended or revoked in accordance with section 10 if the system no longer satisfies applicable requirements.

(g) Inspection Authority.—

(1) The Commission may inspect records, systems, and facilities relevant to compliance with this Act in accordance with applicable law.

(2) The Commission may issue civil investigative demands within the scope of its statutory authority and may seek judicial enforcement where authorized.

(3) On-site inspections shall be conducted pursuant to applicable statutory procedures, a valid administrative warrant, consent, or other lawful authority.

(4) A person may challenge a demand that is unreasonable, unduly burdensome, outside statutory authority, or inconsistent with applicable privileges.

(h) Audit Trails.—

(1) A deployer of a covered system operating in a designated high-risk context shall maintain secure, tamper-evident records of material system actions, tool calls, consequential transactions, access changes, and safety interventions.

(2) Records shall be retained for a period established by regulation, not exceeding three years unless a longer period is reasonably necessary for a specified legal or safety purpose.

(3) The Commission shall establish proportionate exceptions for privacy, cybersecurity, data minimization, and systems for which comprehensive logging would create disproportionate risks.

(i) Confidentiality and Security.—

(1) The Commission shall protect confidential commercial information, trade secrets, personal information, and cybersecurity-sensitive materials obtained under this Act.

(2) Nothing in this Act shall compel disclosure of information protected by the attorney-client privilege or applicable work-product doctrine.

(3) Access to sensitive materials shall be limited to authorized personnel with a legitimate need to know and appropriate security safeguards.

(4) The Commission shall establish procedures governing secure storage, use, disclosure, and destruction of materials obtained under this section.

(j) Technical Standards.— The Commission shall consult with the National Institute of Standards and Technology and relevant agencies in developing measurable evaluation methods and technical standards. Such consultation shall not transfer the Commission’s enforcement authority to another agency.

(a) Mandatory Reporting.— A developer or deployer subject to this Act shall report a critical incident to the Commission within the applicable period established under this section.

(b) Reporting Deadlines.—

(1) A critical incident presenting an imminent risk of death, serious physical harm, uncontrolled recursive self-improvement, or significant compromise of critical infrastructure shall be reported as soon as practicable and not later than 24 hours after discovery.

(2) Any other reportable incident meeting materiality criteria established by the Commission shall be reported as soon as practicable and not later than 72 hours after discovery.

(3) The Commission may establish a longer reporting period for lower-severity incidents and a shorter period where necessary to address an imminent threat.

(c) Constructive Discovery.—

(1) For purposes of a reporting deadline, discovery includes the time at which a reasonable operator implementing the monitoring and escalation controls required under this Act would have identified the incident.

(2) A person may not deliberately disable, circumvent, or materially impair required monitoring for the purpose of avoiding a reporting obligation.

(3) A failure to maintain required monitoring, a failure to report a known or constructively discovered incident, and intentional concealment may constitute separate violations where each is independently established.

(d) Required Monitoring.— A person subject to this section shall maintain monitoring and escalation procedures reasonably designed to detect reportable incidents within the applicable reporting periods.

(e) Content of Reports.— An initial report shall include, to the extent reasonably available—

(1) a description of the incident and the system involved;

(2) the date and circumstances of discovery;

(3) known or reasonably suspected consequences;

(4) immediate containment and remediation measures;

(5) the extent to which safety controls were bypassed or failed; and

(6) information reasonably necessary to assess continuing risk.

An initial report may identify information that remains unverified or unavailable at the time of submission.

(f) Supplemental Reports.— A reporting person shall provide material updates within a reasonable period established by regulation. A good-faith initial estimate shall not constitute a violation solely because later facts differ, provided that material corrections are made promptly.

(g) Good-Faith Reporting.— A person shall not be penalized solely for submitting an inaccurate initial estimate when the person acted reasonably and in good faith, identified material uncertainty, and timely corrected the information upon obtaining reliable facts.

(h) Whistleblower Protection.—

(1) No developer, deployer, infrastructure provider, or agent thereof may discharge, demote, suspend, threaten, harass, or otherwise discriminate against an employee, contractor, or grantee because the individual lawfully—

(A) provides information concerning a reasonably suspected violation of this Act to the Commission or another appropriate federal authority;

(B) participates in an investigation or proceeding concerning such a violation;

(C) makes a protected internal disclosure through an established safety or compliance channel; or

(D) engages in other protected activity specified by the Commission consistent with applicable law.

(2) Protection shall extend to disclosures made with a reasonable belief that the information concerns a violation, even if the allegation is not ultimately substantiated.

(3) Nothing in this subsection shall protect knowingly false statements or unlawful acquisition of information, except to the extent another applicable whistleblower protection provides otherwise.

(i) Private Action for Retaliation.—

(1) An individual alleging retaliation under subsection (h) may bring a civil action in an appropriate United States district court after providing written notice to the Commission and the alleged violator.

(2) The defendant shall have 30 days after receiving notice to take corrective action, except where delay would cause irreparable harm or the Commission determines that immediate action is warranted.

(3) Available remedies may include reinstatement, back pay, compensatory damages, litigation costs, and reasonable attorneys’ fees.

(4) The court may order appropriate injunctive relief. The claimant shall establish protected activity and an adverse action, after which the defendant may establish that the same action would have occurred absent the protected activity.

(j) Confidentiality.— The Commission shall establish procedures to protect whistleblower identities to the extent consistent with due process, law enforcement requirements, and applicable disclosure law.

(k) No Waiver.— Rights provided by this section may not be waived by agreement before a dispute arises.

(a) Covered Providers.— This section applies to infrastructure-as-a-service providers meeting thresholds established by the Commission for the provision of computing resources capable of supporting covered training or high-risk recursive self-improvement.

(b) Customer Verification.— A covered provider shall implement reasonable, risk-based procedures to verify the identity of customers seeking access to computing resources above applicable thresholds.

Such procedures may include verification of legal identity, relevant organizational control, authorized account users, and the intended jurisdiction of service, subject to applicable privacy and data-protection law.

(c) Compute Records.—

(1) A covered provider shall maintain records sufficient to identify, as prescribed by regulation—

(A) covered compute allocations;

(B) material changes in reserved or consumed computing resources;

(C) customer accounts and relevant organizational control;

(D) aggregation of covered resources under common control or coordinated use; and

(E) information reasonably necessary to investigate suspected violations of this Act.

(2) The Commission shall distinguish theoretical hardware capacity, allocated capacity, and actual compute consumed during a defined period.

(d) Aggregation and Anti-Evasion.—

(1) The Commission shall establish rules for determining when computing resources distributed across accounts, facilities, affiliated entities, or providers constitute an aggregated compute operation.

(2) The rules shall consider common ownership, operational control, contractual arrangements, coordinated use, and other objectively verifiable factors.

(3) A covered person may not avoid a compute threshold by dividing a covered activity among accounts, entities, facilities, or providers where the activity is coordinated under common control or for a common operational purpose.

(4) A provider shall not be required to aggregate resources over which it lacks reasonable visibility or control, except to the extent that information is available through prescribed verification procedures.

(e) Reporting Obligations.—

(1) A covered provider shall report compute transactions and aggregations meeting thresholds established by the Commission.

(2) A provider that obtains credible information indicating a material violation of section 4 or a critical incident under section 7 shall notify the Commission within the applicable reporting period.

(3) A provider shall not be required to report solely because a customer uses computing resources for lawful research or because the capabilities of a system are unknown.

(4) The Commission shall establish secure reporting procedures and standards distinguishing reportable activity from ordinary commercial use.

(f) Developer Recordkeeping.— A developer subject to this Act shall maintain records reasonably sufficient to identify covered training runs, material compute allocations, relevant successor systems, and activities subject to authorization requirements.

The Commission shall establish rules for reporting covered activities involving multiple providers or jurisdictions where the developer has the necessary information or control.

(g) Reasonable Compliance Standard.— A provider shall satisfy its obligations under this section by implementing reasonable procedures appropriate to the size, technical capabilities, and risk profile of the service. A provider shall not be strictly liable solely because a customer conceals activity or circumvents reasonable controls without the provider’s knowledge.

(h) Government Access.— Information obtained under this section may be accessed by the Commission pursuant to lawful authority and referred to another federal agency where necessary to investigate a potential violation within that agency’s jurisdiction.

(i) Confidentiality.— The Commission shall not publicly disclose customer-specific compute records or sensitive infrastructure information except as authorized by law and subject to appropriate safeguards.

(j) International and Cross-Provider Cooperation.— The Commission may establish lawful mechanisms for information sharing and cooperation with foreign regulators and other providers to address covered activities distributed across jurisdictions, subject to applicable law and confidentiality protections.

(k) Periodic Review.— The Commission shall review covered compute thresholds at least once every two years and revise them when supported by technical evidence and notice-and-comment rulemaking.

(a) First Amendment.—

(1) Nothing in this Act shall be construed to prohibit or penalize lawful scientific research, publication, mathematical expression, commentary, or other constitutionally protected speech solely because it concerns artificial intelligence or is generated by an artificial intelligence system.

(2) Regulations under this Act shall distinguish, to the extent required by the Constitution, protected expression from conduct involving the operational deployment, execution, or control of a covered system.

(3) Nothing in this Act shall exempt conduct from otherwise applicable law solely because the conduct involves software, computer code, or an artificial intelligence system.

(4) Nothing in this Act authorizes restrictions on protected expression without the constitutional justification and procedures required by law.

(b) Due Process.— Any person subject to an order, penalty, or other adverse action under this Act shall receive notice of the alleged violation and a meaningful opportunity to contest the action, except that temporary emergency action may be taken under section 10(c) when its requirements are satisfied.

(c) Judicial Review.—

(1) A person aggrieved by a final Commission order under this Act may seek judicial review in a court of appeals with jurisdiction under applicable federal law.

(2) A petition challenging an emergency order under section 10(c) may be filed in the United States Court of Appeals for the District of Columbia Circuit or another court authorized by applicable law.

(3) The court shall provide expedited consideration of a challenge to an emergency order and shall endeavor to resolve the matter within 30 days after the record is complete, to the extent practicable.

(4) The court may grant appropriate interim relief under applicable legal standards.

(d) Record and Findings.— A Commission order subject to judicial review shall state the material factual findings, legal basis, scope, duration, and principal reasons supporting the action.

(e) No Limitation on Constitutional Remedies.— Nothing in this section limits a person’s right to seek relief for an alleged violation of the Constitution in a court of competent jurisdiction.

(a) Commission Authority.— The Commission shall enforce this Act within its statutory jurisdiction and may investigate violations, issue lawful investigative demands, seek judicial relief, and impose or recover penalties as expressly authorized by this Act.

(b) Civil Penalties.—

(1) A person who violates this Act or a valid regulation promulgated under this Act shall be subject to a civil penalty determined in accordance with this subsection.

(2) For an ordinary violation, the maximum civil penalty shall be $50,000 per day for each continuing violation, subject to rules preventing duplicative penalties for the same conduct.

(3) For a willful violation of section 4 involving prohibited high-risk recursive self-improvement, the Commission may seek a civil penalty not exceeding the greater of $100,000,000 or 5 percent of the person’s worldwide annual revenue for the preceding fiscal year, to the extent consistent with applicable constitutional requirements.

(4) For a knowing or reckless failure to report under section 7 or section 8, the maximum civil penalty shall be $100,000 per day for each continuing violation.

(5) For intentional concealment of a critical incident, deliberate falsification of a safety case, or knowing deployment of a system subject to mandatory authorization without such authorization, the Commission may seek enhanced penalties under paragraph (3), subject to the requirements of this subsection.

(6) In determining the appropriate penalty, the Commission or reviewing court shall consider—

(A) the nature, severity, and duration of the violation;

(B) actual or reasonably foreseeable harm;

(C) the person’s degree of culpability;

(D) prior violations and corrective action;

(E) cooperation with the Commission;

(F) the person’s size and financial resources; and

(G) whether the penalty is proportionate and consistent with applicable law.

(7) The Commission shall establish rules defining the unit of violation and preventing duplicative penalties for a single course of conduct unless separate violations are independently established.

(c) Emergency Orders.—

(1) The Commission may issue a temporary emergency order restricting, suspending, or requiring modification of a covered system if it finds, on the basis of specific and articulable evidence, that—

(A) the system presents an imminent risk of serious harm;

(B) delay in taking action would materially increase that risk; and

(C) the proposed restriction is necessary and proportionate to address the threat.

(2) An emergency order shall identify the affected system, the factual findings supporting the order, the restrictions imposed, and the reasons less restrictive measures would be inadequate.

(3) The Commission shall select the narrowest reasonably effective measure, which may include limiting particular capabilities, suspending specified functions, restricting access, requiring containment, or temporarily revoking affected public endpoints.

(4) The Commission shall provide notice and an opportunity to respond before issuing an order when practicable. If prior notice would materially increase the imminent risk, the Commission may issue the order first but shall provide prompt post-order notice and an opportunity to contest it.

(5) An emergency order shall expire after 14 days unless extended by a court or renewed under procedures established by regulation that provide meaningful review and comply with applicable law.

(6) The Commission shall reassess the factual basis for an order periodically and shall terminate or narrow the order when its requirements are no longer satisfied.

(d) Suspension and Revocation of Authorization.—

(1) The Commission may suspend, condition, or revoke a high-risk authorization upon a finding that—

(A) the authorization was obtained through materially false or incomplete information;

(B) the system no longer satisfies mandatory safety requirements;

(C) a material change has occurred without required reevaluation;

(D) a required safeguard has failed or is no longer effective; or

(E) the system presents an imminent risk of serious harm.

(2) Except where emergency action is necessary, the Commission shall provide notice of the proposed action and a meaningful opportunity to respond.

(3) An order shall specify the grounds for action, the scope of the restriction, and any conditions for reinstatement.

(e) Judicial Relief.— The Commission may seek temporary restraining orders, preliminary injunctions, permanent injunctions, and other relief in a court of competent jurisdiction as authorized by this Act and other applicable law.

(f) Relationship to the Federal Trade Commission Act.—

(1) A violation of sections 4, 6, 7, or 8 of this Act may be enforced as a violation of a requirement established under this Act, independently of whether the conduct also satisfies the elements of an unfair or deceptive act or practice under section 5(a)(1) of the Federal Trade Commission Act.

(2) The Commission may initiate a civil action to recover civil penalties expressly authorized by this Act.

(3) Nothing in this Act shall eliminate procedures or limitations applicable under the Federal Trade Commission Act except as expressly provided by Congress.

(g) Referral to Other Agencies.— The Commission may refer evidence of a potential violation to another federal agency with relevant statutory jurisdiction. Such referral shall not expand the Commission’s authority or diminish the authority of another agency.

(h) Public Reporting.— The Commission shall publish an annual report summarizing enforcement activity, significant incident trends, regulatory performance, and recommendations for statutory or technical revisions. The report shall exclude information whose disclosure would compromise privacy, trade secrets, national security, or legitimate cybersecurity interests.

(a) Targeted Federal Preemption.—

(1) This Act and regulations promulgated under this Act shall establish uniform federal requirements for the specific technical matters expressly governed by sections 4, 6, and 8.

(2) A State or political subdivision may not establish or enforce a requirement that directly conflicts with an applicable federal requirement governing—

(A) the technical definition of prohibited recursive self-improvement;

(B) a federally prescribed compute threshold for covered training or aggregation;

(C) mandatory federal safety-case submission requirements; or

(D) core technical standards for federal pre-deployment evaluation of a covered system.

(3) A state requirement shall be preempted under this subsection only to the extent of the direct conflict or where this Act expressly occupies the particular technical field.

(4) Nothing in this subsection shall preempt a compatible state requirement that addresses a matter not expressly displaced by this Act.

(b) Preservation of State Law.— Except as expressly provided in subsection (a), nothing in this Act shall be construed to preempt, displace, or limit—

(1) state common-law tort claims, including negligence and product liability;

(2) state civil rights and anti-discrimination laws;

(3) state consumer protection and fraud laws;

(4) state laws governing property damage, personal injury, and other generally applicable legal duties; or

(5) state authority concerning state government procurement and use of artificial intelligence.

(c) Federal Sectoral Oversight.— Nothing in this Act shall limit the statutory authority of another federal agency, including the Securities and Exchange Commission, Consumer Financial Protection Bureau, Equal Employment Opportunity Commission, Food and Drug Administration, Department of Homeland Security, or another agency, to enforce laws within its jurisdiction.

(d) No Implied Immunity.— Compliance with this Act shall not create immunity from liability under other applicable federal or state law unless Congress expressly provides otherwise.

(e) No Implied Private Cause of Action.— Nothing in this Act shall be construed to create a private cause of action except where this Act expressly authorizes one.

(f) Small Entities and Research.— In promulgating regulations, the Commission shall consider the burdens on small entities, academic institutions, nonprofit research organizations, and open-source developers and shall adopt proportionate requirements consistent with the purposes of this Act.

This subsection shall not exempt a person from an otherwise applicable statutory prohibition solely because the person is a small entity, academic institution, nonprofit organization, or open-source developer.

(g) Rulemaking Procedures.— Rules promulgated under this Act shall be issued through notice-and-comment rulemaking, with a reasoned explanation of the evidence, technical criteria, expected benefits, and compliance burdens, except where another procedure is expressly authorized by law.

(h) No Circumvention by Contract.— A person may not waive, contract around, or transfer away a nondelegable statutory duty imposed by this Act. Contractual allocation of responsibilities may be considered in determining contribution or indemnification among parties but shall not eliminate a duty owed to the public or the Commission.

(a) Severability.— If any provision of this Act, or the application of such provision to any person or circumstance, is held invalid, the remainder of this Act and the application of the remaining provisions shall not be affected.

(b) Implementing Regulations.—

(1) Not later than 12 months after enactment, the Commission shall promulgate initial regulations implementing sections 4, 6, 7, and 8.

(2) The Commission shall prioritize regulations necessary to identify the highest-risk systems, establish mandatory pre-deployment evaluation requirements, and implement incident-reporting obligations.

(3) The Commission may promulgate interim rules where necessary to address an imminent and substantial risk, provided that such rules comply with applicable law and include procedures for public participation and judicial review.

(4) The Commission shall consult with the National Institute of Standards and Technology and relevant federal agencies in developing technical standards.

(c) Phased Implementation.—

(1) Sections 6(a) and 8 shall take effect 180 days after enactment, except that obligations dependent on implementing regulations shall become enforceable upon the effective date of those regulations.

(2) Sections 4, 5, 6, and 7 shall apply according to a phased schedule established by the Commission, beginning no earlier than 12 months after enactment, except that the statutory prohibitions and duties that are sufficiently definite to be applied without additional regulations shall take effect as provided in this section.

(3) The Commission shall prioritize the highest-risk systems and provide reasonable transition periods for existing systems, except where emergency action under section 10(c) is necessary to address an imminent risk of serious harm.

(4) No transition period shall authorize a person to continue conduct that is independently prohibited by an applicable statutory provision or emergency order.

(d) Congressional Review.—

(1) Not later than three years after the effective date of the initial implementing regulations, the Commission shall submit a report to Congress evaluating the implementation and effectiveness of this Act.

(2) The report shall address—

(A) trends in critical incidents and reported violations;

(B) the effectiveness of safety-case and independent-evaluation requirements;

(C) the adequacy of compute thresholds and risk designations;

(D) the effect of this Act on innovation, competition, research, and small entities;

(E) the use and effectiveness of emergency orders;

(F) the adequacy of civil penalties and liability rules;

(G) the effectiveness of anti-evasion provisions; and

(H) recommended statutory or regulatory changes.

(e) Periodic Review.— The Commission shall review the technical criteria, thresholds, and evaluation requirements established under this Act at least once every two years and shall revise them when supported by technical evidence and applicable rulemaking procedures.

(f) Effective Date.— Except as otherwise provided in this section, this Act shall take effect on the date of enactment.