WHAT THIS BILL REGULATES · 3 REQUIREMENT TYPES
How Is This Bill Enforced
Verbatim statutory text on the left; plain-language analysis and a per-section checklist on the right. Numbered markers cross-link to the matching checklist row.
This Act may be cited as the Artificial Intelligence Safety Act.
Establishes the short title of the Act as the Artificial Intelligence Safety Act. Creates no compliance obligations.
In this Act: "Artificial intelligence modelArtificial intelligence model"Artificial intelligence model" means an engineered or machine-based component of a system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.Section 5" means an engineered or machine-based component of a system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
"Critical harmCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.
"DeveloperDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5" means a person or organization that has trained, or initiated the training of, at least one frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5.
"Frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5" means an artificial intelligence modelArtificial intelligence model"Artificial intelligence model" means an engineered or machine-based component of a system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.Section 5 that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.
Defines the four key terms used throughout the Act: artificial intelligence model, critical harm, developer, and frontier model. The critical harm definition is notable for its extremely high threshold — death or serious injury to 100+ people or $1B+ in property damages — limiting the bill's scope to catastrophic-scale events involving CBRN weapons or autonomous criminal conduct. The frontier model definition uses a dual threshold: either >10^26 computational operations or >$100M compute cost.
(a) 1 A developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 shall not be held liable for critical harmsCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 if the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 did not intentionally or recklessly cause the critical harmsCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 and the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5: (1) published a safety and security protocol on its website that satisfies the requirements of Section 15 and adhered to that safety and security protocol prior to the release of the frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5; (2) published a transparency report on its website at the time of the frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5's release that satisfies the requirements of Section 20. The requirements of paragraphs (1) and (2) do not apply if the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 does not reasonably foresee any material difference between the frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5's capabilities or risks of critical harmCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 and a frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 that was previously evaluated by the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 in a manner substantially similar to this Act.
(b) 2 If a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 has produced a safety and security protocol in a manner substantially similar to this Act, then the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 shall be deemed to have complied with subsection (a) if the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5: (1) agrees to be bound by the safety and security requirements adopted under Article 56 of the European Union's Artificial Intelligence Act; or (2) enters into an agreement with an agency of the federal government that: (A) enables the agency to access the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5's frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 for the purposes of conducting research and evaluations; (B) facilitates the evaluation of frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5, such as an assessment of cyber and biological risks; and (C) allows for the federal government to release information related to evaluations of frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 that have been publicly released.
(c) 3 A developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 that enters into an agreement with an agency of the federal government under paragraph (2) of subsection (b) shall file a certification with the Attorney General, in a manner prescribed by the Attorney General, attesting that the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 has entered into an agreement that satisfies the requirements of paragraph (2) of subsection (b). It is a violation of this Act to make a materially false or misleading statement, or to omit a material fact, in any certification submitted under this subsection.
Section 10 is the operative core of the bill. It creates a liability safe harbor rather than a traditional regulatory mandate: developers of frontier models are shielded from liability for critical harms they did not intentionally or recklessly cause, provided they published a compliant safety and security protocol and transparency report on their website before releasing the model. A carve-out exempts developers from these requirements where the new model's capabilities and risks are not materially different from a previously evaluated model.
Subsection (b) provides two alternative compliance pathways: agreeing to be bound by the EU AI Act's Article 56 safety requirements, or entering a federal agency evaluation agreement that grants the agency model access, facilitates risk evaluation, and permits public release of evaluation results. Subsection (c) requires developers using the federal-agreement pathway to file a certification with the Attorney General and makes materially false certifications a violation of the Act.
(a) 4 In order to satisfy the requirements of paragraph (1) of subsection (a) of Section 10, a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 shall create a safety and security protocol that documents a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5's technical and organizational protocols to manage, assess, and mitigate risk of critical harmCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 arising from the use and distribution of its frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5. It shall include a high-level summary of: (1) testing procedures that the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 uses to assess risk of a reasonably foreseeable critical harmsCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 arising from the deployment of a frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5; (2) thresholds used by the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 to identify and assess whether a frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 poses a critical risk, including: (A) how the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 will assess whether a threshold has been attained, which may include multiple tiered thresholds; and (B) an illustrative summary of the potential actions the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 may take if each threshold is attained; (3) the mitigations that a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 takes to mitigate reasonably foreseeable critical harmsCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 and how the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 assesses the effectiveness of those mitigations; (4) whether and how a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 will use third parties to assess risk of critical harmCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 and capabilities or the effectiveness of mitigations of critical harmsCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5; (5) the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5's cybersecurity practices relating to frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 and how the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 secures unreleased frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 weights from unauthorized modification or transfer by internal or external parties; (6) to the extent that the frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 is deployed by the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5, a description of how the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 will monitor for critical harmCritical harm"Critical harm" means the death or serious injury of 100 or more people or at least $1,000,000,000 of damages to rights in property caused or materially enabled by a frontier model, through either: (1) the creation or use of a chemical, biological, radiological, or nuclear weapon; or (2) engaging in conduct that: (A) acts with no meaningful human intervention; and (B) would, if committed by a human, constitute a criminal offense that requires intent, recklessness, or negligence, or the solicitation or aiding and abetting of such a crime.Section 5 and how a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 would respond; and (7) the process by which the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 determines when a frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 presents material new risks such that it should be subject to additional assessments.
(b) 4 A developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 may make appropriate redactions to the information required under subsection (a), as necessary, to maintain model security and integrity, trade secrets, and proprietary information.
Section 15 specifies the content requirements for the safety and security protocol that developers must publish under Section 10(a)(1). The protocol must document technical and organizational procedures for managing, assessing, and mitigating risk of critical harm, and must include high-level summaries across seven enumerated categories: testing procedures, risk thresholds with tiered response actions, mitigation measures and their effectiveness, third-party risk assessment plans, cybersecurity practices including protection of unreleased model weights, post-deployment monitoring and response procedures, and the process for identifying when a model presents material new risks requiring additional assessment.
Subsection (b) permits appropriate redactions for model security, trade secrets, and proprietary information — an important concession that ensures the public disclosure requirement does not compromise the security it is meant to promote.
(a) 5 In order to satisfy the requirements of paragraph (2) of subsection (a) of Section 10, a developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 shall create a transparency report that: (1) identifies the frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 that is the subject of the transparency report; and (2) provides a summary of the results of assessments conducted in accordance with the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5's safety and security protocol and the steps taken to address any identified risks.
(b) 5 A developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 may make appropriate redactions to the information required under subsection (a), as necessary, to maintain model security and integrity, trade secrets, and proprietary information.
Section 20 specifies the content requirements for the transparency report that developers must publish under Section 10(a)(2). The report must identify the frontier model and provide a summary of assessment results conducted under the developer's safety and security protocol, along with steps taken to address identified risks. As with the safety protocol, appropriate redactions for model security, trade secrets, and proprietary information are permitted.
(a) This Act shall no longer apply if the federal government enacts a law or adopts regulations that establish overlapping requirements for developersDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 of frontier modelsFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5.
(b) A developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 shall be deemed to be in compliance with this Act if the developerDeveloper"Developer" means a person or organization that has trained, or initiated the training of, at least one frontier model.Section 5 is in compliance with another state's substantially similar frontier modelFrontier model"Frontier model" means an artificial intelligence model that: (1) is trained using greater than 10^26 computational operations, such as integer or floating-point operations; or (2) has a compute cost that exceeds $100,000,000.Section 5 safety and security framework.
Section 25 contains two scope-limiting provisions. Subsection (a) provides a federal preemption sunset: the Act ceases to apply entirely if the federal government enacts a law or adopts regulations establishing overlapping requirements for frontier model developers. Subsection (b) provides a reciprocity safe harbor: developers in compliance with another state's substantially similar frontier model safety and security framework are deemed in compliance with this Act. Neither subsection creates an independent compliance obligation.