WHAT THIS BILL REGULATES · 2 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.
As used in this act: "Artificial intelligence technologyArtificial intelligence technology"Artificial intelligence technology" means any technology involving the development or implementation of machine-based systems that, for explicit or implicit objectives, infer, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that may influence physical or virtual environments.Section 1" means any technology involving the development or implementation of machine-based systems that, for explicit or implicit objectives, infer, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that may influence physical or virtual environments.
"Catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 or user, such as using deceptive techniques against the frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 or user to subvert the controls or monitoring of its frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1" shall not include: incidents arising from information that a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation modelFoundation model"Foundation model" means an artificial intelligence technology that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.Section 1; or lawful activity of the federal government.
"Catastrophic riskCatastrophic risk"Catastrophic risk" means a foreseeable and material risk of catastrophic harm.Section 1" means a foreseeable and material risk of catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1.
"Critical safety incidentCritical safety incident"Critical safety incident" means: a. unauthorized access to, modification of, or exfiltration of, the model weights of a frontier model that results in death or bodily injury; b. harm resulting from the materialization of a catastrophic risk; c. loss of control of a frontier model causing death or bodily injury; or d. a frontier model that uses deceptive techniques against the frontier developer to subvert the controls or monitoring of its frontier developer outside of the context of an evaluation designed to elicit this behavior and in a manner that demonstrates materially increased catastrophic risk.Section 1" means: a. unauthorized access to, modification of, or exfiltration of, the model weights of a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 that results in death or bodily injury; b. harm resulting from the materialization of a catastrophic riskCatastrophic risk"Catastrophic risk" means a foreseeable and material risk of catastrophic harm.Section 1; c. loss of control of a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 causing death or bodily injury; or d. a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 that uses deceptive techniques against the frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 to subvert the controls or monitoring of its frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 outside of the context of an evaluation designed to elicit this behavior and in a manner that demonstrates materially increased catastrophic riskCatastrophic risk"Catastrophic risk" means a foreseeable and material risk of catastrophic harm.Section 1.
"DeployingDeploying"Deploying" or "deployment" means to make a frontier model available to a third party for use, modification, copying, or combination with other software.Section 1" or "deployment" means to make a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 available to a third party for use, modification, copying, or combination with other software.
"Foundation modelFoundation model"Foundation model" means an artificial intelligence technology that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.Section 1" means an artificial intelligence technologyArtificial intelligence technology"Artificial intelligence technology" means any technology involving the development or implementation of machine-based systems that, for explicit or implicit objectives, infer, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that may influence physical or virtual environments.Section 1 that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.
"Frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.
"Frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1" means a foundation modelFoundation model"Foundation model" means an artificial intelligence technology that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.Section 1 that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation modelFoundation model"Foundation model" means an artificial intelligence technology that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.Section 1.
"Generative artificial intelligence systemGenerative artificial intelligence system"Generative artificial intelligence system" or "GenAI system" means a technology system that is trained on data and can generate derived synthetic content, including text, images, video, and audio, which content emulates the structure and characteristics of the system's training data.Section 1" or "GenAI system" means a technology system that is trained on data and can generate derived synthetic content, including text, images, video, and audio, which content emulates the structure and characteristics of the system's training data.
"Large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1" means a frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1 that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.
"Model weightModel weight"Model weight" means a numerical parameter in a frontier model that is adjusted through training and that helps determine how inputs are transformed into outputs.Section 1" means a numerical parameter in a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 that is adjusted through training and that helps determine how inputs are transformed into outputs.
"OfficeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1" means the New Jersey Office of Homeland Security and Preparedness.
Section 1 supplies the operative definitions for the act. Coverage turns on two compute-and-revenue gates: a frontier developer is one who trained a frontier model using at least 10^26 operations, and a large frontier developer is a frontier developer whose group revenue exceeds $100 million. The obligations in later sections fall only on large frontier developers.
The catastrophic harm definition is notably lower than the common frontier-safety threshold — 25 deaths or serious injuries (versus the more typical 100) or $1 billion in property damage — and enumerates weapon-uplift, autonomous-crime, and control-evasion scenarios, with carve-outs for publicly available information and lawful federal government activity.
a. 1 A large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 who has users in the State of New Jersey shall: a. write and implement technical and organizational protocols to assess and reduce the risk of the development, storage, use, or deployment of a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 materially contributing to a catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1;
b. 2 b. annually submit to the Attorney General a "Risk Management Disclosure" containing a written explanation of technical and organizational protocols the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 has implemented to assess and reduce the risk of the development, storage, use, or deployment of a frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 materially contributing to a catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1;
c. 3 c. annually review and, as appropriate, update the technical and organizational protocols the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 has implemented to assess and reduce the risk of the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's models materially contributing to a catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1; and
d. 4 d. submit to the Attorney General a "New Model Risk Disclosure" before, or concurrently with, deployingDeploying"Deploying" or "deployment" means to make a frontier model available to a third party for use, modification, copying, or combination with other software.Section 1 a new frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 or a substantially modified version of an existing frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1 following the effective date of this act, containing: (1) assessments of catastrophic risksCatastrophic risk"Catastrophic risk" means a foreseeable and material risk of catastrophic harm.Section 1 from the frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1, described in enough detail that a third party may exactly replicate the assessments; (2) the results of the assessments required pursuant to paragraph (1) of this subsection; (3) the extent to which third-party evaluators were involved in the assessments required pursuant to paragraph (1) of this subsection; and (4) other steps taken to fulfill the requirements of the rules issued pursuant to section 3 of this act.
Section 2 is the operative core of the bill. A large frontier developer with users in New Jersey must maintain a written and implemented catastrophic-risk management program, submit an annual Risk Management Disclosure describing those protocols to the Attorney General, review and update the protocols annually, and file a New Model Risk Disclosure before or concurrently with deploying any new or substantially modified frontier model.
The New Model Risk Disclosure must contain catastrophic-risk assessments described in enough detail to be exactly replicable by a third party, the assessment results, the extent of third-party evaluator involvement, and other steps required by the Attorney General's rules. These are structured governance and reporting duties keyed to frontier-model capability rather than any particular downstream use.
a. 5 a. The Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, shall issue rules requiring that the "Risk Management Disclosure" submitted pursuant to subsection b. of section 2 of this act contain language or a visual guide which maps the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's actions to the suggested actions contained in the most recent edition of the National Institute of Standards and Technology's AI Risk Management Framework, or to a relevant use-case profile based on the framework. For each suggested action, the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 shall state whether the suggested action is relevant to the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's operations, whether or not the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 has adopted the suggested action or a modified version of the suggested action, a description of how the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 has implemented the suggested action or a justification for the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's decision not to implement the suggested action, including the criteria the developer used to determine compliance with each suggested action, and which employees or other persons the developer has made responsible for implementation of the suggested action.
b. b. The Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, may substitute the suggested actions in the National Institute of Standards and Technology's AI Risk Management Framework with other national standards, international standards, or industry-consensus best practices if the Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, determines that: (1) the AI Risk Management Framework is out of date; and (2) that the National Institute of Standards and Technology is no longer producing updated editions of the framework.
c. 6 c. The Attorney General shall, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, issue rules requiring developers to clearly identify which sections of the "Risk Management Disclosure" submitted pursuant to subsection b. of section 2 this act were written, edited, or otherwise contributed to by a generative artificial intelligence systemGenerative artificial intelligence system"Generative artificial intelligence system" or "GenAI system" means a technology system that is trained on data and can generate derived synthetic content, including text, images, video, and audio, which content emulates the structure and characteristics of the system's training data.Section 1.
d. d. The Attorney General shall publish on Department of Law and Public Safety's Internet website all disclosures submitted pursuant to section 2 of this act, with redactions as necessary to protect the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's trade secrets, the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's cybersecurity, public safety, the privacy of large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's employees, or the national security of the United States or to comply with any federal or State law.
Section 3 directs the Attorney General, in consultation with the Office of Homeland Security and Preparedness, to issue rules shaping the content of the Risk Management Disclosure. The disclosure must map the developer's actions to the suggested actions in the most recent NIST AI Risk Management Framework, stating for each action whether it is relevant, whether it was adopted, how it was implemented (or why not), and who is responsible. The Attorney General may substitute other standards if the NIST framework becomes outdated and is no longer being updated.
The rules must also require developers to identify which sections of the Risk Management Disclosure were written, edited, or contributed to by a generative AI system, and the Attorney General must publish submitted disclosures with redactions for trade secrets, cybersecurity, public safety, employee privacy, and national security.
a. a. A large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 who fails to submit any disclosures required pursuant to section 2 of this act, or who makes materially false or misleading statements in the frontier developerFrontier developer"Frontier developer" means a person who has used, or initiated the use of, a quantity of computing power of at least 10^26 integer or floating-point operations to train a frontier model, including computing used for the original training run and for any subsequent fine-tuning, reinforcement learning, or other modifications the person applies.Section 1's "Risk Management Disclosure" or "New Model Risk Disclosure," shall be liable for a civil penalty in the amount of $100,000 for each violation, which penalty shall be collected by the Attorney General in a summary proceeding before a court of competent jurisdiction pursuant to the provisions of the "Penalty Enforcement Law of 1999," P.L.1999, c.274 (C.2A:58-10 et seq.).
b. b. A large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 who makes materially false or misleading statements about catastrophic riskCatastrophic risk"Catastrophic risk" means a foreseeable and material risk of catastrophic harm.Section 1 in the large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1's "Risk Management Disclosure" or "New Model Risk Disclosure" shall not be liable for a civil penalty pursuant to subsection a. of this section if the statements are found to be reasonable and made in good faith.
Section 4 establishes the enforcement mechanism: a $100,000 civil penalty per violation for failing to submit required disclosures or making materially false or misleading statements, collected by the Attorney General in a summary proceeding under the Penalty Enforcement Law of 1999. A good-faith safe harbor shields a developer from penalty for false or misleading statements about catastrophic risk where those statements are found to be reasonable and made in good faith.
The Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, may: a. audit a large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1 for compliance with the provisions of this act; or b. enter into a contract with a private entity to conduct an audit pursuant to subsection a. of this section.
Section 5 authorizes the Attorney General, in consultation with the Office of Homeland Security and Preparedness, to audit a large frontier developer for compliance and to contract with a private entity to perform such audits. This is an enforcement-authority grant, not an affirmative compliance duty on developers, though it establishes that developers must be prepared to be audited.
The Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, shall adopt and promulgate, pursuant to the "Administrative Procedure Act," P.L.1968, c.410 (C.52:14B-1 et seq.), rules and regulations as may be necessary for the implementation of this act.
Section 6 directs the Attorney General, in consultation with the Office of Homeland Security and Preparedness, to adopt implementing rules under the Administrative Procedure Act. This delegates future regulation of the act's disclosure requirements but imposes no direct obligation on covered developers.
No later than one year prior to the expiration of sections 1 through 5 of this act, the Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, shall submit to the Legislature a list of recommendations for revisions to the definitions of catastrophic harmCatastrophic harm"Catastrophic harm" means the death or serious injury of 25 or more people or $1 billion in damage to, or loss of, property arising from a single incident involving a frontier model: a. providing expert-level assistance in the creation or use of an illegal firearm, lethal autonomous weapon, chemical weapon, biological weapon, radiological weapon, nuclear weapon, or explosive device; b. engaging in conduct with no meaningful human oversight, intervention, or supervision that, if it had been committed by a human, would constitute the crime of murder under N.J.S.2C:11-3, assault under N.J.S.2C:12-1, or theft under chapter 20 of Title 2C of the New Jersey Statutes, including theft by deception under N.J.S.2C:20-4 and theft by extortion under N.J.S.2C:20-5; or c. evading the control of its frontier developer or user, such as using deceptive techniques against the frontier developer or user to subvert the controls or monitoring of its frontier developer or user outside of the context of an evaluation designed to elicit this behavior. "Catastrophic harm" shall not include: incidents arising from information that a frontier model outputs if the information is otherwise publicly accessible in a substantially similar form from a source other than a foundation model; or lawful activity of the federal government.Section 1, large frontier developerLarge frontier developer"Large frontier developer" means a frontier developer that together with its affiliates collectively had annual gross revenues in excess of $100 million in the preceding calendar year.Section 1, artificial intelligence technologyArtificial intelligence technology"Artificial intelligence technology" means any technology involving the development or implementation of machine-based systems that, for explicit or implicit objectives, infer, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that may influence physical or virtual environments.Section 1, foundation modelFoundation model"Foundation model" means an artificial intelligence technology that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks.Section 1, and frontier modelFrontier model"Frontier model" means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, including computing for the original training run and for any subsequent fine-tuning, reinforcement learning, or other material modifications the developer applies to a preceding foundation model.Section 1, and any other provisions of this act the Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, sees fit.
Section 7 requires the Attorney General, in consultation with the Office of Homeland Security and Preparedness, to submit recommendations to the Legislature for revising the act's key definitions no later than one year before the act's expiration. This is a reporting duty on the agency, not on developers.
This act shall take effect six months after the date of enactment. The Attorney General, in consultation with the officeOffice"Office" means the New Jersey Office of Homeland Security and Preparedness.Section 1, may take any anticipatory administrative action in advance as shall be necessary for the implementation of this act. Sections 1 through 6 of this act shall expire five years after the effective date. Section 7 of this act shall expire upon the submission of a list of recommendations to the Legislature required pursuant to section 7 of this act.
Section 8 sets the operative timing: the act takes effect six months after enactment, with anticipatory administrative action permitted in advance. Sections 1 through 6 sunset five years after the effective date, and Section 7 expires upon submission of the required recommendations to the Legislature.