AI safety regulation
25 deaths or $1 billion sets New Jersey's AI harm line
Senator Raj Mukherji's proposal uses that threshold to define catastrophic harm for frontier models. It also points to weapons help, violent conduct without human oversight and systems that slip past their own controls.
The bill draws a narrow line around the worst-case failures of advanced AI. If those systems create enough death, injury or property damage, covered developers would have to explain their risk controls and disclose the results to the Attorney General.
- Frontier AI developers would have to file safety disclosures with the attorney general.
- The bill uses a narrow trigger: 25 deaths or injuries, or $1 billion in damage.
- It names weapons help, violent conduct and control evasion as warning signs.
- False or missing filings would carry a $100,000 civil penalty.
- In New Jersey, the people most exposed to frontier AI risk are not the developers alone
In New Jersey, the people most exposed to frontier AI risk are not the developers alone. They are the users, the businesses that depend on these systems and anyone who could be harmed if a powerful model goes badly wrong. Senator Raj Mukherji's proposal in Trenton would force large frontier developers with users in the state to make formal safety disclosures to the attorney general.
The measure is aimed at models that do more than autocomplete text. It focuses on frontier models, advanced foundation models trained with very large amounts of computing power, and treats their failures as a public-safety problem, not just a technical one.
The harm line
The bill draws a hard line around what counts as catastrophic harm. That means the death or serious injury of 25 or more people, or $1 billion in damage to, or loss of, property, from a single incident involving a frontier model.
The examples are concrete. They include expert-level help in creating or using an illegal firearm, a lethal autonomous weapon, a chemical, biological, radiological or nuclear weapon, or an explosive device. They also cover conduct with no meaningful human oversight that would amount to murder, assault or theft if a person did it, along with a model evading its developer's or user's control by slipping past monitoring with deceptive techniques.
A report before the model goes live
The core requirement is a new-model risk disclosure. Before, or at the same time as, deploying a new frontier model or a substantially modified version of an existing one, a covered developer would have to submit assessments of catastrophic risk, the results of those assessments, the role of any outside evaluators and other required steps.
Large frontier developers would also have to maintain and update technical and organizational protocols meant to reduce the chance that a model materially contributes to catastrophic harm. The bill defines artificial intelligence technology broadly enough to cover machine-based systems that generate predictions, content, recommendations or decisions from inputs.
Public record, private secrets
The attorney general would publish the disclosures on the Department of Law and Public Safety website, with redactions for trade secrets, cybersecurity, privacy, public safety and national security concerns. The attorney general could also audit covered developers directly or contract with a private entity to do it.
Missing filings, or statements that are materially false or misleading, would bring a civil penalty of $100,000 per violation. If enacted, the measure would take effect six months later and expire five years after that.