Lawfare Analysis Calls for Evidence Rules in State AI Incident Reporting

State laws are beginning to create a paper trail for the most serious frontier-AI failures. But a new Lawfare analysis argues that a report alone may not leave regulators with enough evidence to reconstruct an event or prevent its recurrence.

California’s SB 53, New York’s Responsible AI Safety and Education Act, and Illinois’s SB 315 require frontier developers to report specified critical safety incidents to state authorities. The largest covered developers, defined by revenue and high training-compute levels, must also publish AI safety frameworks; Illinois adds an annual independent audit requirement.

The laws set deadlines after an incident reaches a high threshold, not a general channel for every model problem. Covered events include unauthorized access to model weights causing death or bodily injury, catastrophic-risk harm, loss of model control causing death or injury, and deceptive behavior that materially increases catastrophic risk.

The gap begins after notification

The analysis focuses on the next step. It says the laws do not require developers to preserve the model, weights, execution environment, prompts, logs, or other material that could be needed to reconstruct an incident. If those records are not retained, the author argues, later investigators may be unable to establish what happened.

New York and Illinois require their report-receiving bodies to review submissions. Yet the analysis says none of the three regimes requires a developer to investigate an incident or explicitly gives another actor investigatory powers. That is a policy assessment of the statutes, not a current requirement for a new state investigative body.

A record built before the failure

• Require developers to retain incident data so an event can be reconstructed later.
• Require large frontier developers to establish dedicated units to preserve evidence.
• Have those units help investigators access and interpret the preserved material.

Those proposals would move the state regimes beyond notification toward a system designed to support later examination. California is already operating its reporting mechanism, while New York and Illinois are scheduled to start in 2027. Whether lawmakers add retention or investigative duties remains unresolved.