Illinois AI audits ab 2028: first U.S. state analysis

The calendar entry is no longer a conference panel. It is a release dependency. Illinois has moved AI safety compliance from the familiar territory of internal documentation into the less forgiving world of independent audits. That matters because auditability is not something a team sprinkles over a model after launch. Logs, evaluation records, incident workflows, governance signoffs, and vendor evidence all have to exist before a third party can verify them. The practical question is not whether regulators like safety frameworks. It is whether your product process can prove one operated when it counted.

What Illinois actually enacted

According to Skadden, Illinois Gov. JB Pritzker signed the Artificial Intelligence Safety Measures Act on July 6, 2026, making Illinois the third state, after California and New York, to impose transparency, safety, and reporting obligations on large AI developers. Skadden says Illinois goes further by requiring developers to retain an independent third party to audit compliance annually. DLA Piper similarly describes Illinois as the first state to require third-party audits of AI models, while Governing reported earlier that the bill would require major AI developers to disclose risks, report safety incidents, and submit to annual independent audits. The headline version is simple: Illinois is first on external verification. The lawyer version is narrower and more useful: covered large frontier model developers face annual independent compliance audits, plus a public framework obligation that begins on January 1, 2028. If your compliance memo says only that the model card is done, it is missing the point.

Who is in the compliance lane

Skadden says the Act centers on a frontier AI framework and applies to a large frontier developer. Beginning January 1, 2028, that developer must write, implement, comply with, and conspicuously publish on its website a framework describing its approach to catastrophic risk management. Skadden also says the framework must be reviewed at least annually, and any material modification must be published with a justification within 30 days. Crowell & Moring frames the same law as transparency and safety obligations for frontier AI systems, and says Illinois joined California and New York in adopting standards for the most powerful AI systems. That is a useful boundary marker for builders who are not training frontier models but rely on them. You may not be the direct statutory target, but procurement has a long memory and a short questionnaire.

What changes for builders Cooley describes

AISMA as a move from transparency to verification. Its analysis contrasts prior frameworks, including California's TFAIA, New York's amended RAISE Act, and portions of the EU AI Act, which generally require developers to assess and disclose how they identify, evaluate, and manage AI risks. Illinois adds the awkward part: an independent third party has to check whether the process actually operates as described. Translated into product work, that means compliance evidence has to be designed into the system. The release checklist needs ownership for risk assessments, model or system cards, incident reporting paths, public framework updates, and audit cooperation. A vendor contract touching a frontier model should ask for documentation delivery, incident notice, and audit support, not because every customer is directly regulated, but because regulated duties travel through procurement. DLA Piper notes that AISMA resembles California and New York on transparency obligations, with the prominent exception of the audit provisions. That exception is where product teams should spend time. A public framework is a document. An audit is a document plus proof, timestamps, people, controls, and the occasional uncomfortable meeting.

Where jurisdictions now diverge

Cooley's comparison is the problem builders will recognize: California, New York, and parts of the EU AI Act emphasize assessment, disclosure, governance frameworks, transparency reports, model cards, risk assessments, and incident reporting. Illinois keeps much of that architecture, then adds independent verification. Crowell & Moring says that even without federal action, California, New York, and Illinois have created what is essentially a national framework for AI safety and transparency. That does not mean the rules are identical. It means a large developer serving U.S. customers may have to satisfy the strictest operational expectation even where another state is content with self reported controls. The compliance center of gravity is moving from what a developer says it does to what an auditor can verify it did. For readers building with frontier systems, the next useful step is boring and immediate: map which evidence your current development lifecycle already creates, which evidence disappears in chat threads, and which evidence a third party could actually inspect. By 2028, the interesting companies will not be the ones saying they welcome the rules. They will be the ones whose launch process already leaves an audit trail.