Illinois Signs America's Toughest AI Safety Law, Requiring Annual Third-Party Audits of Frontier Models

Illinois Governor JB Pritzker on July 6 signed the Artificial Intelligence Safety Measures Act — Senate Bill 315 — into law, making the state the first in the United States to require that large artificial intelligence developers submit to mandatory annual independent audits. The legislation sets a new national standard for AI oversight and is likely to accelerate pressure on Congress to act before a patchwork of competing state laws creates compliance chaos for the industry.

What the Law Actually Requires

SB 315 targets what it calls “large frontier developers” — a category defined as companies developing frontier AI models and generating more than $500 million in annual gross revenue. Under this threshold, the law’s primary obligations fall squarely on a handful of companies: OpenAI, Anthropic, Google DeepMind, and Meta AI are the obvious candidates, along with any emerging player that crosses the revenue threshold.

The law imposes three core requirements:

Annual third-party safety audits. Large frontier developers must submit to independent external audits every year to verify that their safety policies are real, implemented, and effective. This is the law’s most consequential provision. California and New York have enacted their own frontier AI transparency requirements, but neither mandates recurring external review — only initial or one-time assessments. Illinois’s annual audit cycle means continuous accountability rather than a one-time checkbox.

Public disclosure of safety frameworks. Covered companies must publish policies explaining how they identify and assess “catastrophic risks” — defined in the statute as scenarios involving death or physical injury to 50 or more people, $1 million or more in property damage, or assistance in developing weapons of mass destruction. The disclosure requirement is designed to make safety practices legible to the public, policymakers, and researchers, not just to regulators.

Rapid incident reporting. Covered developers must report safety incidents within 72 hours of discovery, with a tighter 24-hour window for incidents that pose an imminent threat to life or public safety. Periodic summaries of internal risk assessments are also required.

Penalties and Enforcement

The financial penalties are substantial. A first violation carries a fine of up to $1 million. Subsequent violations carry fines of up to $3 million each. The law does not create a private right of action — meaning individuals who believe they have been harmed by a covered AI system cannot sue under SB 315 directly — but civil enforcement authority rests with the Illinois Attorney General.

The law also includes whistleblower protections for employees who report internal violations. This provision could prove to be its most disruptive element: AI safety researchers and engineers inside large companies now have a clearer legal pathway for escalating concerns to regulators without fear of retaliation.

Industry Reactions Reveal a Fracture

The industry response to SB 315 was unusually divided, and the fault lines are telling. OpenAI and Anthropic both publicly supported the bill, a striking departure from the industry’s typically unified opposition to mandatory oversight. Both companies have faced sustained public pressure over their safety practices and appear to see regulated transparency as a competitive advantage — a way to signal seriousness that smaller or less safety-focused competitors cannot easily replicate.

TechNet, a trade organization that represents a broader cross-section of the tech industry including companies with less invested in frontier AI safety branding, opposed the bill. The organization argued that the law risks creating compliance burdens based on “subjective determinations requiring AI safety compliance without established national standards.” That concern is not entirely without merit: the absence of agreed-upon audit methodologies means that early audits may produce inconsistent results across firms.

The bill passed the Illinois House unanimously and cleared the Senate with only five Republican votes in opposition — an unusually bipartisan outcome for any technology regulation, let alone one touching AI.

A National Template in the Making

The significance of SB 315 extends well beyond Illinois. The state has a long history of setting national precedent on data and privacy regulation; the Illinois Biometric Information Privacy Act, passed in 2008, eventually forced major tech companies to change facial recognition practices nationwide and spawned thousands of class-action lawsuits. Illinois’s move into AI safety regulation signals that the same dynamic may be beginning to play out in a much higher-stakes arena.

Representative Daniel Didech, one of the bill’s principal sponsors, was explicit about the urgency. “We have already seen the first AI-inspired mass shooting,” he said on the floor, referring to attacks where AI-generated content or planning tools were implicated. Didech framed the legislation as closing a gap that federal inaction has left open.

At the federal level, U.S. AI regulation remains fragmented. The White House has issued executive orders on AI safety, and the FTC is currently seeking public comment on an AI accuracy policy statement. But Congress has not passed comprehensive AI legislation, leaving states to fill the vacuum. At least eight other states — including Texas, Colorado, and Massachusetts — are reportedly monitoring SB 315’s rollout as a template for their own legislation.

The Audit Problem

The annual audit requirement is ambitious, but implementing it will require solving a problem that the AI industry and academic researchers have not yet solved: there is no agreed-upon methodology for auditing a frontier AI model’s safety properties. Evaluating whether a model reliably refuses to assist with bioweapon synthesis, accurately predicts its own failure modes, or exhibits unacceptable emergent behaviors is vastly more complex than auditing financial statements.

Illinois’s law does not specify audit methodologies, leaving that to regulators and the emerging third-party audit ecosystem to work out before the January 1, 2028 effective date. That 18-month runway is tight. Industry observers note that METR, Apollo Research, and Redwood Research — among the few organizations with meaningful AI safety evaluation experience — would struggle to absorb audit mandates for multiple large frontier developers on an annual cycle without significant capacity expansion.

For the AI industry, SB 315 represents the clearest signal yet that the era of self-regulation is ending. The companies that helped write the rules they will now be audited against have until the start of 2028 to make sure the auditors find what they promised was there.