Illinois Passes America's First Law Mandating Third-Party AI Safety Audits

When the Illinois House of Representatives voted 110 to 0 in favor of Senate Bill 315 on May 27, 2026, it wasn’t just a legislative milestone for one state — it was the first time any legislature in the United States had passed a law requiring frontier AI companies to submit to mandatory, annual, independent safety audits. Governor J.B. Pritzker, who has publicly called the bill a model for responsible AI governance, has said he will sign it into law.

The Artificial Intelligence Safety Measures Act targets the most powerful AI systems in existence and imposes obligations that the largest AI labs in America have so far managed to avoid at the federal level. With Washington’s approach to AI policy remaining deregulatory under the current administration, Illinois has stepped into the vacuum — and the reception from the industry’s two largest players has been, unusually, supportive.

What the Law Requires

SB 315 establishes four core obligations for covered companies:

Catastrophic-Risk Frameworks: Frontier developers must create, publish, and annually update a comprehensive safety framework that addresses catastrophic-risk assessment, proposed mitigations, cybersecurity posture, internal governance structures, third-party evaluations, and risks arising from the company’s own internal use of frontier models. The framework must be publicly available — not a confidential submission to regulators — meaning competitors, researchers, and the public can scrutinize each company’s stated approach.

Annual Independent Audits: Each year, covered developers must hire an independent third-party auditor to verify that their safety practices match their published frameworks. This is the provision that has no precedent in existing US law; financial audits for public companies are a rough analogy, but AI safety auditing is a considerably less mature field, and the bill’s implementation will require developing standards that do not yet fully exist.

Incident Reporting: Companies must report significant AI safety incidents to Illinois state officials within 72 hours of discovery. If an incident poses an imminent risk of death or serious physical harm, that reporting window compresses to 24 hours. The specificity of the thresholds — particularly the 24-hour rule for imminent harm — suggests drafters were thinking about scenarios like AI systems providing dangerous instructions, enabling biological or chemical attacks, or producing outputs that directly trigger physical harm.

Whistleblower Protections: The law creates formal protections for employees who raise internal safety concerns, report incidents to regulators, or otherwise surface potential violations. Given the power asymmetries in the AI industry — where a handful of companies control frontier capabilities and employ relatively few workers — this provision could be consequential if it creates a functional channel for researchers with safety concerns to act without career risk.

Who Is Covered — and Who Isn’t

The bill’s scope is deliberately narrow. It applies only to companies with more than $500 million in annual gross revenue that also build or deploy models meeting a frontier-scale compute threshold. In practical terms, that combination captures OpenAI, Anthropic, Google, Meta, and a small number of other large developers. It does not apply to the thousands of companies that deploy AI built on top of frontier models without developing those models themselves.

This targeting choice was intentional and reflects a judgment that the primary safety risks from AI systems arise at the frontier — the most capable, most resource-intensive models — rather than across the entire AI software stack. Critics of the approach have argued that concentrating obligations on frontier developers creates a compliance threshold that startups will structure themselves to avoid, but proponents counter that the resource requirements for frontier model development effectively limit the relevant population to a manageable set of large organizations.

Civil penalties for noncompliance are set at up to $3 million per violation, enforced solely by the Illinois Attorney General. The per-violation structure, rather than a flat fine, means that a company with systematic compliance failures across multiple requirements could face aggregate penalties substantially larger than the per-instance cap suggests.

Unusual Industry Support

The most striking feature of SB 315’s passage was not its content — many of its provisions mirror frameworks proposed by AI safety researchers and policy advocates — but rather who supported it. OpenAI and Anthropic, the two companies that would face the most significant compliance obligations if the law takes effect, both publicly backed the bill.

Their support likely reflects a combination of strategic and genuine safety motivations. Strategically, large frontier labs benefit from regulatory frameworks that impose costs their smaller, less-resourced competitors cannot easily absorb. A mandatory annual audit regime with a $500 million revenue threshold effectively grandfathers the current frontier developers into a regulatory structure that would be much harder for new entrants to satisfy. Anthropic in particular has long argued that external accountability mechanisms would help demonstrate that safety-focused labs are actually doing what they claim — supporting audits is consistent with that positioning.

The opposition came primarily from NetChoice, a technology trade association representing a broader set of AI companies, which raised concerns about vague definitions and implementation uncertainty. The trade group’s arguments — that terms like “frontier model” and “catastrophic risk” lack sufficient legal precision — are not without merit, but they are also the standard arguments made against most early-stage technology regulation.

The Federal Vacuum Illinois Is Filling

SB 315 exists in the context of a deliberate federal retreat from AI regulation. The Trump administration’s National AI Policy Framework, released earlier in 2026, positioned the US approach as deregulatory and explicitly expressed concern about state-level AI laws creating a compliance patchwork that disadvantages American companies. Several Congressional proposals to preempt state AI regulation have been floated, though none has advanced to a vote.

Illinois’s unanimous House passage — across party lines — suggests that AI safety is achieving a bipartisan political salience that cuts against the deregulatory consensus at the federal level. Governor Pritzker’s framing has leaned heavily on the argument that voluntary commitments by AI companies are insufficient: “We cannot afford to wait for a catastrophic failure to find out whether the largest AI companies are actually managing the risks they acknowledge exist.”

The state has a particular vantage point on frontier AI: while it is not a major AI startup hub in the way that California’s Bay Area is, Illinois is home to major financial institutions, healthcare systems, and critical infrastructure operators that are among the largest enterprise users of frontier AI systems. A safety failure in those sectors would have direct consequences for Illinois residents.

Implementation Realities

The most significant practical challenge the law faces is that AI safety auditing does not yet have established professional standards. Financial auditing has the PCAOB; environmental auditing has EPA certification frameworks. AI safety auditing currently has advocacy documents, research papers, and a small number of early-stage firms attempting to build the practice — but nothing resembling the kind of standardized, legally recognized auditing infrastructure that SB 315 envisions.

The law takes effect in early 2027, giving the industry and the auditing ecosystem roughly one year to develop workable standards. Illinois regulators will need to define what constitutes a qualifying audit, what credentials third-party auditors must hold, and what constitutes an auditable safety practice versus a non-verifiable internal policy commitment. These are genuinely hard questions, and the answer to whether SB 315 produces meaningful accountability or performative box-checking depends heavily on how regulators answer them.

For the frontier AI companies that will be subject to the law, the near-term compliance question is not primarily financial — $3 million per violation is not a meaningful constraint on companies operating at OpenAI’s or Google’s scale — but operational. Building the internal systems to produce auditable safety documentation, maintain incident logs that meet the 72-hour reporting requirement, and manage third-party auditor relationships is non-trivial. Companies that have not already been building these capabilities for their own internal governance will need to build them for Illinois.

A National Model or a One-State Experiment?

The precedent question hangs over every state-level AI regulation: does Illinois’s law become the first of a national wave, or a one-state outlier that the federal government eventually preempts?

The California comparison is instructive. When California passed CCPA and then CPRA, the rest of the country eventually moved toward privacy frameworks that roughly paralleled California’s approach — not because the federal government preempted CCPA, but because the market cost of California-specific compliance made universal adoption the rational corporate choice. If Illinois’s AI safety requirements become the default standard that frontier developers adopt nationally to avoid a patchwork of state variations, the law will have achieved its goals even without federal action.

The unanimous House vote, the governor’s enthusiastic support, and the unusually broad industry backing suggest that SB 315 has more political durability than most state-level tech regulation. Whether it produces real safety accountability depends on the implementation. But as the first law of its kind in the United States, it has already done something important: it has established that mandatory AI safety audits are legally feasible, politically viable, and supported — at least in principle — by the companies that would be subject to them.