The frontier AI safety law the labs helped write

Illinois has just passed the first US law that makes frontier AI labs submit their safety claims to outside audit. It is a real piece of regulation, and it was endorsed by the companies it regulates. Both of those things are true, and the second one is the part worth sitting with.

What the law does. SB 315 covers developers of frontier models — companies above roughly $500 million in revenue, training at high compute thresholds, which in practice means OpenAI, Anthropic, Google DeepMind, and a short list of peers.¹ They must publish a safety framework, submit it to annual independent third-party audit, report critical safety incidents within 72 hours, and not retaliate against whistleblowers. The Illinois attorney general can fine violators up to $3 million per breach. The law takes effect in 2028.²

This is more than California SB 53 or New York's RAISE Act, both of which stop at disclosure. SB 315 is the first US statute that says: you do not just have to tell us your safety plan, someone independent has to check whether you are following it.

Who was in the room. Anthropic's policy lead Cesar Fernandez and OpenAI's Jamie Radice both publicly supported the bill while it was being negotiated.³ The Wired write-up frames this as labs welcoming oversight. I think it is more interesting than that, and more worth examining.

The Trump administration cancelled federal AI model vetting earlier this year. Whatever you think of the previous federal regime, its disappearance left a vacuum. Into that vacuum walked state legislators in Sacramento, Albany, and Springfield. The labs now face a choice they did not face eighteen months ago: which state-level regime do they prefer to be governed by, and which can they help shape before it hardens?

I do not think Anthropic and OpenAI supported SB 315 because they suddenly decided regulation was good for them. I think they supported it because the alternatives — a federal regime they cannot predict, a California bill with sharper teeth, a fifty-state patchwork written by people who do not return their calls — are worse. That is a rational corporate position. It is also a position that should make the rest of us look carefully at what got written into the statute and what did not.

What did not get written in. The people most exposed to frontier AI systems are not the labs. They are workers whose employers are deploying these models into hiring, scheduling, and performance management. They are patients whose clinicians are running diagnostic copilots. They are students, tenants, benefits claimants, defendants. None of those constituencies had a seat at the table when SB 315 was negotiated. The bill's named public supporters are the regulated companies.

This is not a scandal. It is how almost all technical regulation gets written. But it shapes what the law is for. SB 315 is about catastrophic risk, model capabilities that could cause large-scale harm, and about whether labs are honestly assessing those capabilities. It is not about whether the AI system your landlord uses to screen tenants is doing what your landlord says it does. Those are different harms, and SB 315 does not touch the second kind.

When the press release says Illinois is "setting the standard for the rest of the country," it is worth asking: a standard for what? For frontier model safety, yes. For the ordinary, accumulating harms of AI deployment in workplaces and public services, no. Those remain unregulated, and the people living inside them remain unconsulted.

The audit problem. The hinge of SB 315 is the independent third-party audit. Everything depends on whether that audit is real.

Here is the gap. There is no established profession of frontier AI safety auditors. The Big Four accounting firms have the statutory experience and the client relationships, but they do not have the technical capacity to interrogate a lab's claims about model capabilities or evaluation methodology. The specialist evaluators who do have that capacity, METR, Apollo Research, a handful of others, are small, underfunded, and have never operated as statutory auditors for billion-dollar clients.⁴

One of two things will happen. Either the Big Four hire the specialists, the specialists get absorbed into a compliance industry, and the audits become real but expensive. Or the audits become a paperwork exercise — a firm with the right letterhead and the wrong expertise signs off on a framework it cannot actually evaluate. Which one we get is not specified by the statute. It will be specified by the implementing regulations, by the attorney general's enforcement posture, and by what the labs are willing to pay for.

The whistleblower clause. The most consequential sentence in the bill may be the one getting the least attention. SB 315 protects employees at covered labs who report safety concerns. Lab employees have historically faced NDAs, equity clawbacks, and reputational pressure for raising internal alarms; OpenAI's 2024 non-disparagement clause episode is the canonical case. A statutory shield, even one limited to Illinois, changes the calculus for an employee deciding whether to speak.

It also raises a hard question the statute does not answer. Most frontier-lab employees do not live in Illinois. Whether the protection extends to a San Francisco engineer reporting a concern about a model deployed to Illinois users is the kind of jurisdictional question that will be litigated for years. The protection is real. Its reach is unclear.

What to watch. Three things. First, who the attorney general's office certifies as qualified auditors, and whether that list includes any organisation with the technical depth to push back on a lab's claims. Second, whether the whistleblower protection gets tested — and how the courts handle the multi-state employment question when it does. Third, whether the next state to pass an audit law writes one that the labs did not help draft.

The benefits of SB 315 are real. Independent audit of safety claims, even imperfect audit, is more than the federal government is currently providing, and more than any other state has yet mandated. I am not arguing against the law. I am arguing for reading it with clear eyes.

A law endorsed by the people it regulates can still do useful work. It is just less likely to do the work the regulated do not want done. The question for the next legislature, in the next state, is whether the people who actually live downstream of these systems get to be in the room when their version is written.

Footnotes

¶