Reverse federalism, and the AI rules that reach a business your size

A law passes in a state you have never billed a customer in. The coverage calls it landmark. You read three paragraphs, cannot work out whether it is about you, and close the tab. Nine months later the same law has been delayed, amended, or repealed and rewritten. Nothing in that cycle tells a business of four people whether it needs to change anything before Tuesday.

There is a shape underneath it, and it is simpler than the coverage suggests. Sort the laws by who they bind and you get two piles: rules for the handful of companies that train the largest models, and rules for anyone who points an AI system at a member of the public. You are in the second pile, and your side of it is short enough to finish this afternoon. This guide sorts the piles, names what the first one gives you for free, and lists what to change. It is not for anyone training foundation models, who needs counsel rather than a guide, and it is not legal advice for a regulated business in health, lending, insurance or hiring, where everything below is a floor and your sector regulator sets the ceiling.

The labs asked for the rules, and gave the strategy a name

In May 2026, OpenAI’s global affairs team published a post arguing that California, New York and Illinois were converging on a single frontier-safety baseline, and that this was a feature rather than a mess. “Think of it as reverse federalism: states leading in a way that helps move the country forward as we head further into the Intelligence Age” [1]. The same post endorsed independent third-party audits, singling out Illinois’s SB 315 for “going even further by including independent third-party audits” [1].

In July the company repeated the argument in its own newsroom, describing reverse federalism as “states helping establish a shared direction through a common framework” and naming three elements it wants present in each state law: documented safety frameworks with risk assessments, reporting of serious safety incidents, and independent audits for governance and accountability [2]. It asked that “policymakers should also guard against mission creep”, argued that “states should not be asked to manage significant national security risks”, and was blunt about the alternative on offer: “Neither an undefined federal process nor a patchwork of state laws will produce a coherent frontier safety regime” [2]. It still wants Congress to act. “A federal framework remains essential” [2].

You can read a company lobbying for its own regulation cynically, and the cynical read is not wrong: shaping a rule while it is being drafted beats complying with one you had no hand in. The more useful read is that the motive does not change the signal. A developer volunteering for expanded incident reporting and outside audits has concluded that some version of both is arriving regardless. That is a firmer prediction of where mandatory requirements are heading than any forecast, because it comes from the party with the strongest incentive to argue the other way.

The laws with the big numbers are aimed several sizes above you

California’s Transparency in Frontier Artificial Intelligence Act, SB 53, was chaptered on 29 September 2025 [3]. It defines a frontier model as “a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations”, and a large frontier developer as one that “together with its affiliates collectively had annual gross revenues in excess of five hundred million dollars ($500,000,000)” in the preceding calendar year [3]. Those developers must publish a frontier AI framework, publish a transparency report at or before deployment, and “report any critical safety incident pertaining to one or more of its frontier models to the Office of Emergency Services within 15 days of discovering the critical safety incident” [3]. An incident that “poses an imminent risk of death or serious physical injury” goes within 24 hours to “an authority, including any law enforcement agency or public safety agency with jurisdiction” [3]. Developers also may not block employees who disclose catastrophic risks to the Attorney General or federal authorities, and civil penalties run to $1,000,000 per violation, enforced by the Attorney General [3].

New York followed. Governor Hochul signed the RAISE Act on 19 December 2025, in a law that “requires large AI developers to create and publish information about their safety protocols” and to “report incidents to the State within 72 hours of determining that an incident occurred”, with an oversight office inside the Department of Financial Services and penalties “up to $1 million for the first violation and up to $3 million for subsequent violations” [4]. Illinois went further on the audit question. Governor Pritzker signed the Artificial Intelligence Safety Measures Act on 6 July 2026, effective 1 January 2027, making Illinois “the first state in the nation to require regular independent third-party safety audits of covered AI systems, ensuring oversight is conducted by qualified experts without financial conflicts of interest”, alongside public disclosure of safety practices, incident reporting, and whistleblower protections [5].

Three laws, three states, one shape. And none of them is about you. The trigger in California is the compute used to train the model, not how much of it you buy or how many customers you serve [3]. These are supply-chain laws. They regulate the companies whose models you rent.

What the frontier laws hand you as a buyer

The side effect is the part worth using. Documents that were previously voluntary blog posts are now legal artefacts published at a known address. SB 53 requires a large frontier developer to “write, implement, comply with, and clearly and conspicuously publish on its internet website a frontier AI framework” describing, among other things, its thresholds for identifying capabilities that could pose a catastrophic risk, the mitigations it applies, how it reviews those assessments “as part of the decision to deploy a frontier model”, its use of “third parties to assess the potential for catastrophic risks”, its “cybersecurity practices to secure unreleased model weights”, how it identifies and responds to critical safety incidents, and its internal governance practices [3]. Separately, “before, or concurrently with, deploying a new frontier model or a substantially modified version of an existing frontier model, a frontier developer shall clearly and conspicuously publish on its internet website a transparency report” naming release date, supported languages, output modalities, intended uses and restrictions, with large developers also summarising catastrophic risk assessments, results, and third-party evaluator involvement [3].

That is free due diligence on your critical supplier. Before you make a model load-bearing in your business, open the transparency report for the exact version you are calling, not the vendor’s marketing page for the family. Restrictions and intended uses in that report are the vendor’s own account of what the thing is not for, which is the sentence you want on file if a customer ever asks why you used it.

The incident clocks are the second useful piece. A vendor problem that materially affects the model under your product now has to reach a state body on a schedule, 15 days to California’s Office of Emergency Services and 72 hours to New York, rather than at a moment of the vendor’s choosing [3][4]. That does not put the news in your inbox, but it converts “we will hear about it eventually” into a process with a deadline attached.

Do not over-read it. None of these laws is about accuracy, uptime, price, latency or how much notice you get before a model you depend on is retired. They are about catastrophic risk and disclosure. Your ordinary operational risks are still yours to manage and still sit in the vendor’s contract, not in the statute.

Two duties reach you, and both are short

The first is disclosure. On companion chatbots, state legislators “introduced more than 100 bills and enacted 14” in 2026, and most of the new laws build on the two California and New York passed in 2025, which “require operators to include warnings that chatbots are not human and to address certain risks, including sexual content involving minors and self-harm” [8]. Most of those laws target companion products rather than the support widget on your pricing page, and scope varies by state, but the compliance action sitting at the centre of all of them is a line of text.

The second is decisions about people. Colorado is the clearest current statement of it. SB 26-189 was signed on 14 May 2026 and, in the legislature’s own summary, “repeals and reenacts” the state’s earlier provisions with new requirements for automated decision-making technology in consequential decisions [7]. A consequential decision is one that relates to “an individual’s access to, eligibility for, or compensation related to education, employment, housing, financial or lending services, insurance, health-care services, or essential government services and public benefits” [7]. Deployers, meaning anyone using the technology rather than building it, must give “clear and conspicuous notice to consumers at the point of interaction” and, within 30 days of a consequential decision that produces an adverse outcome, “a plain language description” of the technology’s role [7]. The requirements begin 1 January 2027 and the Attorney General enforces them, but “before initiating an action before January 1, 2030, the attorney general must provide the developer or deployer with a 60-day notice and opportunity to cure the alleged violation, if a cure is deemed possible” [7].

Strip the two piles down and your obligations are the same in both: say it is AI at the moment somebody meets it, and be able to explain what the system did when it went against someone. If you screen applicants, score enquiries for eligibility, set a price per customer, or triage anything about a person, you are inside the second one whether or not the word “AI” appears in your marketing.

The rules move faster than a compliance folder

Colorado is also the worked example of volatility. Its earlier automated-decision provisions were repealed and reenacted in May 2026, replaced by a different structure with a start date of 1 January 2027 [7]. Compliance work keyed to the wording of the previous version bought nothing. The habit of writing down where automation touches a person carried straight across.

Pressure runs the other way too. On 11 December 2025 an executive order directed the Attorney General to establish, within 30 days, an AI Litigation Task Force “whose sole responsibility shall be to challenge State AI laws inconsistent with the policy set forth in section 2 of this order”, told Commerce to issue guidance under which “States with onerous AI laws identified pursuant to section 4 of this order are ineligible for non-deployment funds, to the maximum extent allowed by Federal law”, and ordered a legislative recommendation “establishing a uniform Federal policy framework for AI that preempts State AI laws” [6]. So the reverse-federalism bet and a federal preemption push are running at the same time, on the same laws.

The friction shows in the count. “As of July 1, states have enacted 109 AI and 28 data center laws”, against 121 AI and 27 data centre laws by the same date the year before, and 29 states have enacted AI legislation this year where 39 had done so by that point in 2025 [8]. Fewer states legislating, and the ones that do steering toward categories least likely to be challenged.

The operator conclusion is not to track any of this. It is to refuse to build a compliance programme around a date. Build the two habits that survive every version, every court order and every preemption attempt: a notice where a person meets the system, and a record of what the system did. Both of those are also just good practice, which is the test of whether a compliance task is worth doing at your size.

Price the first two items before you assume they are expensive.

For most small businesses that lands under half a day, which is the real finding. The disclosure duty is cheap, and the reason it goes undone is that nobody has listed the surfaces, not that anybody weighed the cost and declined.

What still goes wrong

Every date, figure and quotation above was read from its source on 5 September 2026, and this is the fastest-moving body of law an operator has had to think about. Two of the statutes cited here were signed in the last four months, one of them repealing and reenacting an earlier Colorado law [7], and a federal preemption effort is actively aimed at all of them [6]. Treat the specifics as a snapshot with a short shelf life and the structure as the durable part.

The scope question is genuinely hard, and a guide cannot answer it for you. Whether your support widget is a companion chatbot in a given state, whether your lead scoring is covered automated decision-making technology in Colorado, whether a sector carve-out swallows the general rule, whether the law follows your customer or your incorporation, are all questions with per-state answers and defined terms that do not match each other. Sorting the piles tells you where to look. It does not tell you where you landed, and if the answer changes what you can charge or who you can hire, that is a question for a lawyer in that state.

The largest limit is that the whole arrangement described here might not hold. Reverse federalism is a bet by one company that state convergence produces a workable national floor, made while the same company argues a federal framework is essential and the administration works to preempt the states outright [2][6]. If preemption succeeds, the published frameworks and transparency reports become voluntary again and the leverage a small buyer gets from them evaporates. The two habits survive that outcome intact, which is the honest reason to invest in them rather than in tracking the bills. Disclosure and explainability were worth doing before any of these laws existed, and they will be worth doing after whichever of them survives.