On a gray morning in Lansing, a legislative aide scrolled through yet another briefing packet about machines that write, hire, diagnose, and sometimes invent facts. The packet was thick. The calendar was thin. Committees had hearings scheduled, then postponed, then rescheduled around budget fights that always seem more urgent than code. Across the country, the same pattern holds: executives who built the tools are asking for guardrails, and the people who write statutes have not agreed on one. Michigan AI regulation sits in that gap, named in speeches and still missing from the statute books in any form a company, a school, or a patient could rely on.
A request that surprises almost nobody who builds the tools
Chief executives at the largest artificial intelligence firms have spent the past two years telling Congress, governors, and anyone with a microphone that they want rules. The request is not altruism dressed up as civic duty. Clear law is cheaper than fifty different lawsuits and easier to explain to investors than a rumor about what a state attorney general might do next spring. A federal floor, they argue, would tell engineers what they must test, what they must disclose, and what they may not ship into a hospital or a hiring desk.
Washington has not delivered that floor. An executive order can steer agencies for a season and vanish with the next administration. President Trump has favored a lighter federal hand and has not signed a comprehensive statute, because Congress has not sent him one. Hearings multiply. Drafts circulate. Nothing with the force of a single national standard has cleared both chambers and survived the politics of tech donors, labor unions, and state officials who do not want to be told their own laws are void.
What the chiefs say they can live with
Strip away the conference slogans and the ask is narrower than the headlines suggest. Company leaders tend to accept rules on safety testing for the most powerful systems, on labeling when a person is talking to a machine, and on liability when a product causes a concrete harm that existing tort law already knows how to name. They fight harder against open ended duties to predict every misuse, and against state bans that would force them to build a different product for Michigan than for Ohio.
That distinction matters. A rule that says a model used in medicine must log its sources and flag low confidence answers is the sort of requirement a general counsel can implement. A rule that says the same model must never produce a wrong sentence is a wish, not a statute. Executives who ask for regulation are usually asking for the first kind. Lawmakers who stall are often stuck between constituents who want the second kind and industry lawyers who will litigate anything that looks like a ban.
Lansing and the unfinished file
Michigan has not been silent. Lawmakers have floated bills on deepfakes in elections, on automated decision tools in public benefits, and on the use of student data by education software. None of those efforts has produced the comprehensive framework advocates describe when they say Michigan AI regulation. The file remains a stack of partial ideas: a disclosure here, a study commission there, a press release after a hearing that draws more lobbyists than voters.
The practical result is familiar to anyone who has watched autonomous vehicle policy in this state. Michigan knows how to convene. It is less practiced at finishing. Auto suppliers, hospitals in Detroit and Grand Rapids, and universities with large research budgets all operate systems that already sort resumes, draft clinical notes, or flag fraud. They do so under privacy statutes written for an earlier internet, employment law written for human managers, and consumer rules that never contemplated a chatbot inventing a refund policy.
Why one standard keeps slipping away
Federalism is the polite explanation. States may regulate health, schools, elections, and consumer fraud. Congress may regulate interstate commerce. Artificial intelligence touches all of those at once, so every committee can claim a piece and none can claim the whole. The impolite explanation is coalition math. A bill strict enough for civil rights groups loses moderate Republicans. A bill loose enough for industry loses Democrats who remember social media promises that aged badly. A bill that tries to preempt state law angers attorneys general who see preemption as a gift to companies headquartered elsewhere.
Trump has not forced the issue with a detailed legislative proposal of his own, and Democratic leaders have not assembled a majority that can survive the Senate. The result is not a libertarian paradise and not a precautionary state. It is drift. Agencies issue guidance. Courts inherit disputes that statutes should have settled. Companies write their own policies and call them compliance.
Workers who meet the software before the law does
In a call center outside Pontiac, a supervisor described software that suggests replies and scores tone. She did not use the language of model weights. She talked about quotas and about a new hire who trusted a suggested sentence that promised a credit the company does not offer. Nobody in that office can point to a Michigan statute that tells the vendor how accurate those suggestions must be, or that tells the employer when a score may be used in a performance review.
National labor groups have asked for notice when monitoring tools watch keystrokes or calls, and for a human appeal when a machine rejects a candidate. Those are old principles in new clothing. They do not require a legislature to understand transformers. They require a legislature to decide that an automated decision is still a decision, with a person responsible for it. Until that sentence is law, the person responsible is whoever happens to be on shift when the error lands.
Schools, clinics, and borrowed trust
Public trust arrives before public rules. Teachers in Ann Arbor and Flint already see essays that read like polished brochures and homework that collapses when a student is asked to explain a paragraph aloud. Hospital systems pilot tools that draft discharge notes. Pastors and counselors hear congregants quote a chatbot as if it were a second opinion on grief. The category of spiritual life is not exempt from a technology that answers every question in a confident voice.
Faith leaders are not asking Lansing to theologize. They are asking, in quieter rooms, whether a tool used by a counseling ministry must say it is a tool, and whether a church office that stores pastoral notes in a cloud service has any duty beyond the privacy policy nobody reads. Those questions rhyme with the ones hospitals ask. They will be answered by contract and habit if they are not answered by statute.
The cost of fifty partial answers
A patchwork is not the same as no rules. California, Colorado, and a handful of other states have moved on automated decision making, political deepfakes, or developer disclosures. Michigan employers who sell across state lines already comply with the strictest customer on the invoice. That is a kind of regulation, chosen by procurement departments rather than by voters.
It is also unstable. A firm can exit a small market. It cannot easily exit the idea of liability. When a Michigan resident is harmed by a system trained in another state and hosted in a third, the lawsuit will not wait for a model bill from a national association. Judges will stretch old doctrines. Juries will do what juries do. Executives who say they want Michigan AI regulation, or a federal substitute that covers Michigan, are partly trying to avoid that lottery.
What a usable statute would actually contain
A usable law would not try to define intelligence. It would name uses. Hiring. Credit. Housing. Medical advice. Election advertising. Policing. For each use, it would require a plain language notice, a way to reach a human, and a record that survives an audit. It would set different duties for a hobby chatbot and for a system that denies a benefit. It would say which state rules yield when a federal rule covers the same harm, and which state rules remain because they protect elections or consumer fraud in ways Congress left alone.
None of that is exotic. Europe has tried a risk ladder. American agencies already tier oversight for drugs, aircraft, and banks. The novelty is political, not technical. Someone must disappoint a donor. Someone must tell a colleague that a favorite symbolic ban will not survive court review. Someone must schedule the vote on a week when the budget is not on fire.
Voices from the industry, without the applause line
It is fair to be skeptical when a powerful firm asks to be regulated. History is full of incumbents who wrote the rules to lock the door behind them. Small developers in Ann Arbor and Detroit worry that a heavy testing regime will be a tax only large labs can pay. That worry is legitimate, and it is a reason to scale duties to risk and revenue, not a reason to do nothing.
It is also fair to notice what the chiefs are not asking for. They are not asking for a ban on open research. They are not, in most public letters, asking for criminal penalties aimed at teenagers who jailbreak a toy. They are asking for predictability. Predictability is not the same as wisdom. A bad predictable rule can freeze a harmful practice in place. A missing rule freezes nothing except accountability.
Where Trump and the states leave the map blank
The White House under Trump has emphasized competition with China, energy for data centers, and a suspicion of what officials call European style bureaucracy. Those are governing choices, and voters can judge them. They are not a substitute for a sentence that tells a Michigan hospital what it must disclose when software drafts a note in a patient’s chart. State leaders who wait for Washington are making a choice too. Waiting is a policy. It allocates risk to the people least able to negotiate a master service agreement.
Local politics will not stay local if a campaign deepfake swings a state House race, or if an automated benefits error hits a county that already distrusts its government. The next scandal will produce a rush bill. Rush bills are how bad definitions enter the code and stay there for a decade.
A narrower bargain still available this session
Michigan does not need to solve the philosophy of machine minds before adjournment. It could pass a disclosure law for political ads that use synthetic media, a requirement that state agencies inventory the automated tools they already buy, and a simple right to a human review when those tools deny a license, a benefit, or a job with the state. Those three steps would not satisfy every executive who wants a national shield. They would give residents something more solid than a task force.
Federal action could still overtake that work. If it does, a clean inventory and a human review right are unlikely to clash with a sane national floor. If it does not, the state will at least have stopped pretending that silence is neutrality.
The stall is the story
Readers should not mistake corporate requests for a moral awakening, and they should not mistake legislative delay for careful craft. The record, in Washington and in Lansing, is mostly absence. Michigan AI regulation remains a phrase in search of a statute. Company chiefs will keep asking for rules they can price. Residents will keep meeting the software anyway, in the call center, the clinic, the classroom, and the quiet hour when someone asks a machine a question they used to ask a person.
A standard, even a modest one, would not end that encounter. It would say who answers when the encounter goes wrong. Until then, the answer is whoever is closest to the screen, and that is not a policy worthy of a state that still knows how to build things.