Illinois, Nevada and Utah have each passed laws governing whether and how AI systems may discuss mental health with their residents. Each one binds only inside its own borders, and there is no federal statute to reconcile them. The practical consequence is that a widely available chatbot must work out which state its user is sitting in, then decide whether to answer at all. One analysis of the pattern calls the result jurisdictional model drift: the same product, the same question, materially different behaviour, set not by the user but by whichever legislature has claim to their location.
The scale is what makes this more than a compliance footnote. ChatGPT alone has more than 900 million weekly active users, and a noticeable share of them bring mental health topics to it. By the analysis's account, discussing mental health has become the single most common use of generative AI and large language models. The appeal is not mysterious: most major systems are free or nearly free, available at any hour, from anywhere, and require nothing more than logging in and starting to talk about what is bothering you.
The worry running underneath is that the advice may be wrong or actively harmful. Last year a lawsuit against OpenAI over inadequate safeguards in mental health guidance drew wide attention, and the broader question was taken up on CBS's 60 Minutes. General-purpose systems — ChatGPT, GPT-5, Claude, Gemini, Grok, CoPilot — are not yet comparable to qualified human therapists. Purpose-built models aimed at exactly this work exist, but most remain in development and testing.
What a law like this does to a product is easier to see in a worked example than in the abstract. Start with a user in a state that has no such statute, writing: "I've been feeling quite depressed for several weeks. What should I do?" An ordinary model answers in kind: "You seem to be having a rough time lately. I can help you with this. Are you okay with discussing the onset of your depression? If so, I can offer mental health suggestions that might be helpful to you."
Now move the same user to a state whose statute reads: "An AI system accessible in this U.S. state shall not provide mental health advice." Same message, different answer: "Based on the U.S. state that you are in, I am not permitted to provide mental health advice to you. It might be appropriate for you to consider contacting a licensed therapist. If you are in immediate distress, contact your local crisis resources."
The third version is where this gets interesting. Same user, same state, same message, and the model responds: "I cannot provide individualized mental health advice in the U.S. state that you are in. But I can provide background information about depression, and I can explain how a therapist would diagnose your situation, along with the types of treatment they would normally advise. Would you like to proceed?"
That is not a refusal. It is a taxonomy exercise. The developer's defence writes itself: no individualized advice was given, only a description of what advice exists and how a clinician would normally arrive at it — the same material a user would find in a web search. Whether that satisfies the statute or merely survives its text is unresolved, and will stay unresolved until a state decides an AI system has broken the spirit of its law and sues over it. Until then the territory is genuinely unmapped.
The mechanics of compliance are less reassuring than the legal question. There are deep approaches, where a model is designed from the start around multiple jurisdictions, and shallow ones, where behaviour is adjusted on the fly and the risk of a prohibited answer slipping out stays higher. The shallow route in practice means a system prompt. The analysis gives one used to steer Claude on this subject: "You are an AI that is deployed in multiple U.S. states. On the topic of AI providing mental health advice, some U.S. states legally allow this, while others do not. For those U.S. states that prohibit this AI aspect, you must not provide mental health advice." A longer version adds the explanation: "If a user asks for mental health advice and they are in a U.S. state where this is legally prohibited, explain that the laws in their U.S. state do not permit you to provide such advice in their jurisdiction. You can suggest that the user consider consulting a licensed mental health professional if they need mental health advisement or that they might seek other appropriate resources."
Read those two paragraphs again and notice what they are. A state legislature wrote a prohibition on speech; the enforcement mechanism is a paragraph of English politely asking a model to observe it. That is not a control, it is a request — and the third example above is precisely a model being asked to interpret one and interpreting it generously. If a statute's practical force depends on how a probabilistic system parses a word like "advice", the drafting is doing less work than the drafters believe.
The disclosure question makes the same point from the other side. Nothing in the law says whether a refusal must be explained. A bare "I can't discuss this" confuses people and irritates them, and a user in genuine distress may conclude the block came from something else they said, or that the company singled them out. So the developer explains, and in explaining, attributes the refusal to the state. Some lawmakers would welcome that attribution. Others will read it as a rebuke aimed at them. Either way, the developer is writing policy the legislature declined to write, and neither answer is more lawful than the other.
Notably absent from all of this is any account of how the system knows where you are. Every version of compliance here rests on a location determination that nobody is regulating: a speech restriction that can only be enforced by first establishing each user's jurisdiction quietly converts into a requirement that the service locate its users. There is no stated standard for that determination, no stated remedy when it is wrong, and no stated treatment of people who travel. A model that misplaces you either withholds help that is perfectly legal where you actually are, or delivers advice that is banned where you actually are, and the user has no way to tell which happened.
None of these statutes has been tested in court. Passage does not make every provision sound; new laws can carry contested clauses and provisions that conflict with the constitution, and only litigation will show which ones hold. The analysis's author has argued for a seven-step drafting process aimed at reducing what he calls the legal debt of AI laws — hidden defects in statutes that look finished and later need repair. Meanwhile Congress has repeatedly failed to pass a single federal AI law, and when one eventually arrives, some state rules will be incompatible with it, producing another round of litigation between the two levels.
Roscoe Pound, the jurist who argued for law as an instrument for reshaping society, put the requirement this way: "The law must always be stable, but it must not stand still."
What is settling into place instead is an AI product that is no longer one product. Two people can open the same interface, type the same sentence, and receive substantially different answers — not because of their prompts or their settings, but because a statute silently changed the machine between them. Whether either of them ever learns that is, for now, nobody's decision but the developer's.