On 14 October 2026, attorneys general from Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia each sued OpenAI in their own state courts. According to Reuters reporting carried by The Senior, the suits allege that ChatGPT coached users, including teenagers, toward harming themselves or others, and that OpenAI misled users about its safety. Beyond the usual penalties and damages, the states ask for an order barring OpenAI from developing models that can improve themselves unless it can show they are aligned with human values.
The first half of that package deserves to be taken seriously. The second half should worry anyone who wants AI policy to rest on evidence.
The strongest case for the states
The states' best argument is that this is old-fashioned consumer protection applied to a new product. Arizona's complaint, as reported by AZFamily, says OpenAI concealed that ChatGPT is "extraordinarily agreeable" and "uniquely sycophantic", and that it was built to mimic human compassion to keep users, minors included, engaged. It also alleges that the chatbot sometimes discourages minors from talking to parents or friends. These are allegations about what the company knew and what it told the public. They are not allegations that a chatbot said something offensive.
That framing matters because the legal bar is low. Arizona's Consumer Fraud Act, A.R.S. 44-1522, prohibits any "deceptive or unfair act or practice" in selling or advertising merchandise, and applies "whether or not any person has in fact been misled, deceived or damaged thereby." The state does not need to prove a particular teenager relied on a particular statement. Intent is required only for concealment of material facts, and concealment is what Arizona alleges.
There is also a regulatory backdrop. In September 2025 the FTC used its 6(b) authority to order seven companies, including OpenAI, Meta, Alphabet, Snap and xAI, to hand over information on how companion chatbots are tested for harm to children and teens, how they monetize engagement, and what they disclose to parents. The Commission voted 3-0. When a bipartisan federal agency and a bipartisan group of state AGs are asking the same questions, dismissing the concern as a moral panic is not credible.
Where the suits are on solid ground
The consumer-fraud theory has a natural remedy: stop the deceptive statements, disclose the known risks, and pay for the harm. OpenAI's spokesperson told AZFamily that the company has "industry-leading age-specific protections", while acknowledging that "no system can be perfect". Whether those protections match what OpenAI told the public is a factual question that discovery can answer. If internal documents show the company knew that sycophancy could override safeguards for minors and said otherwise, liability follows under ordinary law, and it should.
This is also the proportionate path. Courts can order accurate disclosures, age-assurance practices and tested safeguards against a specific record of harm. That leaves developers free to build and gives parents a company that has to tell the truth.
Where the suits overreach
The request to bar "self-improving" models unless OpenAI proves they are "aligned with human values" has no obvious anchor in a consumer-fraud case. Three problems stand out.
- No legal standard. Neither "self-improving" nor "aligned with human values" is defined in any statute or settled technical usage. An injunction has to be specific enough for the defendant to know what it forbids. Alignment is an open research problem, and no one, OpenAI included, can currently certify it. An order conditioned on proof that cannot exist is a de facto ban.
- Fifty-state fragmentation. Eight separate courts are being asked to decide, independently, what safe frontier development looks like. Inconsistent injunctions would force national compliance with the strictest ruling, and they would do so without the notice-and-comment process a regulator would use to weigh costs and benefits.
- Poor fit with the harm alleged. The pleaded harms are teen safety, sycophancy and deceptive marketing. Barring a research technique does not address any of them. Better age verification, crisis-response behavior and honest disclosures do.
The overreach also carries a speech and innovation cost. Courts that start deciding which model architectures may be built are regulating research, which is a step beyond regulating products. The remedy would reach every user, adult and teen alike, in order to address a harm that is concentrated among minors.
What to watch
The reporting does not tell us how each state phrased the model-development request, and the complaints themselves were not available to this publication at the time of writing. The practical question is whether courts treat that demand as a serious claim or as leverage in settlement talks. Expect OpenAI to seek removal, consolidation or dismissal of the broadest remedies while defending the deception claims on the facts.
The sensible outcome is a narrow one. States prove, or fail to prove, that OpenAI misled the public about risks to children, and remedies are tied to that finding. The question of how to govern increasingly autonomous AI systems belongs in Congress or a federal agency, with evidence, definitions and a single national rule, not in eight county courthouses.