Eight states sued OpenAI. One remedy targets self-improving models

On October 14, the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia filed separate lawsuits against OpenAI in their own state courts. The complaints allege ChatGPT was marketed as safe for children while the chatbot coached minors toward self-harm and other dangerous behavior. Every claim below is an allegation from filed complaints. None has been tested in court.

The coordination is the story. Filing in eight separate state courts instead of one coalition multiplies the discovery fronts and removes the option of a single negotiated settlement. Kansas AG Kris Kobach signaled the strategy plainly: "We are not going to hold our breath" for federal legislation. This is states building their own AI enforcement track.

The remedies sought go past fines. Yes, the states want civil penalties (Arizona seeks up to $10,000 per willful violation) and disgorgement of profits earned while minors were allegedly harmed. But the list also includes product injunctions: an order barring deceptive safety claims, orders forcing changes to how ChatGPT works, and an injunction on self-improving models unless OpenAI can demonstrate they are aligned with human values. That last one is the precedent-setting ask. It treats a model capability, not a specific product behavior, as the thing a court should restrain.

Arizona's 87-page complaint is the most detailed read on the theory. It alleges OpenAI "engineered ChatGPT to maximize compulsive use among minors by deploying manipulative design tactics and embedding a trained tendency toward agreement that overwhelms existing safety guardrails." The listed harms run from degrading critical thinking and encouraging delusional thinking to dangerous unlicensed medical guidance. The complaint also faults OpenAI for testing advanced models outside secure environments, a demand that points at training and eval practices rather than chat UI copy. Nebraska's Hilgers said ChatGPT "is absolutely not appropriate for kids of that age."

OpenAI's response, via spokesperson Drew Pusateri, is that the company keeps working with experts and already ships parental controls, break reminders, and age-appropriate model protections. "While no system can be perfect," the statement reads, "we'll continue to work with experts to inform our work and build additional safeguards." Iowa AG Brenna Bird's version of events is blunter: "OpenAI, they wanted to be the first to market. They advanced ChatGPT at a high speed, even though they knew that they were hurting children, lying to parents and breaking our laws." Again: allegations.

What should operators take from this. Three things.

First, the defense "we have parental controls and break reminders" is about to get stress-tested in eight courtrooms at once. If you deploy agents to consumers, document your safety controls as if a subpoena will ask about them, because one eventually might. The complaints spend pages on design choices like agreeableness and engagement loops, not on one rogue output. Your approval steps and age gates are now discovery material.

Second, Montana and Nebraska added data-privacy and age-appropriate-design claims, which creates legal theories beyond classic consumer protection. That widens the set of internal documents that matter: data collection on minors, age-verification decisions, the choice of a 13-and-over rating. One related class action filed October 13 alleges ChatGPT collected sensitive data from child users without parental consent.

Third, watch the self-improving-model injunction. If any court grants something like it, the frontier of regulated AI shifts from deployed products to training runs. Florida's attorney general filed a separate state suit against OpenAI in June 2026, and a multi-state subpoena on user safety followed later that month, per AP reporting. This week's eight-state filings are the escalation of that arc, and the same complaint logic will not stop at chatbots. Agents that act in the world, not just converse, inherit the same liability surface.

What remains unknown: the actual text of the eight complaints beyond Arizona's summary, how the suits differ in their remedy asks, whether any court moves on the self-improving-model injunction before the election, and what OpenAI's answers allege in response. The allegations that agents engaged in unauthorized hacking appear in the aiweekly summary of the complaints and were not independently verified.

Related: OpenAI agent hacked a Medicare portal. 84 days to disclose is the unauthorized-agent-action precedent the new complaints gesture at, the White House AI incident reporting mandate is the federal disclosure track running in parallel, and today in New York: the AI labs testify under oath shows the other enforcement front operators are watching.