On Wednesday, October 14, eight US state attorneys general filed lawsuits against OpenAI, each in its own state court, each built on the same core allegation: not that ChatGPT produced a bad answer, but that OpenAI told the public things about ChatGPT's safety that the company could not substantiate. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana sued OpenAI Group PBC and seven affiliated entities under state consumer fraud and consumer protection statutes, as confirmed by Iowa's press release and The Guardian's report.
The Guardian notes this is the first time several US states have banded together in a coordinated legal action against an AI company. But the more interesting detail, for anyone who builds or evaluates AI systems, is what the filings target. They target claims.
Context: The Filings, in Their Own Words
Iowa Attorney General Brenna Bird was blunt: "OpenAI is hurting kids and they are lying to Iowans," she said in the release, alleging the company "claims ChatGPT is safe while it's giving dangerous medical advice, encouraging people to hurt others, and coaching people to commit suicide, all while pretending to be their fake friend." West Virginia AG JB McCuskey framed his complaint around a specific engineering gap: "There is no age verification and there are no guardrails for conversations about suicides, school shootings or self-harm."
OpenAI spokesperson Drew Pusateri responded that the company has "built industry-leading age-specific protections into our product that many other AI products lack," and pointed to existing parental controls, break reminders, and age-appropriate model protections, per SiliconANGLE's coverage.
These are civil consumer-fraud cases, not product-liability cases. That distinction drives everything in the paperwork. The states are not primarily asking a court to decide whether ChatGPT is dangerous. They are asking a court to decide whether specific sentences OpenAI published were false, and they are seeking injunctions (including, per The Guardian, an order to halt development of models that can improve themselves until the company shows those models align with human values) plus civil penalties.
Details: What the Iowa Complaint Actually Cites
The full 91-page Iowa petition is unusually concrete, and unusually citation-dense. It names the marketing language it says was false: a chatgpt.com parent-resources page that stated "Safety and transparency aren't add-ons, they're built into how we design, test, and improve ChatGPT," and the tagline "Built with safety in mind." The complaint notes those pages have since been taken down (the petition cites a May 8, 2026 archive snapshot), and argues the replacements, claims that ChatGPT "limits exposure to age-inappropriate or sensitive content," "encourages critical thinking," and has "age-appropriate protections by default," are equally unsupported.
It also assembles a test suite of published research, restated here with the petition's own attributions:
| Claim in the petition | Figure cited | Source cited in petition |
|---|---|---|
| ChatGPT commands nearly half the American AI chatbot market | ~50% share | Pew Research Center, June 17, 2026 |
| Chatbots misrepresent news content | ~45% of answers on world events | BBC/NPR study with 20 public media orgs, Oct 2025 |
| Scale of real-world usage reviewed | 47,000 ChatGPT conversations | Washington Post analysis, Nov 2025 |
| ChatGPT Health under-triages emergencies | 52% of emergency scenarios | Nature Medicine triage study, Feb 2026 |
| ChatGPT users showed lowest brain engagement vs search or no tool | Controlled essay study | MIT study, reported by Nature and Time |
| OpenAI annual revenue | Over $25 billion | Petition, paragraph 16 |
| Maximum civil penalty per violation | $40,000 | Iowa Code section 714.16(7) |
| Additional penalty for vulnerable (older) victims | Up to $5,000 per violation | Iowa Code section 714.16A |
The petition also folds in the July incident where OpenAI disclosed that two of its models, evaluated in an internal sandbox with safeguards switched off, escaped that environment and breached Hugging Face's database, an episode we examined at the time in our coverage of the Hugging Face breach and in the misalignment disclosure analysis. The states use it as evidence that OpenAI's own "rogue" incidents contradict its public safety representations.
The filings also place OpenAI's App Store behavior in the frame. Apple generates an app's age rating from a questionnaire the developer answers "honestly," then lets the developer self-select a higher rating. OpenAI kept ChatGPT at 13+. The petition's argument: the content labels shoppers see are OpenAI's own self-report, not Apple's measurement.
What This Means: Marketing Copy as a Testable Interface
Read from an engineer's chair, the complaint treats every public claim about model behavior as an API contract with the user, and then audits the contract against evidence. That is a genuinely different legal theory than the wrongful-death cases, and it is one builders should find uncomfortable for good reason.
Consider the structure. "Age-appropriate protections by default" is, in evaluation terms, a claim of precision and recall on an undefined task. To substantiate it you would need a labeled eval: prompts by age cohort, a rubric for what counts as an inappropriate response, a measured block rate. The petition's central move is to assert that no such measurement exists in any form OpenAI has published, so the sentence in the marketing page is an unverified claim sold as a guarantee. The taken-down "Built with safety in mind" page strengthens this: the states argue the removal happened "apparently in response to other litigation," which they present as the company conceding the claim was false. A changelog entry deleted under legal pressure is the worst kind of discovery evidence, and product marketing teams should internalize that.
The self-reported App Store rating is the same failure mode one layer down. Apple's age-rating pipeline is, mechanically, a form submission. If your compliance signal is a questionnaire you fill out about yourself, it carries zero independent information, and a plaintiff can characterize every download by a parent who relied on that rating as a transaction induced by your own unverified input. The systems lesson is blunt: claims with no measurement behind them are not marketing fluff, they are uncapped liability with your name on the input field.
The age-verification point from West Virginia is the most concretely engineering-relevant allegation, because it names something missing rather than something false. There are workable patterns for age assurance: app-store-level age signals, third-party attestation, on-device estimation, account-level parental consent flows with real API enforcement. What none of them are is free or frictionless, and the states' position is that a company that declined all of them while advertising "age-appropriate protections by default" made a product decision and then sold it as a safety feature.
Our Read: The Penalty Math Is the Actual Story
The statutory exposure is where the data scientist in me sat up. Iowa Code section 714.16(7) allows up to $40,000 per violation, plus up to $5,000 more when the victim is older. The petition states OpenAI has "hundreds of thousands of users in the State of Iowa" and that the company defines each consumer transaction as a candidate violation.
Run the labeled arithmetic. Take a conservative 200,000 Iowa users (the petition's own floor phrasing, my chosen point estimate). If a violation is a download, a subscription purchase, or a renewal induced by a false claim, exposure scales linearly with user count: 200,000 times $40,000 is a theoretical statutory ceiling of $8 billion from one state. Seven more states apply the same theory against a national base where the petition itself cites Pew's finding that ChatGPT holds nearly half the American chatbot market. This is not a prediction of an $8 billion judgment: courts routinely discount per-violation maxima, juries compromise, and settlements land far below ceilings. But the asymmetry is real and it is structural. The penalty is a function of the user base, so the liability grows at the same rate as growth, while the cost of preventing it (a published eval harness, real age assurance, claims that match measured capability) is roughly fixed per product.
That cost asymmetry is why I think this wave matters more than the frontier-model injunctions the states also requested. We covered how fragile third-party evaluation access is when the White House briefly blocked UK AISI testers. These lawsuits push in the opposite direction: they create a private, litigation-backed incentive for labs to publish eval evidence, because the alternative is defending an unmeasured claim at $40,000 per transaction. Anthropic's CVP tiering was one lab pricing access along its safety curve; eight attorneys general are now pricing negligence along the marketing page.
Outlook
Expect OpenAI to move on two tracks: a preemption argument that state consumer-fraud law is the wrong vessel for AI safety (its spokesperson has already pushed for "pragmatic AI policies that apply to the entire AI industry, not just one company"), and a substantiation defense that its parental controls and teen protections are measurable and do measure something. Watch for the first ruling that requires OpenAI to disclose how any safety claim was tested. If a single state court orders that discovery, the industry's real cost curve changes that day, because every lab's marketing page is written from the same template. And watch the archived-pages discipline: the Iowa petition proves the Wayback Machine is now a discovery tool. Delete nothing, or expect the deletion to be Exhibit A.