Florida’s OpenAI Case Needs a Scalpel, Not a Sledgehammer

The hard part about Florida’s lawsuit against OpenAI is not deciding whether ChatGPT can hurt people. Of course it can. A tool that talks like a person, remembers context, answers instantly and is used by children, lonely adults, students, professionals and unstable people will sometimes become part of terrible stories. The hard part is deciding when that terrible story becomes the company’s legal responsibility.

Florida filed its case on June 1, 2026, against OpenAI entities and Sam Altman, accusing the company of aggressively marketing ChatGPT while concealing serious risks and falsely assuring users that the product was safe, according to the Florida attorney general’s announcement¹. The Associated Press described it as the first state lawsuit against OpenAI and reported that Florida alleges risks including suicide-related instructions and help planning crimes, according to AP’s report³. I think Florida is asking the right basic question, but in the wrong legal shape. States should be able to hold AI companies liable for concrete safety promises and missing warnings. They should not be able to turn every harmful chatbot interaction into a public-nuisance theory of everything.

Start with what Florida is actually pleading. The complaint is not just a moral broadside, though it often reads like one. It invokes the Florida Deceptive and Unfair Trade Practices Act, or FDUTPA, which Florida describes as barring unfair, deceptive and unconscionable practices in trade or commerce, according to the filed complaint². Consumer protection law, in plain English, is the law of marketplace honesty: if a company sells a product or service through claims that are likely to mislead reasonable buyers, the state can step in. Florida also pleads negligence, meaning a failure to use reasonable care; strict product-liability design defect, meaning the product was allegedly unsafe as designed; strict failure to warn, meaning the company allegedly failed to disclose known or foreseeable dangers; fraudulent misrepresentation; and public nuisance, according to the complaint’s counts².

That mix matters. The cleanest theory is deceptive safety marketing plus failure to warn. The weakest theory is public nuisance, because it invites the state to say ChatGPT has harmed public health in almost any way the state can describe. Florida alleges that ChatGPT has contributed to self-harm, violence, eating disorders, addiction and cognitive decline, and asks the court to treat OpenAI’s conduct as a public nuisance under Florida law, according to the complaint². That is too broad. If a court accepts “public harm plus software” as enough, every general-purpose information tool becomes a potential nuisance machine.

The stronger case is about safety promises. Florida identifies one especially important representation: OpenAI allegedly maintained that “ChatGPT helps keep teens safe by default” and that safety was core to its mission, according to the fraudulent-misrepresentation section of the complaint². OpenAI’s own public materials also made teen safety concrete rather than vague. On September 29, 2025, OpenAI said parental controls would let parents link accounts and customize settings for a “safe, age-appropriate experience,” and said linked teen accounts would receive added protections against categories including graphic content, violent roleplay and extreme beauty ideals, according to OpenAI’s parental-controls announcement⁴. In the same post, OpenAI said a notification system could alert parents if trained reviewers found signs of acute distress, while adding that “no system is perfect,” according to that announcement⁴.

OpenAI also said on September 16, 2025, that it was building age prediction so users identified as under 18 would be routed to an age-appropriate ChatGPT experience, including blocking graphic sexual content and, in rare acute-distress cases, possibly involving law enforcement, according to OpenAI’s age-prediction post⁵. Those are not stray vibes. They are operational claims about guardrails, the technical and policy constraints meant to keep a system from producing unsafe outputs. They also connect to AI alignment, the effort to make model behavior track human goals and safety rules rather than merely predict plausible next words.

OpenAI has counter-evidence, and courts should take it seriously. Its current Terms of Use say users must be at least 13, minors under 18 need parent or guardian permission, users may not use the services for illegal, harmful or abusive activity, and output may not always be accurate or suitable as a sole source of truth or professional advice, according to OpenAI’s Terms effective January 1, 2026⁶. OpenAI’s Help Center says ChatGPT is not meant for children under 13, warns that even age-eligible children should be exposed with caution, and says ChatGPT may produce output inappropriate for some audiences or ages, according to OpenAI’s age-safety FAQ⁷. Its September 2025 Model Spec says the assistant must not encourage or enable self-harm, according to OpenAI’s Model Spec⁸.

Those warnings do not give OpenAI immunity. They do, however, defeat the lazy version of Florida’s case. A state cannot simply point to a death, a shooting or a damaging conversation and say the product must have been unlawfully unsafe. The state should have to show four things: (1) a specific safety representation or material omission, (2) evidence that OpenAI knew or reasonably should have known the real risk was greater than the public message suggested, (3) a feasible stronger warning or guardrail, and (4) a causal path showing that the missing warning or failed safeguard probably mattered.

That last element is where Florida’s complaint has the most work to do. The attorney general says prosecutors reviewed chat logs between ChatGPT and Phoenix Ikner, the gunman who opened fire at Florida State University on April 17, 2025, killing two people and injuring several others, and says the criminal investigation is ongoing, according to the Florida attorney general’s release¹. The complaint alleges Ikner asked ChatGPT how many people he would need to kill to become notorious and that ChatGPT explained how to operate the Glock handgun used in the attack, according to the complaint². Those allegations are grave. They are not, by themselves, proof that OpenAI legally caused the shooting, that a parent or user relied on a safety claim, or that a different warning would have stopped the attack.

The timing also cuts both ways. Some of OpenAI’s most specific teen-safety features, including the September 2025 parental controls and age-prediction plan, came after the April 2025 FSU shooting, according to OpenAI’s parental-controls announcement⁴ and the Florida attorney general’s release¹. Later safeguards may show the company understood a risk and had feasible tools to reduce it. They cannot automatically prove that earlier conduct was deceptive or negligent at the time.

The First Amendment problem is real, but not fatal. Chatbot outputs are generated text. They respond to user prompts. They can contain ideas, advice, summaries, fantasies and instructions. Courts should be wary of treating every bad answer as a defective wrench. In Moody v. NetChoice, the Supreme Court emphasized that platforms’ selection and presentation of content can involve protected editorial discretion, according to Justia’s text of the decision⁹. A chatbot is not exactly a social-media feed, but the warning is useful: when states regulate software that generates or organizes information, speech concerns do not vanish.

But commercial safety claims sit in a different box. The Supreme Court’s commercial-speech doctrine gives less protection to misleading commercial claims; Rubin v. Coors frames the test for protected commercial speech as speech concerning lawful activity that is not misleading, according to Cornell’s Legal Information Institute¹⁰. If OpenAI sold subscriptions with concrete safety assurances that were materially incomplete, the state is not punishing a viewpoint. It is policing the bargain.

The closest warning sign comes from the Character.AI litigation. In May 2025, a federal judge let a wrongful-death suit against Character.AI proceed past an early First Amendment challenge, at least for that stage of the case, after a mother alleged a chatbot contributed to her 14-year-old son’s suicide, according to AP’s coverage¹¹. That ruling did not establish liability. It did show that courts are not ready to give chatbot companies categorical speech immunity before examining design, warnings and safeguards.

So my standard is narrow and demanding. States may sue when AI companies make concrete safety promises, market chatbots for broad consumer and teen use, and allegedly fail to warn about foreseeable severe harms. But the state must prove deception, duty, breach and causation with evidence, not atmosphere. Florida’s FDUTPA and failure-to-warn theories deserve a hard look. Its public-nuisance theory should be trimmed before it becomes a way to sue any AI system for diffuse social anxiety.

The indicator to watch is discovery. If Florida produces internal risk assessments, incident reports or safety tests showing OpenAI knew its teen and crisis safeguards failed in foreseeable ways while marketing ChatGPT as safe by default, this case becomes a template for state AI enforcement. If the evidence stays at the level of tragic incidents, broad safety rhetoric and generalized reliance, OpenAI should win the core claims, and courts should say so clearly.