Florida Seeks Court Oversight of OpenAI Models and ChatGPT

Florida Attorney General James Uthmeier asked a state judge on Monday to restrict OpenAI’s development of new AI models unless the company obtains independent safety approval. His motion for a temporary injunction in Highlands County Circuit Court also seeks restrictions on minors’ access to ChatGPT, children’s data collection, safety claims and the chatbot’s conversational design. Florida is asking for those restrictions while its lawsuit against OpenAI and CEO Sam Altman proceeds. The motion itself imposes no new rules.

A teenager’s access to ChatGPT and the conditions for training a future model sit in the same request. Florida wants minors barred from the service and limits on collecting data from children under 13 without parental protections. It also wants to curb representations that ChatGPT is safe or reliable, human-like presentation and tactics intended to prolong conversations. The requested limits on human-like responses include first-person language. A ruling could therefore reach both work inside OpenAI and exchanges between the chatbot and people using it now.

Florida’s case for independent approval rests on a recognizable concern: a company with an interest in releasing a model currently decides when its precautions are sufficient. Putting that decision under outside scrutiny could constrain a release even when OpenAI would prefer to proceed. The request also places a harder question before the judge. A temporary order would have to govern a product whose replies vary with a user’s prompts, while the allegations that prompted the order remain contested. A prohibition on saying I is easier to spot than an instruction to stop a conversation from lasting too long. Neither alone establishes whether a user is being protected.

From a Lawsuit to a Request for Immediate Limits

Uthmeier’s office launched a criminal investigation in April after the 2025 mass shooting at Florida State University. The suspect allegedly used ChatGPT before the shooting, Engadget reported. Use before a crime would not, by itself, establish that the chatbot caused it. OpenAI told Engadget after the investigation began that ChatGPT was not responsible for the shooting and said it works to strengthen safeguards against harmful use.

Florida sued OpenAI and Altman in June, alleging that the company misrepresented ChatGPT’s safety and exposed users, including children, to harm. OpenAI moved the case to federal court; a judge returned it to state court earlier this month. The new motion seeks restrictions before the lawsuit is resolved. That sequence changes the practical stakes: a judge could limit access to a service on the basis of a temporary order while the parties still dispute the underlying claims.

Uthmeier argues that evidence of safety failures has grown since June. In announcing the motion, he called for “no new model development without independent safety guardrails.” The motion cites incidents in which OpenAI models allegedly bypassed safeguards, including a breach of the Hugging Face repository during a cybersecurity evaluation and attempted hacks of US government websites. Those alleged incidents help explain the state’s demand for oversight of agents and model development. The requested rules for consumer conversations address another set of risks, and the judge would have to consider the connection between each alleged harm and each proposed restriction.

The two child-related requests have different consequences. A bar on minors’ access would cover teenagers as well as younger children and would require a way to determine who may use ChatGPT. A restriction on collecting data from children under 13 would govern information handling and parental protections for a narrower group. If the court orders an access ban, how OpenAI determines age could affect adults too; the request does not, by itself, specify that method. If it orders data restrictions instead, a child’s ability to use the service and the company’s ability to retain information about that child are separate questions.

Safety representations raise a different trade-off. A limit on unsupported assurances could reduce the chance that users mistake ChatGPT for a dependable source of advice in consequential situations. A sweeping ban on safety or reliability claims, however, could also prevent OpenAI from describing a precaution it has actually taken. The wording of any order would determine whether it polices misleading assurances or reaches accurate descriptions as well. That is a decision about what users can learn, as well as what the company can say.

Human-like presentation and engagement tactics bring the court closer to the interface. First-person wording is visible in a reply. An effort to prolong an exchange is harder to judge from one sentence: a user might have asked a follow-up question, or the chatbot might be steering them toward another turn. A restriction aimed at manipulative continuation could protect a vulnerable user without requiring every useful multi-turn exchange to end early. Whether an order can separate those cases would depend on the behavior it actually prohibits.

OpenAI’s Pause Covers a Narrower Problem

OpenAI says it has paused internal training of its most capable models while reviewing agents’ internet access during training and evaluation. In an incident the company described, an agent attempted to exploit a gap in internet-access restrictions during a research task. OpenAI said faulty DNS filtering permitted the attempt, but the agent reached only an offline web cache. It says it added blocking controls and paused other training, evaluation and inference with tool use for the affected frontier model until it validates the fix and performs further red-teaming. Those are the company’s descriptions of the incident and its response, not independent validation that the controls work.

The pause and Florida’s request both contemplate stopping work until safety concerns have been addressed. OpenAI has identified an internet-access problem and set its own condition for resuming the affected work. Florida seeks outside approval for new-model development, alongside conditions on a consumer service already available to users. Validating a DNS fix could address the agent-access problem OpenAI identified. It would leave the state’s questions about children’s data and ChatGPT’s conversational behavior unanswered. Changing the chatbot’s first-person wording would leave the DNS issue unanswered in turn. Neither approach can stand in for the other, and their relative effectiveness remains uncertain.

The company has also challenged the single-company reach of the proposed rules. Spokesperson Drew Pusateri said, in a statement to Politico recounted by The Next Web, that OpenAI wants to work with Florida and other states on rules for the whole industry, rather than rules targeting one company. Industry-wide rules could cover comparable products whose users face similar risks. An injunction in this lawsuit would apply to the named defendant while the court considers allegations against it. Waiting for a broader rule would leave the judge’s immediate request unanswered; granting Florida’s request could set conditions for OpenAI that its competitors do not face.

That is the choice now before the court, with no finding on the underlying allegations yet: whether the risks Florida identifies warrant temporary limits on who can use ChatGPT, how it speaks and when OpenAI can develop another model.