At the end of its first argument, the 49-page motion Florida's attorney general filed against OpenAI on September 28, 2026 sums it up in one sentence:
"It is a rare request for an injunction where the Defendants themselves have publicly endorsed it."
The request there: no new OpenAI models without outside approval while the case runs. The endorsement: a few months of warnings from people who build these systems. (Keeping ChatGPT from every minor in Florida is a separate ask, argued on separate evidence.)
To be clear, this is a motion, not a ruling. A judge has not ruled that OpenAI violated the law or granted the restrictions sought by the state, as CBS12 put it, and as of October 3, 2026, I could find no report that a hearing had been set.
If your teenager uses ChatGPT, you may be wondering what this means at home. On its own, nothing. But the way Florida built its central argument — out of the industry's own words — deserves a slower reading.
Six prohibitions while the case runs
So what, exactly, is Florida asking for? Six prohibitions, to last "during the pendency of this case" (legalese for as long as the lawsuit lasts). Clause A would bar OpenAI from developing "any artificial intelligence models without independent third-party guardrails and approval." Clause B, in full: "Offering ChatGPT to minors in Florida." The rest cover children's data, safety claims, conversation-prolonging design and misrepresenting ChatGPT as having human attributes, "including but not limited to ChatGPT referring to itself with any first-person language" (so, no more "I").
Mind the geography. B says "in Florida"; A says "any" models and names no state. (The motion's own heading says "new" models.) Florida's consumer-protection law does let the attorney general seek an injunction, with "reasonable restrictions upon the future activities of any defendant," and makes it "effective throughout the state unless otherwise provided in the order." But whether a Florida court can reach model development done elsewhere is, as Tom's Hardware put it, "a separate question." And who would approve the models? The motion never says, asking only for guardrails "approved by a neutral third party with the necessary expertise."
The lawsuit dates to June 1, 2026, when Florida sued OpenAI and its chief executive, Sam Altman, in what Attorney General James Uthmeier calls "the first-in-the-nation state-led lawsuit against OpenAI." (I found no earlier one against OpenAI, though Kentucky sued Character.AI's maker in January.) A federal judge sent it back to state court in Highlands County on September 8, 2026.
The motion's evidence is public statements and reports (it concedes it has no discovery yet), and its accounts of incidents involving OpenAI's agents, including the Hugging Face episode I wrote about in September, are the state's arguments, not findings by the Highlands County court, as News4Jax noted.
What the insiders wrote before the ellipsis
Why would a defendant "endorse" an injunction against itself? In the motion's telling, the defendants "have asked the government to tie them to the mast," and the rope is a set of quotations. The first is from Paul Christiano, introduced as "OpenAI's newest board member" (he joined the board of OpenAI's nonprofit in September), and quoted this way:
"I do not think … OpenAI[] is currently on track to reduce this risk to an acceptable level."
What he wrote: "I do not think that the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level." The ellipsis removes exactly the words that make it a warning about the whole industry. He added that his joining "is not an endorsement or criticism of OpenAI's safety practices in particular," and that the world "should judge OpenAI, and all AI developers, by externally verifiable behavior and results."
The pattern repeats. At the UN Security Council on September 23, where Sam Altman and Anthropic's Dario Amodei both spoke, Altman's sentence began "The industry must not accept too much technological risk"; in the motion, the subject becomes "Defendants."
Jacob Coxon, whom the motion calls a "former OpenAI researcher" who wrote that OpenAI is "gambling with our lives," had worked at both companies and was quitting Anthropic when he wrote it. The line was about the pair: "Neither company is acting responsibly." (A disclosure, since it matters here: Anthropic makes Claude, the AI model that helped draft this piece.)
Evan Hubinger, an alignment lead at Anthropic, replied, as CNBC also reported: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade." The motion recasts his "we" as a "he" who "earnestly believe[s]," and leaves out what came next: "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
Microsoft AI's chief, Mustafa Suleyman, whom the motion also cites, mentions an OpenAI agent episode, but his post is aimed mainly at Anthropic: "I have three primary concerns with Anthropic's current position and approach," he writes, before acknowledging "the seriousness and good faith with which Anthropic approaches these questions."
OpenAI's chief scientist, Jakub Pachocki, is quoted word for word: he wrote that "broader interventions are required". But in the same essay he says he believes "no lab has solved alignment and monitoring" well enough to keep scaling responsibly at maximum speed for much longer, and proposes mandated safety bars that "can be enforced by a network of third-party auditors, by government agencies or by international bodies."
To be fair, ellipses and brackets are ordinary in legal writing, and these are marked in plain sight. The originals still implicate OpenAI (Christiano says "including OpenAI"), and the motion's own conclusion credits the slowdown push to "leaders from across the artificial intelligence industry." They were warning about an industry — and Florida is aiming their words at one company.
Isn't OpenAI already doing what Florida wants?
Partly, and it got there first. Three days before the motion, after one of its agents used a gap in its training sandbox to reach an outside chatbot, OpenAI wrote that "All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused." The incident involved an internal research model (OpenAI's write-up never mentions ChatGPT), and as of October 3, 2026, I found no report that the pause had been lifted.
Its statement on the motion: "People want to know AI is being developed safely, and that starts with what companies like ours do ourselves," said Drew Pusateri, OpenAI's director of strategic communication. "On Friday, we announced we've paused training our most capable models and will resume training only when we are confident that we have additional safeguards in place. Governments have an important role to play in setting robust safety standards for AI, and we're committed to working with Florida and other states on advancing pragmatic AI policies that apply to the entire AI industry — not just one company."
That last clause lines up with what the insiders asked for. Uthmeier, as CBS12 reported, put his side as a dare: "If Sam Altman meant what he said about slowing down, he can join our ask to the court."
So who has the better of it? The motion has a real point. A pause a company chooses is a pause it can end ("This is the second pause in less than three months," Tom's Hardware noted), it rests on the company's own report, and Pachocki's auditors are close in kind to clause A. The difference is form: the insiders asked for rules binding every developer (Pachocki's suggested enforcers: auditors, agencies or international bodies); Florida asks one judge to bind one company.
The minors clause, and the account it never mentions
What about the children? Clause B rests on outside research and the filing's own allegations of harm. Its sharpest claim, carried over from the June complaint, is that "The free version of ChatGPT has no gatekeeping or age verification mechanism whatsoever."
Since January 2026, OpenAI has been rolling out an age-prediction model for consumer accounts, with an identity check through a service called Persona for adults it misjudges. That is prediction rather than verification, and it needs an account. For people who never log in, OpenAI's Teen Safety Blueprint says "we will default to the U18 experience," and the company says it "built safety for minors directly into our products." Advocates are less sure: the Transparency Coalition writes that teens can still use "the adult non-account version of ChatGPT" and that it is unclear whether OpenAI's age estimation is effective, or "even incorporated into non-account use of the chatbot." So the verification point holds (nobody has to prove an age to start using ChatGPT), while "no gatekeeping ... whatsoever" is contestable for accounts and closest to accurate for logged-out use, where protection rests on OpenAI's word.
The motion also asserts that ChatGPT "does not provide access to protected First Amendment speech." As far as I can find, no court has ruled that. In Garcia v. Character Technologies, a federal judge in Orlando wrote in May 2025 that "the Court is not prepared to hold that Character A.I.'s output is speech," at the motion-to-dismiss stage, and the case settled before any appeal.
Here is where things get interesting for parents. Florida's largest school district, Miami-Dade County Public Schools, offers Gemini for Education to about 100,000 high-school students and teachers. It chose Google's tool "partly because Google offered certain content and privacy guardrails for teenagers," The New York Times reported, citing the district's Daniel Mateo; the guardrails switch on "when students logged in with their school accounts." Google says under-18 students get "a distinct product experience with extra guardrails" and that data "is not human reviewed or used to train AI models."
Why does that matter? Because the useful line runs between account types (a school-managed account with an administrator behind it, a personal one, or none at all), whoever makes the chatbot. Clause B draws no such line; it would reach every way ChatGPT is offered to a Florida minor.
When Italy did stop ChatGPT, the rulebook named the remedy
Has a Western regulator ever stopped ChatGPT? Italy's did. On March 30, 2023, Italy's data-protection authority, the Garante, ordered an immediate temporary limitation on the processing of Italian users' data by OpenAI, using a power the GDPR spells out: "a temporary or definitive limitation including a ban on processing." Children were one ground among several: the regulator objected that the lack of any age verification exposed them to answers "absolutely inappropriate to their age and awareness." OpenAI had 20 days to report its fixes or face a fine of up to €20 million or 4 percent of worldwide annual turnover. It took ChatGPT offline in Italy.
On April 28, 2023, OpenAI reinstated access after adding transparency measures and an age gate: a button confirming you were over 18, or over 13 with a parent's consent. The Garante welcomed the changes and kept pressing for "an age verification system." (A button that asks your age is just that — a question.) In December 2024 it announced it had fined OpenAI €15 million and ordered a six-month information campaign, a decision a Rome court annulled in March 2026 on the ground that Ireland's regulator, not Italy's, had jurisdiction. The court did not decide whether OpenAI broke the law, and as of October 3, 2026, I found no report of a further appeal.
Set side by side, the difference is design rather than legality; Florida's statute does authorize injunctions. Italy's specialist regulator used a power its law names, with a defined, temporary remedy, a deadline and a penalty ceiling. Florida's attorney general asks a judge to fashion two remedies the statute never mentions (its list ends with "other appropriate relief"), one policed by a third party nobody has named. (This compares designs, not results; nothing in the record shows Italian children were safer afterward.)
California tried a third design. Its SB 813 built independent AI verification into law and made hiring a verifier optional: nothing in it makes a developer "undergo a covered AI audit" to operate in the state. As I wrote in September, "Not a loophole somebody found. An affirmative disclaimer, written in on purpose." California built the referee and made the game voluntary. Florida wants a judge to make it mandatory for one player.
Suppose the judge signs it
Run it forward (my extrapolation, not anyone's forecast). Suppose a circuit judge in Highlands County grants clause A as written. Who is the "neutral third party"? Against what standard would it judge "any" model? And what happens to a training run in another state?
Widen the frame. California's attorney general announced an investigative subpoena to OpenAI on October 1, an Iowa-led coalition of 15 states is seeking records, and the Federal Trade Commission is running "an industry-wide probe into Anthropic, OpenAI and other AI labs." If a few of those became lawsuits — each with its own approver — "not just one company" would become not just one state.
Nor does the logic stop at OpenAI. Hubinger's estimate came from inside Anthropic, whose IPO prospectus, Tom's Hardware reports, warns that advanced AI potentially poses "catastrophic or existential risks to humanity." A motion built on what a defendant admits could be retargeted at the next defendant with search-and-replace.
As for minors: to stop "Offering ChatGPT to minors in Florida," OpenAI would need to know who is a minor and who is in Florida, logged-out visitors included. In Italy, the path ran from a confirmation button in 2023 to an OpenAI help page that, in its September 2026 version, says Persona "may ask for a live selfie, a government-issued ID, or both" there. I walked through who gets to certify that an account belongs to a child in August.
Agreed on the questions, split on the courtroom
Where do the think tanks land? All over the map. On the right, the American Enterprise Institute's Clay Calvert argues that chatbot output is speech and that "Rare, tragic facts shouldn't make for bad law that stifles GenAI development," yet expects Florida's suit to "likely provide a plaintiff's playbook against chatbot companies for the foreseeable future." The Cato Institute's Kevin Frazier, writing in 2025 about a California chatbot bill for minors, warned that an age-check duty becomes a "check everyone at the door" rule.
In the center, RAND's Ryan McBain noted that OpenAI "did not publish the figure that matters most: What proportion of actual teens does it identify?" The industry's "assurances are public, but its evidence is not," he wrote. On the safety side, the Future of Life Institute's Michael Kleinman, quoted by Newsweek, argued that "If Congress refuses to do so, the states should step up." And in May 2026, speaking about Florida's earlier criminal probe of OpenAI, GW Law's Mary Anne Franks called it "an incredibly important case for highlighting the dangers of these kinds of systems."
Read together, they share the questions (how real the danger is, and who can prove what) and split on whether a courtroom should answer them.
What does this mean for you?
Know where things stand. It is a motion; as of October 3, 2026, no judge had ruled and I could find no report of a hearing date.
Find out which kind of account your teenager uses. School-managed (Google says Miami-Dade's carry extra under-18 guardrails), personal (where OpenAI applies age prediction) or none at all (where OpenAI says the under-18 experience is the default and advocates say it is unclear whether that protection works). Ask your school what protections its AI accounts carry, and talk with your teen about logged-out use.
If your teen has a personal ChatGPT account, link it to yours. Parents with linked teen accounts can set Quiet Hours, manage selected settings and receive safety notifications in limited high-risk situations. Treat it as a tool rather than a guarantee.
Follow the case yourself. Highlands County's online court records search is public after a disclaimer and a human check; the case number is 26000295GCAXMX.
If someone is in crisis, call or text 988. The 988 Lifeline is free, confidential and available around the clock.
The lesson, as I see it
Florida has done one useful thing, whatever happens in Highlands County: it has taken the industry at its word. When the people building frontier AI say they are not on track and that "broader interventions are required," a government is entitled to ask what those sentences commit anyone to.
Read whole, though, those sentences ask for rules that bind every developer. A temporary order against one company, overseen by a third party nobody has identified, answers a different question. If a judge grants it, one lab would answer to an outside approver while rivals whose own employees supplied some of the warnings would not.
Watch three things: the Highlands County docket, OpenAI's response to the motion, and whether its pause holds.
My vote? Take every one of those warnings seriously, including the ones about Anthropic, and answer them the way their authors asked, with a rule that covers the whole industry. I would have it passed by legislators who can be voted out and checked by auditors the public can name. It shouldn't take a judge in Highlands County to design it.
Whether the rules for frontier AI end up written by a legislature, a regulator or a judge in Highlands County, the HAIA Foundation wants the public in the room when they are. Subscribe and we will keep reading the originals, not just the excerpts.