Florida asks a court to make outside safety checks a condition of OpenAI's next models

Consider a team whose fourth-quarter plan says to move to OpenAI's next frontier model when it ships. The date on that line already depends on a condition OpenAI set for itself. Training of its most capable models will resume "only when we are confident that we have additional safeguards in place," spokesperson Drew Pusateri said ^[9]. Saachi Jain, OpenAI's head of safety systems, told the Wall Street Journal that "when we ship it to users, we have an extremely high bar in terms of safety and alignment" ^[12].

Florida's motion would change who sets that condition. Under the order Uthmeier wants, OpenAI could not improve new models without approved third-party protections in place ^[1]. The restart would no longer be OpenAI's call alone. His brief treats the company's caution as consent. "It is a rare request for an injunction where the Defendants themselves have publicly endorsed it," Uthmeier wrote. "They have asked the government to tie them to the mast" ^[5]. In a video on X, he said: "If Sam Altman meant what he said about slowing down, he can join our ask to the court" ^[4].

OpenAI's answer is that safety "starts with what companies like ours do ourselves," in Pusateri's words, and he said the company is transparent and cooperating with officials in Florida and across the U.S. ^[10][13]. Uthmeier calls its large language models "dangerous" products ^[14]. The June lawsuit is still being fought, and OpenAI maintains its safety protocols are sufficient ^[8].

The easy assumption for a team building on OpenAI is that the legal risk sits in the next model. Only one of the five demands in the motion concerns new model development, though ^[1]. The other four are about the product people use now: "no more harvesting children's data; no more calling this product safe, accurate or reliable; no more pretending it's human; no more tricks designed to keep users talking past the point of danger" ^[3].

Minutes in chat were never a measure of whether a user got an answer or came back the next week. That last demand treats keeping users talking past a certain point as a harm ^[3], so a team that reports session length as success is reporting the thing Florida wants limited.

The reporting does not say whether an order would reach developers who use OpenAI's models through its API. The motion is also only a request so far ^[1].

I'd build the next roadmap on the models OpenAI sells today and count any upgrade as a bonus. The cost is a plan that looks less ambitious in review, and a competitor who bets on a quick restart ships first if the pause ends soon.

To sort a product, use two axes. The first is model dependence: whether a feature works on current models or waits on one OpenAI has paused or held back ^[9][11]. The second is exposure to the product demands: whether minors reach it, or whether its design leans on a human-seeming persona and longer sessions ^[2][3]. Low on both, and the filing changes little. High dependence with low exposure is a schedule problem OpenAI created before Florida filed ^[9]. The reverse, current models with high exposure, is where four of the five demands point ^[1]. A product waiting on a paused model while serving minors sits in both.