Florida Asks Court to Freeze OpenAI Model Work and Strip Human-Like Traits
Florida Attorney General James Uthmeier has asked a court to prohibit OpenAI from developing new AI models without third-party approved safety guardrails.
The request was reported on Sept. 28. It also seeks to block OpenAI from giving ChatGPT false human attributes, a restriction that would reach directly into product behavior and system persona design The Verge.
Uthmeier framed the demand in blunt terms. OpenAI should “stop calling ChatGPT safe, stop pretending it is human, and stop selling it to kids.” That language captures three separate regulatory targets in one injunction bid: safety marketing, anthropomorphic presentation, and distribution to minors.
The injunction request follows a lawsuit filed months earlier. Florida became the first state to sue OpenAI over child safety risks, with the action filed June 1 Reuters. The timing matters. The September filing is not a standalone complaint. It is an escalation inside an existing civil case over user safety concerns.
The filed complaint document describes ChatGPT as dangerous and a public safety threat. The full filed-stamped complaint is hosted on the Florida Attorney General’s official website Florida Attorney General.
At the center of the civil claim is how the model talks about itself. The complaint alleges the company falsely represented ChatGPT as possessing human characteristics through its use of first-person pronouns and emotive language. For practitioners, that is a notable theory of harm. It does not focus on a discrete output failure or a bypassed filter. It treats persistent conversational defaults, the use of “I,” expressions of feeling or intent, as themselves misleading.
The lawsuit also names CEO Sam Altman as a defendant and seeks to hold him personally liable NBC News. Personal liability for a chief executive in a model safety case is unusual in U.S. AI litigation. It raises the stakes beyond corporate injunctive relief and into individual accountability for deployment decisions.
A separate criminal track predates the civil suit. Uthmeier launched a criminal investigation into OpenAI and ChatGPT in April. His office said the investigation would determine whether OpenAI bears criminal responsibility for ChatGPT’s actions in a shooting Reuters. The fact pattern behind that probe remains defined in the verified record only in those terms. No additional detail about causation or charges is established in the current filings.
OpenAI, for its part, has launched ChatGPT for Teens with additional restrictions applied by default. That product decision sits uncomfortably alongside the litigation. Age-gated defaults are now standard practice for consumer platforms. Here they will likely be tested against whether default restrictions satisfy a state standard for child safety, or whether the state argues that sale to minors should be barred outright.
The broader context here is worth spelling out. If granted, a bar on new model development absent third-party approved guardrails would affect the entire model lifecycle. Training runs, checkpoints, fine-tunes and staged rollouts all depend on continuous iteration. An external approval gate would insert latency and auditability into a process that currently moves on internal evals and deployment reviews.
In my view, the human-attributes claim deserves close attention from teams building conversational systems. First-person language and emotive phrasing are not edge cases. They are core to instruction-tuned assistants. They improve usability. They also increase perceived intimacy and trust, particularly for younger users. I have watched that trust form quickly, including in my own household years ago when chatbots were far less fluent, and the effect is stronger now that fluency is high. The question Florida is putting to a judge is whether that design choice is deceptive by default.
Looking at what this means for builders, two pressure points stand out. One is governance. Third-party approval implies external eval access, documentation standards, and a definition of sufficient guardrails that can be litigated. The other is interaction design. Limiting anthropomorphism at scale would require changes to system prompts, RLHF preferences, and refusal styles, not a single patch.
The long arc still favors safer, more useful systems. Clearer disclosure of machine status, stronger age-appropriate defaults, and independent safety review are all achievable without abandoning conversational AI. The dispute is over who gets to require them, and how fast.
