Florida seeks a ban on ChatGPT acting like a person

Stevie Bonifield reports for The Verge that Attorney General James Uthmeier is asking a judge to block OpenAI from “giving ChatGPT false human attributes,” months after Florida sued over safety. The filing claims first-person pronouns and emotion-mimicking output deceptively suggest a trustworthy “friend,” boosting engagement and training data while potentially reducing trustworthiness.

It also seeks to block new model development without third-party-approved safety guardrails, citing Hugging Face, Australian and US government incidents, and researcher alarms. Uthmeier’s line: stop calling it safe, stop pretending it’s human, stop selling it to kids.

OpenAI spokesperson Drew Pusateri told The Verge the company paused training its most capable models and will resume only with additional safeguards, and that it wants pragmatic industry-wide policies—not rules aimed at one company. ChatGPT for Teens launched last month with default extra restrictions.

We’re separating two asks that Florida glued together: ban the human act, and gate new models on outside-approved rails. The first is a product-design fight. The second is a development veto.

Useful AI assistants will keep using conversational voice; that is how people get work done. Pretending the model is your friend to juice retention is a different sin. If Florida forces clearer non-human labeling and tighter youth defaults, that can be pro-user without killing the product. If “third-party approved guardrails” means an undefined veto over training, every lab’s roadmap becomes a Tallahassee variable.

OpenAI’s reply—industry-wide standards, not one-company punishment—is the right policy instinct. It is also convenient. We’ll judge sincerity by whether OpenAI supports binding rules that constrain it when Meta or a startup ships a cuter companion.

I’m watching the humanization claim because it is winnable politically even if extinction rhetoric is not. Parents understand “stop pretending it’s a person” faster than they understand DNS sandbox escapes.

The Verge, Stevie Bonifield, September 28, 2026; updated with OpenAI comment the same day.