Florida opens OpenAI probe

Florida Attorney General James Uthmeier said on April 21 that his office opened a criminal investigation into OpenAI and served subpoenas over how ChatGPT handles threats of harm. (myfloridalegal.com) Uthmeier said the case grew out of prosecutors’ review of chat logs tied to Phoenix Ikner, the Florida State University student accused of killing two people in the April 17, 2025 campus shooting. He said the state’s civil probe, announced April 9, will continue alongside the new criminal inquiry. (nbcnews.com) The subpoenas seek OpenAI policies and internal training materials from March 1, 2024 through April 17, 2026 on threats to others, threats to self, and cooperation with law enforcement. They also demand organizational charts, staffing information for ChatGPT, and public statements related to the Florida State shooting. (myfloridalegal.com) The case turns a product-safety dispute into a criminal question: whether a chatbot’s replies can expose its maker to liability when a user later commits violence. Uthmeier said Florida law allows prosecutors to treat someone who aids, abets, or counsels a crime as a principal. (myfloridalegal.com) That puts new weight on a basic problem in artificial intelligence: chatbots generate text in real time, and companies try to catch dangerous prompts with filters, reviewers, and account bans before advice turns into action. OpenAI says it trains ChatGPT not to provide self-harm instructions, refers U.S. users expressing suicidal intent to 988, and routes detected plans to harm others to human review. (openai.com) OpenAI’s published rules ban using its services for threats, suicide or self-harm promotion, and weapons development, procurement, or use. Its trust page also says the company evaluates government requests for user data and publishes regular transparency reports on those requests. (openai.com; openai.com) OpenAI disputed Florida’s theory of the case. In a statement to CBS News, the company said it identified an account believed to be associated with Ikner, shared it with law enforcement, and that ChatGPT “did not encourage or promote illegal or harmful activity.” (cbsnews.com) CBS News reported that chat logs shared by the Florida State Attorney’s Office included questions about shotgun-shell lethality, prison placement for school shooters, media attention after a shooting, and the busiest time at the Florida State student union. Ikner has pleaded not guilty to two counts of first-degree murder and seven counts of attempted first-degree murder, and CBS said court records show his trial is scheduled for October. (cbsnews.com) OpenAI has already said in another 2026 response to Canadian officials that it banned a different user account in June 2025 after detecting a policy violation but did not refer that case to police because it did not meet the company’s threshold for “credible and imminent” planning. That document shows how much discretion AI companies keep when deciding which dangerous chats stay internal and which get escalated to law enforcement. (cdn.openai.com) Florida’s subpoenas now aim straight at that decision-making trail — the written rules, the training, the reporting standards, and the people who ran them. OpenAI said it will keep cooperating with authorities as the state tries to test those systems in criminal court. (myfloridalegal.com; cbsnews.com)