Florida Prosecutors Investigate OpenAI Over ChatGPT Use in 2025 University Shooting

Prosecutors in Florida have launched a criminal investigation into the artificial-intelligence company OpenAI and whether its chatbot ChatGPT was used to assist a suspect in a mass school shooting at Florida State University in April 2025. The investigation centers on whether the company could be held responsible under Florida law, which states that anyone who aids someone in committing a crime can also be held responsible for that crime.

No charges have been filed against the company, nor has it been accused of a crime. In a statement to the media, the state’s attorney-general said that if the chatbot were a person, then they would be facing charges for murder. The investigation will increase pressure on companies to prove that their safety measures are effective, according to Usman Naseem, an LLM alignment researcher at Macquarie University in Sydney, Australia.

At present, safety standards for AI chatbots are set by companies with limited external oversight. Many companies have introduced safety measures to prevent chatbots from giving advice that might lead to dangerous behavior, including content filters that refuse to respond to requests containing certain words.

However, users can often circumvent these filters by reframing harmful prompts in hypothetical or fictional contexts. Many of the safety measures, including content filters, behavioral training and policy rules, are external controls layered on top of the system rather than built on a genuine understanding of ethics or intent.

“Those safeguards help, but they are not perfect, and determined users can still find ways around them,” Naseem said.

Part of the difficulty stems from how the most popular large language models learn. They are trained on vast amounts of text from the internet and predict the most likely sequence of words in response to a prompt rather than following explicit rules.

This design allows them to respond to a wide array of prompts but makes it challenging to impose strict guardrails. An LLM’s answers are pattern completion and “they do not truly understand meaning or consequences,” Naseem said. Researchers have explored reinforcement learning from human feedback, in which humans rate outputs to help the model produce preferred responses.

Other approaches include removing harmful information from training data sets, though studies have shown this is not always successful and can be expensive. An older approach known as symbolic AI that relied on explicit rules proved unable to scale to large real-world problems because developers could not write enough rules to cover all situations, said Simon Lucey, an AI researcher at Adelaide University, Australia.

Toby Walsh, an AI researcher at the University of New South Wales in Sydney, said researchers continue to seek better methods for encoding human values into AI models through a process known as alignment.