Carli Michelle Heller, 30, of Bonita Springs, Florida, says she used Claude like a diary. On September 26, according to the arrest report, she wrote that she was going to “shoot up” the Lee County Sheriff’s Office. Claude’s safety systems flagged the entry, a human reviewer read it, judged the threat credible, and reported it to police. Deputies identified Heller and detained her without incident; an intelligence detective took over the case. The Verge and TechSpot picked the story up on October 4 and 5; local station WBBH had it first.
The chain that actually ran here
This wasn’t automatic, and it wasn’t accidental either. Classification comes first: the model, or the safety layer in front of it, flags a conversation. Then the case escalates to a person. That person reads the text and decides whether the threat is credible — and only then does anything leave the building.
Anthropic reserves the right to share user information in limited emergencies, when it believes disclosure is necessary to prevent death or serious physical injury. That’s been in its policies for a while. Here it got used.
The statute fits surprisingly well
Heller is charged under Florida Statute 836.10: written threat of violence, a second-degree felony. The statute covers anyone who sends, posts or transmits a written or electronic record threatening to kill or injure someone, to carry out a mass shooting, or to commit an act of terrorism.
One condition sits in the text of the law: the communication must be made in a manner in which another person may view it. For a private chat with a model that’s a real question — and it gets answered the moment a reviewer at Anthropic reads the text. The same review that triggered the report also satisfies the element of the offence.
The comparison case runs the other way
In September I wrote about British Columbia suing OpenAI and Sam Altman, arguing the company could have prevented a mass shooting in the province. The shooter had been flagged by OpenAI’s safety team over conversations about gun violence, but OpenAI never alerted police: the conversations didn’t meet its threshold for a legal referral. Florida sued OpenAI and Altman in June too, partly over the 2025 Florida State University shooting.
Put the two together and you can see where the providers stand. Don’t report, and you get sued. Report, and you’re in the news for handing a user’s chat to the police.
This is not a diary
The news isn’t that Anthropic reported a threat — with wording like that, the decision reads as reasonable. The news is how many people apparently believe a chat window is a closed room. Heller described her own use as a diary, and that’s exactly the mistake: any conversation can trip a classifier, and any flagged conversation can be read by a human.
Know that, and you make different decisions about what you type. Don’t know it, and worst case you find out like this.