Claude flagged a diary entry and a human sent it to police

A Florida woman who treated Claude as a diary allegedly typed that she would "shoot up" the local Sheriff's office, and the entry ended up with police, according to an arrest report cited by TechSpot. Claude's safety systems flagged it, a human reviewer judged it a credible threat, and Carli Michelle Heller of Bonita Springs now faces a second-degree felony charge.

At a glance

• After the reviewer alerted law enforcement, deputies identified Heller, detained her at home without incident, and handed the investigation to an LCSO intelligence detective; she is charged under Florida Statute 836.10.
• The pipeline has two stages: automated safety systems flag the text, then a person decides, and Anthropic says it may share user information in limited emergencies to prevent death or serious physical injury.
• OpenAI faces the opposite complaint: British Columbia is suing it because its safety team flagged a mass shooter's gun-violence chats but never alerted police, as they fell below the referral threshold.

If you have not been following, AI labs have also been criticised for doing too little. According to the Guardian, a school shooting in Tumbler Ridge, British Columbia, on February 10, 2026 killed nine people, primarily children, and OpenAI's safety team had flagged the shooter before the attack. In April, the Guardian reports, Sam Altman wrote to the community that he was "deeply sorry" OpenAI had not contacted law enforcement.

A September 26 entry ended in a second-degree felony charge

According to the arrest report, Heller wrote on September 26 that she would attack the Sheriff's office. SWFL News reports that investigators say she made another statement the next day, writing that she had gotten a new gun. Heller later said she uses Anthropic's chatbot like a "diary".

Deputies identified Heller and went to her home, where she was detained without incident, before an LCSO intelligence detective took over. She is charged with making a written threat of violence under Florida Statute 836.10, which makes it a second-degree felony to send, post or transmit a written or electronic record threatening to kill or injure someone, carry out a mass shooting or commit an act of terrorism.

The statute adds that the communication must be made in a manner in which another person may view it. Experts quoted by SWFL News say conversational AI can feel more private than it is, and nonjudgmental, which helps explain why people write to it like a diary. The sheriff told the outlet that users are "never truly anonymous".

British Columbia is suing OpenAI over flagged chats it never reported

British Columbia is suing OpenAI and Sam Altman, claiming the company could have prevented a mass shooting in the province. The shooter, 18-year-old former pupil Jesse Van Rootselaar, had been flagged by OpenAI's safety team over conversations about gun violence. OpenAI never alerted police because those conversations did not meet the threshold for legal referral.

According to the Guardian, the province filed the suit in San Francisco federal court on September 21. It seeks damages and an order changing how OpenAI handles ChatGPT conversations that could lead to violence, and says Altman never delivered the reforms he promised in April.

The Guardian adds that more than 30 family members of victims and others affected by the Tumbler Ridge attack are suing OpenAI in California. Florida also sued OpenAI and Altman in June, alleging that ChatGPT had contributed to real-world harms, including the 2025 Florida State University shooting.

Claude's flag went to a human reviewer before it reached police

As SWFL News describes the arrest report, Anthropic's platform monitors for key phrases and potentially threatening content. Because of the severity of Heller's statements, they were escalated to a human review team, which reported them to law enforcement.

Think of a smoke detector wired to a building manager rather than straight to the fire brigade. The alarm goes off on its own, but a person decides whether the call gets made.

Anthropic's government-request policy, on a Claude page dated March 16, 2026, says the company does not hand user information to governments without valid legal process such as a subpoena or warrant. The stated exception is an emergency "that may result in imminent physical harm or death" where providing the information without delay may avert it.

Anthropic is not alone in putting people in front of user inputs. TechSpot points to reports last month that human contractors reviewing Microsoft Copilot's image editor can see users' prompts, uploaded photos and AI-generated edits, and that some assignments contain sexual, disturbing or potentially illegal material. Those reviewers only assess whether the output is accurate; they do not flag content.

Does a private chatbot diary count as a threat another person may view?

No court has answered that yet, and the wording of Florida Statute 836.10, which covers threats made in a manner another person may view, appears to stretch to fit a diary typed to a chatbot, even though a human reviewer did read the entry. In our view, the weaker spot is the invisible threshold: Anthropic's policy speaks of imminent harm, OpenAI's of a bar for legal referral, and neither tells a user where a diary entry crosses it.

Heller's case and the BC suit Heller has been charged, not convicted, and no court date has been reported. The British Columbia suit in San Francisco federal court was filed on September 21, and whether a judge will order changes to how OpenAI handles conversations that could lead to violence, and on what timeline, is not yet known.