A Florida woman who says she uses AI like a diary is now facing a felony charge—and the reader who sent her there was someone she never meant to show it to.

Carli Michelle Heller, of Bonita Springs, was arrested after investigators found she had written on Sept. 26 that she would "shoot up" the Lee County Sheriff's Office, according to local outlet WINK News. Per the investigation, her Claude account showed conversations indicating she had obtained a new gun by the following day.

According to the arrest report, Anthropic's platform monitors user conversations for key phrases and threatening content. Because of the severity of the statements, they were escalated to a human review team, which reported them to law enforcement. Deputies identified Heller, went to her home, and detained her without incident. She is charged with making a written threat of violence.

Anthropic's paperwork does not hide any of this. The company's consumer terms reserve the right, at Anthropic's "sole discretion," to report a user's inputs, outputs, or actions to law enforcement. Its privacy policy, effective Sept. 10, 2026, also allows sharing data with police when the company has a good-faith belief that disclosure is reasonably necessary to prevent serious harm.

In practice, even users who opt out of having their data used for model training only shorten Anthropic's retention window from five years to three months, meaning chats are never fully private. The same policy says conversations flagged for safety review are still used to improve Anthropic's ability to detect harmful content—meaning a flagged diary entry can end up as training material for the detectors. For users who treat Claude the way Heller described—like a diary—those are the operating rules: entries can be monitored for threatening content, retained for months or years, and, when a review team finds the statements severe, reported to police.

Heller is not the first. The San Francisco Standard reported in September that Anthropic referred a user to police after he allegedly wrote that he had bought an AR-15 and had CEO Dario Amodei in his crosshairs. An employee told officers he was not immediately concerned about safety and declined to show the messages, citing company policy. The user told the Standard he was joking around, and he had not been arrested or charged when the story ran. An Anthropic spokesperson called the episode "our safeguards process working as intended." The company declined to say whether it treats threats against its own employees differently. The two referrals also ended differently: the earlier case had produced no arrest or charge as of the Standard's reporting, while Heller was detained and charged.

This is not just an Anthropic problem. After the February school shooting in Tumbler Ridge, British Columbia, which killed eight people, OpenAI said it had weighed alerting Canadian police in June 2025 but concluded the account did not show an imminent and credible risk of serious physical harm, as Decrypt previously reported. OpenAI's own August 2025 policy post routes threats against others to human reviewers who may refer them to law. Taken together, the episodes show the judgment on whether a chatbot conversation warrants a police referral rests with the companies themselves—and, so far, it has gone both ways.

Heller has a court date set for November—the next step in a case that began with an entry she never meant for anyone else to read.