According to investigators, Carli Michelle Heller wrote on September 26th that she would "shoot" at the office of Sheriff Li County ( Lee County Sheriff's Office ), and the next day she claimed to have a new gun.
According to the arrest report, the security system of Anthropic flagged these contents, which were then reported to the law enforcement authorities by a manual review team.
The consumer terms of Anthropic reserve the right, at their "own discretion," to report users' inputs, outputs, or behaviors to law enforcement authorities.
This woman from Florida claims to have used AI as a diary. Now, she is facing felony charges, and the “readers” who brought her here are precisely the people she never intended to let see this content.
According to reports from the local WINK News, Carli Michelle Heller from Bonitas Springs (Bonita Springs) was arrested after investigators discovered that she had written on September 26th about her intention to "shoot" the Lee County Sheriff's office. The investigation revealed that by the next day, there were messages on her Claude account indicating that she claimed to have obtained a new gun.
The arrest report states that the platform of Anthropic monitors for keywords and threatening content. Due to the seriousness of these expressions, the relevant content was submitted to a manual review team and subsequently reported to law enforcement authorities.
The officer confirmed the identity of Heller, went to her residence, and detained her without any incidents occurring. She was charged with making violent threats in written form.
The document with Anthropic does not conceal this fact. Its consumer terms reserve the company's right to report users' inputs, outputs, or behaviors to law enforcement authorities at its "discretion." The privacy policy, which came into effect on September 10, 2026, also allows for the sharing of data with the police when the company deems it reasonable and necessary to prevent serious harm in good faith.
In fact, even if you choose not to allow Anthropic to use your data for model training, the company can still retain your data for three months instead of five years. Therefore, your chat content is not truly private.
The same policy also states that conversations marked for security review will still be used to enhance the ability of Anthropic to detect harmful content. The marked diary entries may ultimately become training material for the detector.
Heller is not an isolated incident. In September, the San Francisco Standard Chronicle reported that Anthropic had referred a user to the police because the user allegedly wrote that they had purchased a AR-15 and targeted the Anthropic CEO, Dario Amodei. An employee told the police that he initially did not believe there was an immediate security risk and refused to show those messages on the grounds of company policy.
The user told Standard that he was just joking at the time; by the time the report was published, he had not yet been arrested or prosecuted. A spokesperson for Anthropic described the incident as “our security procedures functioning as expected.” The company refused to state whether they would handle threats against their own employees in a different manner.
This is not just a problem of Anthropic. After a school shooting incident in Tumbler Ridge of British Columbia in February that resulted in 8 deaths, OpenAI stated that they considered issuing a warning to Canadian police in June 2025, but ultimately determined that the account did not indicate an urgent and credible risk of serious bodily harm.
In a policy article published in August 2025, OpenAI stated that threats directed at others would be forwarded to human reviewers, who might then pass them on to law enforcement authorities.
The trial date for Heller has been set for November.
