She used Claude as a diary and wrote, “Going to the sheriff's office to shoot.” Anthropic's review system flagged it, and after manual review, law enforcement was notified.

Police in Lee County, Florida, reported that Carly Michelle Heller, a 30-year-old resident of Bonita Springs, was arrested for allegedly threatening violence in writing. Police said Heller wrote a shooting threat targeting the Lee County Sheriff's Office in a conversation with Claude, Anthropic's chatbot. Anthropic's safety system flagged the content, and after human review, law enforcement was notified. Heller told investigators she had been using Claude as a “diary.”


According to the arrest report cited by local media, Heller wrote in the conversation in the early morning of September 26 that she was “going to the sheriff's office right now to shoot.” The next morning, she said she had gotten a new gun that day and wrote, “This is the last chance, I've had enough.”

After receiving the report, police went to her residence on Dean Street in Bonita Springs and detained her without incident. The case is being handled by the intelligence division of the Lee County Sheriff's Office. She currently faces one charge of written threat of violence and has not been convicted.

Florida law stipulates that threatening to kill, cause bodily harm, carry out a mass shooting, or commit an act of terrorism in written or electronic records that can be viewed by others may constitute a second-degree felony. But whether the conversation content meets the statutory requirements and whether prosecutors can prove the elements of the crime will be decided by the judicial process. Local legal analysts believe prosecutors may interpret the text as a direct threat, but the case is not uncontroversial and also involves the boundaries of free speech.

Anthropic's official policy states: If the company believes there is an emergency situation that could result in imminent bodily harm or death, and providing information in a timely manner may prevent danger, it may disclose user information to the government without the usual legal process. Since Claude's launch, Anthropic has used automated detection systems and human review to identify harmful content and enforce its usage policies. In this incident, the process reported by police was exactly that: the system flagged it first, then human review, then law enforcement was contacted.

Users treat the chat window as a private diary, but the platform does not see it that way. What you think is a private whisper may just be a pending alert in Anthropic's review logs. It's not just Claude. Recent reports about Microsoft's Copilot image editor also show that human reviewers may see user prompts, uploaded photos, and generated results.

Similar lawsuits are also increasing. TechSpot noted that OpenAI and its CEO Sam Altman are being sued over a mass shooting in British Columbia, Canada, with plaintiffs arguing the company could have taken measures to prevent the incident. Florida previously sued OpenAI and Altman over harm related to the 2025 Florida State University shooting. These lawsuit allegations do not mean the court has determined liability, but AI companies are indeed caught in the crossfire between user safety, privacy, and prevention of real-world harm.

You see, the uniqueness of this “diary” is that it not only records your thoughts but also proactively contacts the police at critical moments. Before writing, it's safer to assume others might see it.

Original report: [cnBeta](https://www.cnbeta.com.tw/articles/tech/1580684.htm)