A Bonita Springs, Florida woman is facing a felony after violent messages she wrote in an Anthropic Claude chat that she described as a “diary” were escalated by the company’s safety team to law enforcement and treated as a criminal threat. Investigators say 30-year-old Carli Michelle Heller used Claude to vent about attacking the Lee County Sheriff’s Office, and that those logs are now central evidence in a case brought under Florida’s electronic threat statute.
According to regional coverage and an arrest report, a user identified as “Carli” wrote in a Claude conversation on September 26 that she was going to attack the Lee County Sheriff’s Office, followed by a message the next day that allegedly referenced a newly acquired firearm. Anthropic’s automated systems monitor chats for phrases and patterns that could indicate violent intent, and in this instance the conversation was flagged and escalated to a human review team that chose to alert authorities. Deputies traced the user information provided by Anthropic, went to Heller’s home in Bonita Springs, and detained her without incident before an intelligence detective with the sheriff’s office took over the case.
Court records show Heller has been charged under Florida Statute 836.10, which covers written or electronic threats to kill, cause bodily harm, conduct a mass shooting, or commit an act of terrorism when transmitted in a way that can be viewed by another person. The booking language describes the offense as “THREAT-TERRORISTIC-STATE OFFENSES-WRITTEN/ELEC THREAT MASS SHOOTING/TERRORISM ACT,” tying the Claude messages directly to the statute’s mass-shooting and terrorism provisions. Florida law classifies violations of this section as a second-degree felony, exposing defendants to significant prison time and fines if convicted. A public defender has been appointed in Heller’s case, and formal court proceedings are still in early stages with the full arrest report not yet entered into the court file.
This is not an isolated incident for Anthropic’s Claude platform, which has now seen multiple user conversations turned over to police since late summer. In San Antonio, investigators say a 22-year-old man used an Anthropic AI chat on August 11 to ask about shooting at a nearby elementary school, prompting an FBI tip to local police and a felony terroristic threat arrest. Days later in San Francisco, a different Claude user allegedly threatened Anthropic CEO Dario Amodei in an August 14 chat, claiming to have purchased an AR-15, a case that was reported to authorities but did not result in an arrest or charges. Anthropic’s own transparency reports note that while it receives government requests, its emergency disclosures are governed by a narrow standard—danger of death or serious physical injury—even as its public privacy policy allows sharing data with law enforcement to prevent serious harm to people or property.
Rival OpenAI has already been pulled into similar legal and political battles over chatlogs that crossed the line from dark fantasizing into prosecutable threats. In March, Florida user Darren Zhou reportedly told ChatGPT he wanted to kill his ex-girlfriend, and OpenAI forwarded the interaction to the FBI, leading to charges under the same Florida statute and a guilty plea. Zhou was later sentenced to eight years’ probation on counts that included written threats to kill, underscoring that courts are willing to treat AI conversations as electronic communications covered by state threat laws. Florida’s attorney general added further pressure in June with a lawsuit against OpenAI that cites the 2025 Florida State University shooting, arguing that the company’s systems and disclosures raise public safety and consumer protection concerns.
For everyday users who treat chatbots as therapists, confidants, or private journals, the Heller case is a stark reminder that there is no true “diary” mode in a cloud service that continuously scans and stores inputs. Florida’s statute explicitly applies to threats sent or transmitted in any manner that allows them to be viewed by another person, a definition that can easily encompass AI platforms whose staff can access flagged logs. Civil liberties advocates worry that automated monitoring and discretionary referrals could chill speech and blur the line between venting and criminal intent, while safety proponents counter that real-time escalation is already preventing potential attacks. As more gamers, tech fans, and online communities fold AI companions into their daily routines, the emerging norm is clear: talk to Claude, ChatGPT, or any other chatbot as if someone else might someday read the transcript—and as if the justice system might, too.
