Carli Michelle Heller, a 30-year-old Bonita Springs, Florida resident, reportedly described using Anthropic’s Claude as a “diary,” according to reporting based on the arrest report. According to that same report, her words included a stated intention to “shoot up” the Lee County Sheriff’s Office. Claude’s safety systems flagged the conversation, a human reviewer assessed the threat as credible, and law enforcement showed up at her door.
The individual case will play out in court. The broader question belongs to everyone who has ever typed something raw and personal into a chatbot.
What the Arrest Report Says
The alleged messages named a specific target, referenced a firearm, and triggered an escalation to law enforcement.
According to reporting based on the arrest report, Heller allegedly wrote on September 26 that she intended to “shoot up” the Lee County Sheriff’s Office. A subsequent message allegedly referenced obtaining a new gun. Claude’s automated safety systems reportedly detected the statements and escalated them to a human reviewer, who determined the content represented a credible threat and notified law enforcement. Deputies located Heller and detained her without incident, after which an intelligence detective with the Lee County Sheriff’s Office took over the investigation. A court date was reportedly set for November. Heller has not been convicted, and the charges remain allegations.
Anthropic’s Emergency Exception
Anthropic’s emergency-disclosure policy is narrow, but most users have no idea it exists.
Anthropic’s standard process for government requests requires valid legal process, such as a subpoena or warrant. The company’s published policy carves out an emergency exception for situations where it believes imminent physical harm or death is at risk, permitting disclosure without waiting for that process. The reporting in this case is consistent with that exception: a specific target, a specific alleged threat, a referenced firearm. This is not a policy that routes every confessional conversation to a law enforcement inbox.
The problem is not that the exception exists. The problem is that most users would have no idea it does.
Florida Law Covers Electronic Threats
Typing a threat into a chatbot is not the same as writing it in a locked journal.
Florida Statute 836.10 makes it a second-degree felony to transmit an electronic record that threatens to kill or injure another person, conduct a mass shooting, or commit an act of terrorism, when the record can be viewed by another person. Entering a message into a cloud-based AI service does not grant it the confidentiality of a handwritten note locked in a drawer. Whether the alleged messages satisfy every element of the statute is a matter for the courts to decide.
Intimate by Design, Monitored by Policy
AI companies encourage confessional use, and they owe users a straight answer about what that actually means.
The interface Claude presents is designed to feel like a private conversation, closer in experience to a personal notes app than a public forum. That sense of intimacy may be real. The legal and technical confidentiality a user might assume from it is not. Much like a secretly tracking users scenario, the gap between perceived and actual privacy can carry serious consequences.
Two changes would meaningfully address this gap. First, the emergency-disclosure policy should appear somewhere more prominent than a support documentation page that most users will never find. Second, Anthropic and its peers should publish aggregate data on how frequently emergency referrals occur. That data would give users a factual basis for understanding disclosure risk rather than leaving them to guess.
The next time you treat a chatbot like a private journal, consider this: the interface is personal. The infrastructure is not. Those are different things, and now you know it.
