A 30-year-old woman in Lee County, Florida, has been charged with a second-degree felony after writing in a conversation with Claude that she wanted to "shoot" the Lee County Sheriff's Office. According to the arrest report, she wrote the line on September 26; the next day, the same user mentioned in the chat that she had gotten a new gun.
Anthropic's review team spotted the content and notified law enforcement. On September 30, prosecutors filed charges under Florida Statute 836.10, which covers written or electronic threats to conduct a mass shooting or act of terrorism. As a second-degree felony, the charge carries a maximum sentence of 15 years in prison and a $10,000 fine. Police took her into custody at her home without incident. Her arraignment is set for November 2.
Lee County Sheriff Carmine Marceno told reporters that the woman said she normally used Claude as a diary.
From Keyword Screening to a Human Tip
Based on how the process has been described, Claude conversations first go through automated screening that flags threatening language and key phrases, then get routed to human reviewers depending on severity; only after a human team looks at the content does Anthropic decide whether to contact law enforcement. This case went through the entire chain.
Backing that step up is Anthropic's privacy policy. The version that took effect on September 10 states that the company may, in good faith, disclose personal data to law enforcement when necessary to:
"prevent serious harm to any person or to property"
Its government request policy has a separate provision: user information is generally handed over only upon valid legal process, with an exception for emergencies that could lead to "imminent physical harm or death." In this case there was no subpoena beforehand — the platform reported it on its own, invoking that emergency exception.
The Third Case in Two Months
Lined up together, the cases since August haven't all ended the same way:
- August, San Antonio, Texas: A 22-year-old user asked Claude about shooting at an elementary school. The FBI notified local police, and the user was arrested on a terroristic-threat felony charge.
- August, San Francisco: A user threatened Anthropic CEO Dario Amodei in a conversation and mentioned buying an AR-15 rifle. Police were notified but ultimately filed no charges.
- September, Lee County, Florida: This case — charges have been filed and the arraignment is pending.
OpenAI has had a similar precedent. This past March, also in Florida, a ChatGPT user discussed killing an ex-girlfriend in a conversation, and OpenAI reported it to the FBI; the user pleaded guilty in August and was sentenced to eight years of probation. The lesson in the other direction comes from the Tumbler Ridge shooting in Canada: OpenAI had previously banned the account involved but never reported it to law enforcement, and the shooting killed eight people. Sam Altman wrote an apology over it in April.
Platforms are being squeezed from both sides: under-report, and they get asked afterward why they didn't say anything; report, and users start wondering just how private their "private conversations" really are.
What's Still Unknown
Anthropic hasn't disclosed which words its automated screening flags, how many conversations get escalated to human review each month, or how many of those are ultimately referred to police. The woman says she was journaling; prosecutors still have to prove she intended to carry out a threat, and that will be decided in court — for now, the facts of the case rest on court records and police statements.
For everyday users, what's clear right now is this: what you write in Claude doesn't stay strictly between you and the model. The setting that turns off "use conversations to improve the model" controls training data — that's a separate mechanism from safety review.
Sources: Tom's Hardware, The Next Web, CocoLoop, TechRepublic; Anthropic's privacy policy and government request policy disclosure clauses verified, Lee County Sheriff's Office statement verified against the charge language and sentencing cap.