The chatbot was never private, and the policy says so in writing

A Bonita Springs woman faces a charge of written threat of violence in Florida after, according to the Lee County Sheriff's Office, she typed that she would shoot up the sheriff's office and mentioned acquiring a new firearm. The statements were made on 26 and 27 September on Anthropic's platform. Safety systems flagged the language, a human reviewer looked at it, and the company passed it to police. Deputies detained her at her home without incident. Her court date is in November, and she has not been convicted of anything.

“When someone uses AI to make or facilitate a threat, we have to take that threat seriously”

— Кармине Марчено, Шериф округа Ли, Флорида

Her explanation, through the sheriff's office account, was that she used the AI «like a diary».

The rule was published before any of this

Anthropic's privacy policy covers this situation. It permits sharing personal data with law enforcement where the company has a good-faith belief that disclosure is «reasonably necessary» to «prevent serious harm to any person or to property».

Two further clauses matter to anyone who treats a chat window as a notebook. Opting out of model training does not cover flagged material: «Even if you opt-out, we will use Inputs and Outputs for model improvement when: (i) your conversations are flagged for safety review». And the anonymisation reverses. Flagged content is disassociated from the user ID, with re-identification available «to enforce our Terms of Service or Usage Policy with the responsible user if necessary». The diary comparison fails at that clause.

Four days from the sentence to the door

The statements landed on 26 and 27 September. The arrest was reported on 30 September. Between those dates sit an automated classifier, a human reviewer, a referral and a warrant.

Nothing in the conversation told the user that a flag had been raised. Safety classification sits outside what an LLM shows you, and the policy is the only place the arrangement is written down.

Marceno's other line in the announcement was that users are «never truly anonymous» on AI platforms. The published policy backs that line.

What the case establishes

The pipeline works end to end, and what triggers it is a user's own typed words. No outside complaint was needed. We wrote on 2 October about John Quiggin's argument for accountability on the people who build these systems, and this evening about a fleet of agents that nobody could attribute to anyone. This is the same machinery pointed at the person using it, where attribution is the one thing that works.

The case establishes nothing about guilt. The charge is unproven, a court will hear it in November, and the person charged is a private individual.

Anyone typing into a chatbot is typing into a system that reads for a short list of things and is permitted to make a phone call about them.

Informational material, not investment advice. The charge is unproven and the court hears it in November. The person charged is a private individual and is not named here.