A Florida Woman Used Claude as a Diary. An Anthropic Employee Read It and Reported It to Police

What it claims

This is a news report, not a technical release. Its point for engineers is that content typed into Claude's consumer product can be read by Anthropic staff and passed to police. Carli Michelle Heller of Bonita Springs, Florida, used Claude as a diary. On Sept. 26 Heller wrote about planning to "shoot up" the Lee County Sheriff's Office, and the next day wrote about having a new gun. According to the arrest report, as relayed by WINK News, Anthropic's systems flagged the entries and a human review team reported them to law enforcement. Heller was detained without incident, charged with making a written threat of violence, and has a court date in November.

How it works

The arrest report describes the mechanism. It does not come from Anthropic.

flowchart LR
  A[User input] --> B{Automated scan:<br/>key phrases, threats}
  B -- not flagged --> Z[No review]
  B -- flagged, severe --> C[Human review team]
  C --> D[Report to police]
  C --> E[Used to improve<br/>harm detectors]

1. The platform monitors for key phrases and threatening content.
2. Statements judged severe go to a human review team.
3. That team decides whether to refer the content to law enforcement.
4. Per the privacy policy, conversations flagged for safety review are also used to improve Anthropic's detection of harmful content. A flagged entry can therefore become training data for the detectors.

Policy terms the reader should know

• Consumer terms: Anthropic may report a user's inputs, outputs or actions to law enforcement at its "sole discretion."
• Privacy policy (effective Sept. 10, 2026): Anthropic may share data with police when it believes in good faith that disclosure is reasonably necessary to prevent serious harm.
• Retention: the article says that opting out of model training shortens Anthropic's right to use your data from five years to three months, but does not remove it.
• Comparable industry practice:
• OpenAI's August 2025 policy post also routes threats against others to human reviewers, who may refer them to law enforcement.
• After the February school shooting in Tumbler Ridge, B.C., OpenAI said it had considered alerting Canadian police about the account in June 2025. It decided against it because the account did not show an imminent and credible risk.

Limits

• Sourcing: the mechanism comes second-hand from an arrest report via local news. Anthropic is not quoted on the Heller case, and no thresholds, phrase lists or reviewer criteria are given.
• Inconsistency: in an earlier case reported by the San Francisco Standard, Anthropic referred a user who allegedly wrote about buying an AR-15 and targeting CEO Dario Amodei. An employee told officers he wasn't immediately concerned about safety and refused to show them the messages. Anthropic called this "our safeguards process working as intended." It declined to say whether threats against its own staff are handled differently.
• Scope: the article covers consumer terms only. It says nothing about API, enterprise or zero-retention arrangements, so check your own contract before drawing conclusions for a product built on Claude.