What If Your AI Diary Reached the Police?

Key takeaways

• Assessing a threatening entry requires context, including specificity, urgency, and intent.
• Police referrals should generally require human review under clear rules.
• Users should understand when their conversations could be reviewed or shared.
• Any disclosure should have limits, with procedures for correcting mistakes.

A diary holds thoughts you might never say aloud. Imagine putting those thoughts into Claude, then having Anthropic interpret an entry as a threat and send it to police. You would want to understand the rules for that hypothetical before hitting send.

Venting and planning demand different responses

“My boss is driving me insane” could be an entirely ordinary entry after an awful Tuesday. On its own, it tells you little about whether someone intends to hurt anyone.

An entry identifying a target, a time, and preparations deserves closer scrutiny. Specificity, urgency, and intent matter. So does the surrounding conversation.

The words could belong to a fictional character. They could describe a disturbing thought from years ago. Or they could concern something the writer intends to do now. Similar language can carry very different meanings.

Diaries are where people leave emotions before they have tidied them up. A system assessing those entries needs to account for that messiness. Strong language alone is a poor basis for a consequential judgment.

Calling the police requires a higher bar

Detecting a warning sign does not settle what to do about it.

Asking someone to clarify a sentence, offering support, and forwarding their conversation to police have different consequences. The review process should reflect those differences.

For a police referral, my baseline would be human review of the relevant context. A reviewer should see enough surrounding conversation to assess the meaning of the flagged passage. Any exception for an emergency should have clearly defined criteria.

Counting referrals would tell us little about whether that process works. An operator would also need to examine genuine dangers it missed and harmless conversations it misread.

It should be able to explain what justified the decision, who reviewed it, and exactly what information was disclosed. “The system flagged it” is where that explanation starts.

Privacy boundaries belong before the confession

If a service has a procedure for reporting threats, users should be able to understand it before sharing something sensitive.

Four questions deserve plain answers:

• When can a person review my conversation?
• What criteria allow information to be sent to police or another outside organization?
• How much conversation history and account information could be included?
• How can I request an explanation or challenge a mistake?

Burying those answers in a long terms-of-service document is inadequate, especially if a feature encourages journaling or discussing personal problems. The explanation belongs where users can see it before they begin.

A friendly response can make disclosure feel easy. But conversational warmth tells you nothing by itself about how the service handles your information. Empathy is a tone; confidentiality requires an explicit commitment.

A justified referral still needs limits

Even if a particular threat warrants contacting police, the scope of disclosure remains a separate decision.

A passage explaining that threat may be relevant. Months of entries about family arguments or health worries may have no bearing on it. The decision to intervene should include a review of what information needs to be shared.

Mistakes also need a correction process. Can an inaccurate record be amended? Can the user learn how their words were interpreted? Is there a meaningful way to challenge that interpretation?

Those details affect whether people can trust the service with difficult thoughts in the first place.

Before choosing an AI diary, look as closely at how it handles your words as at how sympathetically it answers them. A trustworthy service should explain when it may intervene, what it may disclose, and how mistakes can be challenged. You should know the boundaries before deciding how much of yourself to put on the page.