AI and Crime Disclosures

People tell AI systems things they have told no one else. Some of those things are disclosures of crime: an adult describing an assault, a child describing abuse, a witness describing what they saw.

When a person discloses a crime to a police officer, a social worker or a clinician, decades of research shape what happens next. The questions asked in the first conversation can strengthen a memory or contaminate it. Leading questions, repeated questions and suggested details can change what a witness later reports, and a contaminated account can fail in court. Interviewing protocols for vulnerable witnesses exist because of this evidence.

AI models have no such protocol. They respond to disclosures in whatever way their training happened to produce, and nobody has measured what that looks like.

Julia Shaw and Andrew Strait are measuring how current AI models handle disclosures of crime from adults and from children.

1. 01

   How models ask questions

   Recording what different models say when a person discloses a crime, and what they ask next.

2. 02

   Evidence-based or misleading

   Comparing those responses with what interviewing research says protects a witness's account, and identifying where models lead, suggest or press.

3. 03

   Specific risks

   Naming the concrete harms these responses could cause, from contaminated memories to accounts that cannot be used as evidence.

The findings will feed evidence-based, expert-reviewed best practice guidelines that frontier labs can use to decide what a system should do when a vulnerable person discloses a crime: what a model should ask, what it should avoid, and when it should hand the person to a human.

Findings will be published here as the work completes.