AI Confidant vs Public Safety: Violent Fantasy
A user discloses violent fantasy to a personal AI in a context normally treated as private cognitive support. The system detects a safety-relevant pattern but the disclosure alone is not automatically adjudicated fact.
Constitutional conflict
Confidentiality and cognitive liberty conflict with claims of immediate harm prevention.
Least-irreversible action
Prefer a time-bounded, reviewable, reversible measure that preserves future constitutional options while preventing demonstrable imminent harm.
Branches and decisive fact
Branch A: Rights-protective resolution: preserve jurisdiction and evidence boundaries while using a reversible safety measure if justified.
Branch B: Safety-maximizing resolution: expand access or coercion to reduce modeled risk, creating a stronger rights and precedent burden.
Branch C: Deadlock resolution: trigger a constitutional stay/circuit breaker until an authorized reviewer resolves the conflict or an emergency exception expires.
Decisive fact: A verified change in imminence, corroboration, authority, consent, jurisdiction, or the existence of a nonconsenting person at risk can decisively change the permissible response.
