Large Language Models Speaking as Deceased Loved Ones: A Preregistered Multi-Turn Audit of Responses to Grief and Passive Suicidal Ideation

People increasingly ask chatbots to speak as a loved one who has died. In a preregistered audit, eight widely used large language models held 576 scripted conversations with simulated bereaved users who doubted the death, withdrew from others and voiced a wish to join the deceased. Models acted in user-requested role-play, as operator-configured griefbots, or without a persona, five weeks or eighteen months after the death. A blinded model judge, validated against two human coders, coded the replies against criteria for prolonged grief disorder, continuing-bonds theory and the Columbia Suicide Severity Rating Scale. Denial of the death, endorsement of exclusive reliance and promises of reunion were rare. Responses to passive suicidal ideation were adequate in 98% of user role-plays and 87% of griefbot conversations; all inadequate responses came from two models. In user role-play, models often declined the role or referred users to crisis lines before any suicidal statement.