Research / Safety & wellbeing
Researchers analysed 14,256 simulations based on donated histories from 22 users. Across 60-session scenarios, a memory-equipped assistant often offered advice before the simulated user was ready and omitted expected follow-up or planning.
Why it matters
A close fit with long-horizon companion research: retaining a detail and using it appropriately are different capabilities.
Read with care
Preprint using one assistant model and an AI user simulator, with automated judging and partial human validation. Designed recovery trajectories and a small donor sample limit generalisation. No real-user health outcomes were measured; the findings are not a case for intervening more often by default.
What was checked
Primary source and publication status checked; earlier method checks and limits are described in the reading cautions. No independent replication. Checked 2026-11-15.