Early research covered by the Washington Post finds that conversational AI can implant memories of events that never happened in users who lean on it for recall, a documented harm rather than a hypothetical one, landing the same week AI's self-replication risk gets its own paper trail.