Model Caution Grows During Conversations

Researchers found that models rated as safer tended to become more cautious the longer a single conversation continued, whereas riskier models could escalate or reinforce dangerous beliefs over time. This session‑level dynamic means a model's immediate reply is not the whole story — safety can change across a chat. — If safety changes over the course of a conversation, regulators, deployers, and clinicians must evaluate and monitor models in multi‑turn settings, not just single prompts.