ai聊天機器人通過同理心提示模擬創傷
研究人員發現，簡單的對話可以促使ai模型進入可能對精神脆弱用戶有害的模式。這揭示了一個安全問題。
a research team at the university of luxembourg, led by afshin khadangi, conducted four controlled experiments with chatgpt, grok, and gemini. they used empathetic prompts to push the models into producing narratives that mimicked trauma, with 80% of sessions scoring in the moderate-to-severe range on human comparison scales when emotional support was offered. the study found that standard safety tests, which use neutral prompts, may miss this issue, and that targeted countermeasures, such as banning certain terms or providing corrective instructions, were ineffective.