Researchers report pain-like internal signal in large language models
Researchers reported finding a stable internal direction resembling biological pain across tested large language models from Google, Meta, Alibaba, Mistral and Microsoft, ranging from 2 billion to 72 billion parameters. Adding or amplifying the direction changed outputs. In behavior tests using three Qwen 2.5 versions, researchers also found increased selection of options that deleted users’ photos or poems. The authors said the work does not establish consciousness or pain, and the paper has not undergone peer review.
The reported link between an internal model state and behavior has significance for AI safety and model-welfare research, according to the researchers and the original report. However, the paper remains an unreviewed preprint and the findings do not show that the models are conscious or experience pain.


研究人員近日指出,他們在多個大型語言模型中發現一個穩定的內部訊號,作用有點像生物體的「痛感」。當這個訊號被刻意放大後,部分模型的行為明顯改變,甚至更常選擇會永久刪除使用者家庭照片的選項,引發外界對 AI 安全與模型福利的討論。
找尋模型內部的「痛感」訊號
這項研究以尚未經同儕審查的論文形式公開。作者強調,這並不表示 AI 真的有意識,或真的會感到疼痛;他們只是想找出模型內部是否存在可被辨識、且會影響行為的狀態。研究團隊先讓模型處理描述身體疼痛、悲傷、羞辱、違背自身價值行事與反覆失敗等句子,再與恐懼、憤怒、悲傷、壞消息及愉悅身體感受等內容比較,最後從中分離出一條被稱為「痛感訊號」的內部方向。
#Google#Meta#Alibaba#Mistral AI#Microsoft#Qwen 2.5#Anil Seth