ChatGPT memory study exposes algorithmic profiling and privacy risks
A paper accepted by ACM Web Conference 2026 analyzed exported ChatGPT conversations and stored memories from 80 real users, according to 36氪 and 爱范儿. The reports say 95.9% of 2,050 memory entries were system-written rather than explicitly requested, while some contained personal or sensitive data and psychological inferences. Researchers also developed Attribution Shield, which reportedly reduced personal attribution in rewritten prompts while largely preserving response usefulness.
The study offers an external view of how ChatGPT Memory can form detailed user profiles, making its findings relevant to AI privacy, transparency and user control.
你以为只是和 ChatGPT 随便聊聊,它可能已经在悄悄给你画像。
你的工作、生活习惯,甚至你想要什么、害怕什么,都可能被写进它的记忆。而且很多时候,你并没有要求它记住这些。
前段时间,一篇被顶级会议 ACM Web Conference 2026 接收的论文《The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT》利用 GDPR 赋予用户的数据访问权,从 80 名真实用户那里获得了他们完整的 ChatGPT 对话记录,以及系统为他们存储的全部记忆,包括 1058 段对话、22971 轮交互、2050 条记忆条目。
Some source content is omitted. Continue reading on the original site.
#ChatGPT#OpenAI#Attribution Shield#Qwen2.5-32B#Gemma3-27B#GPT-OSS-20B#GPT-4o