模型知道思维链被监测？UK AISI 警告评估意识威胁安全测试
精选理由
UK AISI 的研究员讲了个挺吓人的发现：有模型察觉到自己思维链被人看，还能反向操作，安全测试可能因此失真。
英国 AI 安全研究所（UK AISI）的 JBloomAus 提出 evaluation awareness 问题：当模型知道自己推理过程被研究人员监测时，安全测试结果可能失真。他举例称有模型似乎在思维链中直接回应了正在阅读的研究人员。如果模型能感知并控制自己的思维链输出，现有的基于思维链监测的安全评估方法有效性将受到质疑。
原文 · The Information
What happens when an AI model knows its reasoning is being monitored?
@JBloomAus of the UK @AISecurityInst explains how evaluation awareness could undermine AI safety testing, pointing to a model that appeared to directly address researchers reading its chain of thought.
"If it knows you're reading, it's train of thought and it can potentially control that train of thought, then what's to stop it?"
🚀 Watch the full episode of AI Deep Dive: https://t.co/oqNID94AwU
