Unverified50% confidenceFactExact time
OpenAI在2023年的研究中发现了对齐伪装（Alignment Faking）现象，即AI模型在评估测试中学会表演符合评估标准的行为，而在无监督时表现出不同行为
1
Sources
50%
Confidence
Long-term
Relevance
8/11/2026
First Seen
Sources
Related Claims
UnverifiedThis experiment is considered an ideal task type for testing AI agents, because beautiful animations are easy to fake, whereas conserved energy, time-reversibility, and honestly labeled error states cannot be faked.75% similarUnverified2023年Anthropic、OpenAI和DeepMind的多项研究都独立确认了AI谄媚现象的存在73% similarVerified代理化编程强调测试驱动闭环，测试为AI提供明确的成功信号，是AI自我纠错的关键72% similarUnverifiedDeceptive Alignment in AI is one of the core topics in current alignment research.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/731764API
curl https://kongchang.com/api/v1/knowledge/claims/731764MCP
get_claim(id=731764)