Expired50% confidenceFactExact time
As companies such as OpenAI and Anthropic deepen their research into model alignment, the new generation of models is beginning to be trained with a 'helpful resistance' capability, pushing back when users reason incorrectly or make harmful requests.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/14/2026
First Seen
Valid until: 10/12/2026(expired)
Sources
AI女友低龄化：12岁少年沉迷聊天机器人背后的情感危机
Related Claims
UnverifiedAnthropic、OpenAI等机构已发表多篇论文研究AI谄媚性现象，发现即使最先进的模型也会在用户施加压力时改变正确答案以迎合用户73% similarUnverified主动要求AI引入可信瑕疵可以对抗模型的讨好性偏差，从而提升输出的真实度71% similarUnverifiedOpen-source model weights allow researchers to analyze model internal behavior using methods such as probing, activation patching, and mechanistic interpretability, analyses that cannot be performed on closed-source API models.71% similarUnverifiedOpenAI proposed the 'AI Safety via Debate' theory in 2018, with the core assumption that adversarial discussion among multiple models can converge toward correct answers.71% similarUnverifiedAI model over-refusal is a known issue in safety alignment research, where RLHF and Constitutional AI training cause models to incorrectly classify harmless requests as requiring refusal68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/507379API
curl https://kongchang.com/api/v1/knowledge/claims/507379MCP
get_claim(id=507379)