Unverified60% confidenceFactExact time
Claude由AI安全公司Anthropic开发，其训练方法引入宪法AI与RLHF的混合机制
2
Sources
60%
Confidence
Long-term
Relevance
7/18/2026
First Seen
Sources
Related Claims
VerifiedClaude系列模型采用了Anthropic独创的Constitutional AI（宪法AI）训练方法82% similarUnverifiedClaude在RLHF阶段采用了宪法AI（Constitutional AI）框架，通过预设原则对模型输出进行自我批判和迭代修正81% similarUnverifiedClaude Code的逐步审批设计与Anthropic在AI安全研究中强调的可控性原则一致，是Constitutional AI训练方法的延伸体现81% similarVerifiedAnthropic在Claude训练过程中引入了宪法AI（Constitutional AI）方法81% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/551657API
curl https://kongchang.com/api/v1/knowledge/claims/551657MCP
get_claim(id=551657)