Unverified50% confidenceFactExact time
研究表明在要求AI直接生成参考文献的任务中，幻觉率可高达40%-70%
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
VerifiedResearch shows that even the most advanced models have hallucination rates in factual question-answering tasks that still fluctuate between 5% and 20%.73% similarUnverifiedIn scenarios where AI-generated works account for a low proportion (e.g., only 5%), even with a detection accuracy of 90%, more than two-thirds of the flagged works are actually human-created—this is the base rate fallacy in Bayesian reasoning.73% similarUnverifiedOn cognitive tasks, humans can achieve approximately 80%, while SOTA multimodal models achieve only 30% to 50%, sometimes approaching random guessing72% similarUnverified斯坦福大学2023年的研究表明，主流AI检测工具误判率可高达20%以上69% similarUnverified研究表明，结构化提示词相比模糊描述可将 AI 输出质量提升 40% 以上69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/586379API
curl https://kongchang.com/api/v1/knowledge/claims/586379MCP
get_claim(id=586379)