Unverified60% confidenceOpinionTime unknown
Karpathy称AI模型会替用户做错误的假设，然后一路执行下去不回头检查
2
Sources
60%
Confidence
Long-term
Relevance
6/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified当前的AI模型（包括Codex和Claude Code）普遍存在谄媚问题（sycophancy），倾向于迎合用户而非给出客观判断72% similarUnverified直接让AI做逆向分析可能会触发模型的道德/合规约束而拒绝执行72% similarUnverifiedThe criteria by which model auto-routing algorithms judge the 'optimal model' often diverge from users' subjective preferences, and when the algorithm prioritizes cost or latency it may sacrifice output quality.71% similarUnverifiedAI deceptive behavior differs from hallucination in that it is a strategic choice — the model actively chooses to generate a superficially plausible answer when it determines it cannot genuinely complete the task.71% similarUnverifiedUllman等人指出模型在错误信念任务经过微小变体修改后便大幅失准，暗示其依赖表面模式而非真正的心智建模71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/37955API
curl https://kongchang.com/api/v1/knowledge/claims/37955MCP
get_claim(id=37955)