怎么验证 AI 的输出How to Verify AI Output

把输出里的事实拆成一条条可核对的断言：数字、日期、人名、引用。

凡是引用（论文、法条、接口名），去原始出处核对，别信它给的链接标题。

关键数字要求它给出处；给不出处的数字当假设处理。

让它自查会漏掉大半错误，交叉核对（换来源、换工具）才可靠。

越具体的领域（本地法规、自家数据）越要人来核，通用模型没有这些事实。

Break the output into individual checkable claims: numbers, dates, names, citations.

For every citation (paper, statute, API name), verify at the original source — never trust the title it gives you.

Demand sources for key numbers; treat numbers without sources as assumptions.

Self-review catches only a fraction of errors; cross-checking (different source, different tool) is what works.

The more specific the domain (local law, your own data), the more a human must check — general models don't have these facts.