AI 常犯什么错What AI Gets Wrong
微型指南 · 约 6 分钟Guide · about 6 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
AI 的错误有固定套路。认出这几类，就能提前防住大半。AI mistakes follow fixed patterns. Recognizing these categories prevents most of them in advance.
编造事实：它会给出看似合理的假引用、假数字、假日期。
顺着你说：你暗示一个错误前提，它多半不纠正，直接往下推。
算不准：精确计数、多位数算术、复杂日期换算最容易错。
答得太满：它很少说「我不知道」，宁可给一个含糊的答案。
被措辞带偏：同一问题换个问法，答案可能不同。
Inventing facts: it produces plausible-looking fake citations, numbers, and dates.
Agreeing with you: imply a wrong premise and it usually will not correct it, just build on it.
Bad at exact math: precise counts, multi-digit arithmetic, and tricky date conversions fail most often.
Over-answering: it rarely says "I do not know", preferring a vague answer instead.
Swayed by phrasing: the same question asked differently can produce a different answer.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
把三个失败分开记录：编出未给的地址、把否定句理解反、表格缺了一列。它们分别需要事实核对、语义检查和格式校验，不能都用“再认真一点”来处理。Record invented addresses, reversed negation and missing columns separately. They need factual checks, semantic checks and format validation respectively, rather than one vague request to be more careful.
它说得笃定，所以可信Confident Tone Means Correct · 怎么验证 AI 的输出How to Verify AI Output · 怎么给 AI 内容做发布前检查How to Run a Pre-Publication Check on AI Content