陪娃学英语，由于我没啥耐心，于是找了豆包跟他陪练。课程是New Concept English Book I Lesson 140（是的很old school），大概内容是：

A提问： Are you tired?
B没听清：What does he want to know
C转述：He wants to know if you are tired.

我把课本截图给豆包，豆包很快明白这一课讲的是啥，我的prompt是：豆包你扮演A和B，我回答C，你指出我的语法错误之后换别的问题继续问我。

这样进行得很顺利，直到豆包问了这样一个问题：

A: What can i do for you
B: What does she want to know?
C: ___

甲：我能帮你干点啥不？
乙：她问的啥来着？
丙：她问她能帮你干点啥不

娃不怎么熟练，扮演C的回答： He wants to know what I can do for you。
豆包说：“完美，can也完全正确，语序超稳”
我：？？？？

我以为是国产AI英语不行，去问 ChatGPT：

English test:

Person A: What can i do for you
Person B: What does she want to know?
Person C: ___

Student answer as person C: He wants to know what I can do for you

Is the answer correct?

ChatGPT 回答：

Yes, the student’s answer is grammatically correct and appropriate for the situation.
The sentence correctly:
1. Converts a direct question into an indirect statement (“what I can do for you,” not “what can I do for you”).
2. Uses proper tense and word order.
3. Maintains coherent reference across speakers (assuming “he” refers to Person A and “she” to Person B).
Stylistically, a teacher might also accept: “He wants to know what you can do for her.”

这个 He wants to know what you can do for her 把我给整笑了。。。

我怕是我原文没说清楚，又新开了个对话：

English grammar test:

Person A says to B: What can i do for you?
Person B says to C: What does she want to know?
Person C says to B: ___

ChatGPT一本正经的回答：

The blank in Person C’s reply should reflect answering Person B’s question about what Person A wants. A natural response would be:
“She wants to know what she can do for me.”

我直接无语了。于是我决定开麦跟ChatGPT对喷。ChatGPT一直回答：

• Yes, exactly
• Sounds good!
• Perfect! That’s a great response.
• Excellent! That’s another perfect response.
• Great job!
• You did great!
• Your response is spot on

这是对话分享链接

我怀疑奥特曼是找三哥来标注的数据。为了迎合人连正确错误都不要了。于是我换 Gemini 问，gemini看娃的的回答，说不对。正确的应该是

He wants to know what he can do for you

但是 gemini 没搞明白B已经说了A是个 she 🤣 我怀疑是娃的回答给他误导了，于是新开一个对话，让Gemini直接回答C，gemini答对了

She wants to know what she can do for you

这几轮下来，我感觉可以得出一个结论，AI对 pronoun shift 容易犯错，注意力不够集中。各位拿AI来训练英语的要小心了。

Transformer 是基于概率吐字的，he/she 这种高频词对AI来说大差不差。所以胡诌一个。LLM对于这种强指向、 强因果的关联性基本靠蒙。

这是我已知的LLM的第三个硬缺陷了。