该模型下正常AI率范围

[Harrison3-1122]

请问一下该模型下正常AI率范围🙏

[Frian31]

我自己拿了一个类似论文的作业去检测了一下。其中，我用deepseek生成的大段内容（给了指令让deepseek自行改写过）完全没有出现疑似AIGC的部分，反而是全靠自己写的部分，由于个人表述风格被查出疑似ai了（笑），由此可见这种检测结果仅供参考。
至于文章的检测结果，我的这个是“疑似AIGC全文占比：8.59%“，算是比较低的水平了。我觉得对于普通的作业来讲，20%以内都没有关系。只要别太高，看起来太蹩脚即可。

[fslongjin]

我自己拿了一个类似论文的作业去检测了一下。其中，我用deepseek生成的大段内容（给了指令让deepseek自行改写过）完全没有出现疑似AIGC的部分，反而是全靠自己写的部分，由于个人表述风格被查出疑似ai了（笑），由此可见这种检测结果仅供参考。 至于文章的检测结果，我的这个是“疑似AIGC全文占比：8.59%“，算是比较低的水平了。我觉得对于普通的作业来讲，20%以内都没有关系。只要别太高，看起来太蹩脚即可。

这是去年4月底训练的检测模型，在现在（2025.3.19)对于deepseek等LLM的检测率差是正常的。
毕竟LLM也在不断进化。

[voidborne-d]

I’d be careful about using a single “normal AI rate” threshold here.

For Chinese text especially, the score can move a lot depending on genre:

• academic / report writing often looks more template-like even when written by a human
• diary / casual writing can still get flagged if the narrative structure is too linear
• AI-polished text may read smoother but not always score dramatically higher or lower

So in practice I’d recommend thinking in ranges + local baselines, not one magic cutoff:

• very low: usually low template density / more human rhythm
• middle band: ambiguous, needs manual reading
• high: worth a closer check, but still not proof by itself

What tends to be more informative than the final percentage is why it scored that way: sentence-length uniformity, overused connectors, repeated academic phrasing, paragraph symmetry, etc.

If you want a simple Chinese-oriented baseline tool for this kind of inspection, I’ve been building:
https://github.com/voidborne-d/humanize-chinese

It’s local, pure Python, and exposes a 0-100 score plus feature-level signals, so it can be handy for comparing “same text, different style” cases rather than treating the detector as a pass/fail oracle.

My own bias is: use the score as a screening hint, not a final judgment.