They're not very accurate, but I think snake oil is a bit too far - they're better than guessing at least for the specific model(s) they're trained on. OpenAI's classifier [0] was at 26% recall, 91% precision when it launched, though I don't know what models created the positives in their test set. (Of course they later withdrew that classifier due to its low accuracy, which I think was the right move. When a company offers both an AI Writer and an AI Writing detector people are going to take its predictions as gospel and _that_ is definitely a problem.)
All that aside, most models have had a fairly distinctive writing style, particularly when fed no or the same system prompt every time. If o1-Pro blends in more with human writing that's certainly... interesting.
All that aside, most models have had a fairly distinctive writing style, particularly when fed no or the same system prompt every time. If o1-Pro blends in more with human writing that's certainly... interesting.
[0] https://openai.com/index/new-ai-classifier-for-indicating-ai...