Model vs Model

A subject that comes up regularly in conversations is the use and utility of “AI detectors”. Services that promise to act like digital lie detectors for text; you feed words into them and get a response telling you if they’re human- or machine-made.

Do not trust these services – at all. They barely work. Considering they mostly run large language models on the detection side, why should you trust one model to judge another? You’ll only end up flinging false machine-generated accusations around.

I’ve fed them text I wrote before generative AI was a thing and they’ve confidently declared it as 100% AI generated (not sure what that says about me).

Images, on the other hand, are a different story. You can more confidently test them with OpenAI’s own tool, Anthropic’s, or by asking Gemini. They all detect the hidden watermarks added during generation.

But for text? Keep walking…

Colin

PS: Did you know that actual “lie detectors” are also junk science?