The AI trap: when human experts are hard to find, and AI experts are everywhere

I ran across a problem recently I’d call the AI trap. I couldn’t find an expert on an important matter so I used an LLM* to get advice. They I used a frontier model** to critique the first LLM and it found lots of errors and omissions. And the frontier model could also be wrong. Luckily I found a human, who confirmed to me what advice I should follow.

But that’s the thing. People struggle to find experts that will take the time to explain things authoritatively to them. LLMs have all the time in the world but they are not authorities: they are at best a source of information to further your research. But people will stop with them rather to go further on.

So my strategy nowadays is to have a way to test any information I get back. I ask the LLM for sources, I run the output from one model through another, I’ll ask a second model to play devil’s advocate on the first model, and I get runnable code and even tests to verify the runnable code. Always be verifying is my motto with AI now. Never settle for the answer you get back from it. Never fall into an AI trap.

(* GPTOSS120B running on the cloud at ollama.com)

(**Claude)

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.