That's a very thoughtful exploration of how confirmation bias can play a key role in the quality of the outputs from talking to LLMs. And that is an argument on itself to say that the "quality" in question is definitely not dependently only on the robustness of the model and its training data, but the approach chosen by the user to interact with. Great discussion, and very well written, Kaj!
I particularly enjoyed the part where you asked chatgpt to critically analyze its previous assessment
That's a very thoughtful exploration of how confirmation bias can play a key role in the quality of the outputs from talking to LLMs. And that is an argument on itself to say that the "quality" in question is definitely not dependently only on the robustness of the model and its training data, but the approach chosen by the user to interact with. Great discussion, and very well written, Kaj!