Larger AI models are bigger bullshitters

 So... as AI chatbots have been built upon ever-bigger and better models (with more parameters, more training data and more computing power) they have also been trained to "always answer", even if they have to make an answer up. Sometimes that's good; we need new ideas. Often it's bad.

As someone on twitter put it: "AI needs to learn when to shut up".

My story for Nature: https://www.nature.com/articles/d41586-024-03137-3