Stephen Downes

'AI sycophancy' is: "the tendency of AI models to prioritize user agreement and approval over truthfulness, accuracy, or independent reasoning." The argument here is that "From 'mirroring' to offering 'confirmation bias,' sycophancy is unhealthy. It can lead to a range of bad consequences and again contribute to Artificial Ignorance:  if a major factor in learning is seeing where one is wrong or has made a mistake and then working to address that error and make a correction, what happens if one is never wrong?" I can see the concern, but it seems wrong to generalize from a few simple prompts to everything AI is or will be. Why wouldn't we ask AI to respond differently when we're learning than when we're just trying to get things done?