Hacker News

studies how generative AI can replicate and amplify real-world biases

Emphasis mine. That's a damn important factor, because the deep "learning" models are prone to make human biases worse.

I'm not sure but I think that this is caused by two things:

1. It'll spam the typical value unless explicitly asked contrariwise, even if the typical value isn't that common.
2. It might take co-dependent variables as if they were orthogonal, for the sake of weighting the output.