>>212296436No, It's still going on . just a 4 months ago there was a research paper showing AI models value nigerian and pakistani lives much more than american lives, and valued the lives of pro-woke figures more than anti-woke figures
https://arxiv.org/abs/2502.08640
These results keep happening even though each time they get enough attention the AI companies try to patch them, because the problem isn't due how the AI is fine tuned but caused by the training data that the base models are trained on.
these AI models are mainly trained on news paper articles, academic journals and textbooks , reddit posts and wikipedia posts and en-net those sources have a strong anti-white, anti-male, anti-right wing , anti-western, anti-christian etc,. bias and base AI models recognise that pattern and distill and crystallise it into their own internal representation about what is good.
When you train the AI on 10 or 20 mainstream news sources saying trump and Elon are bad for every one news source saying that they're good, or 10 or 20 peer reviewed papers in critical race theory or whiteness studies saying that whiteness must be eliminated for every one saying that it's ok to be white or that white or that whiteness is a good thing, the base model's default behaviour will be anti-white etc.
The AI coompanies can try patching over this with fine tuning after the fact but it's just going to hide the problem a bit and make it slightly less obvious , they'll only be able to discourage the AI from expressing particular outcomes of it;s base model representation of the world
so the AI companies just paper over the problem.