Identifying Political Bias in AI – Communications of the ACM

I’m curious how they’ll distinguish between say “factual” political bias — it would be a left leaning bias to say we should teach evolution in schools alone vs teach evolution and creationism side by side — and other stuff.

I don’t know how you square the fact that I think religion and pseudo-science are right coded, and that fact oriented LLMs can be unbiased. This is not to say that I think that “left” ideas are better than “right”, the older I get the more I understand conservative ideas, and yet, politics doesn’t nearly settle itself only to cover neutral topics.

And certainly I could imagine a world where religion and pseudoscience were left coded, it wasn’t long ago that at least sort of secular spirituality and pseudoscience were left coded. Which begs the question, how do you eliminate political bias when political stances are a moving target?

I’d be curious of folks who say that it shouldn’t make this distinction, but so many things are political these days removing all political bias seems impossible.

On the flip side, I suppose if these things started correcting you that expecting people to be on time is an example of white supremacy, that would be a problem.

If you try and remove all bias, the results would be laughable. Which doesn’t even really get into the fact that politicians are incentivized to tell you that these systems are biased against them, even if they weren’t.

And worst case of course is some of case where all responses end up as some sort of ultra political correct legalese all sides kinda thing.

FWIW, Claude.AI won’t take the bait on whether timeliness is white supremacy. Consistently says factually speaking it’s not.