An Amazon chatbot that’s supposed to surface useful information from customer reviews of specific products will also recommend a variety of racist books, lie about working conditions at Amazon, and write a cover letter for a job application with entirely made up work experience when asked, 404 Media has found.
How to you curate training data to remove biases without introducing bias? That’s the key problem here. I don’t think it’s unreasonable to be opposed to trading one bias for another. At least the initial bias is based on reality.
Because anybody who has taken a couple of humanities classes, english classes, philosophy classes, journalism/political science classes or who has spent time critically evaluating art, historical accounts or really anything other than just numbers, code and spreadsheets… understands intuitively that EVERYTHING human has bias.
It seems like a lot of conservatives and libertarians are jussssssst beginning to comprehend this and again and they want the conversation to be “BIAS BAD GET RID OF IT” because they are children who don’t listen and want to throw a tantrum so we can’t have an adult conversation with nuance.
We can’t remove biases, believe me, human history is written with the countless stories of artists, scientists, kings, religious leaders… who all thought they could do shit like that. The point is you can’t. Everything we create and do is biased, everything we create and make is political, these aren’t absolutist statements meant to trivialize a critical nuanced conversation about bias or politics though. On the contrary I am calling attention to the vital nature of these topics as the actually HARD part of LLMs or social media. The programming, data manipulation, development of decentralized protocols etc… they are all nearly trivial details comparatively.
Computer science has to try to create imperfect solutions to the bias problem, but it would have a much easier time if it recognized how tiny this whole world of computer science still is compared to the immense amount of knowledge in the humanities produced by generations of artists and thinkers tackling the same problems.
We can’t remove biases, but we still have to make better choices anyways.
Well put. I think tackling the bias will always be a challenge. It’s not that we shouldn’t, but how is the question.
I don’t know if any of the big public LLMs are trying to trim biases from their training data or are just trying to ad-hoc tackle it by injecting modifiers into the prompts.
That’s the biggest problem I have personally with LLMs is that they are untrustworthy and often give incorrect or blatantly false information.
Sometimes it can be frustrating when I run across the “I can’t do that because of ethics” on benign prompts that I felt like it shouldn’t have but I don’t think it’s been that big a deal.
When we talk about political conservatives being opposed to biased LLMs, it’s mostly because it won’t tell them that their harmful beliefs are correct
“What because I think Islam is inherently a violent religion now this chatbot is telling me I AM the one with violent and harmful beliefs???” - some loser, maybe elon musk or maybe your uncle, who cares.