reply on: On the Inevitability of Left-Leaning Political Bias in Aligned Language Models \ stacker news ~AI

pull down to refresh

75 sats \ 9 replies \ @optimism 13 Oct \ parent \ on: On the Inevitability of Left-Leaning Political Bias in Aligned Language Models AI

I think that at this point, the training data (for the big LLMs) is all-encompassing and what you find in chatbot interaction is the result of training more than base language ¹.

I do think that bias is there because of this, but I agree with the author on the part where the training of an LLM is on-purpose and thus bias is a desired outcome; I do however disagree with then trying to one-dimensionally express the result on a scale of left-to-right. I also think that the most fit-for-purpose LLMs are those that are at least tuned to some level of specialization, and that means that you need bias.

Some of the bias is probably intentional. Grok apparently had to filter out some nazi crap under public pressure. I think that if we'd less ascribe personality to LLMs, and thus also train it less to simulate having a personality, we'd also need less filtering for bias. I guess my ideal LLM is the opposite of the robotic persona that all the Big Tech is now pitching.

I may be mistaken in this because I'm not at a lab training LLMs, but this is how I interpret the post-o1 era we're in now, where reinforcement learning is ultimately what makes it tick in terms of instruction following. ↩

133 sats \ 6 replies \ @Undisciplined 3h

The nature of political polarization is that almost everything becomes politically coded. There may be no good reason for various opinions to be clustered together, but they are.

So, when there's an extreme censorship campaign against one side, as there clearly was, the available training data will be biased towards the side that wasn't censored.

42 sats \ 5 replies \ @optimism 3h

Okay. So let's take DeepSeek. There are some obvious things censored in there. Did this happen during ingestion because data was left out, or post-ingestion through reinforcement learning?

100 sats \ 4 replies \ @Undisciplined 1h

I'm not even talking about the specifics of what was included or excluded for the purpose of training.

We had an intense decade-long period of big tech censorship online. If these models are training on what's available online, that is a very biased dataset and there's no way to include the missing stuff because people began self-censoring to avoid being demonitized.

0 sats \ 3 replies \ @optimism 1h

Meta has been training it on libgen. I really think this is not some magical thing. They teach it to be the way it is.

100 sats \ 2 replies \ @Undisciplined 59m

That doesn't mean much to me.

I'm talking about bias in the information produced for and available to the world. It's not about some specific training set. There's no available unbiased dataset.

42 sats \ 1 reply \ @optimism 15m

What I'm saying is that these bots aren't as much influenced by the dataset that they're ingesting (literally pirated libraries), as by the follow-up training where it is adjusted so that it answers questions correctly.

Some obscure offensive answer popping up is mostly when it doesn't do what it's trained to do, no matter what you find offensive. The general alignment to human speech and thus, most of the bias, is because that is how they tuned it.

0 sats \ 0 replies \ @Undisciplined 10m

Ok, fair enough. I'd be surprised if there weren't substantial viewpoint bias amongst the trainers, too.

Just about everyone with an advanced degree comes from the establishment left.

100 sats \ 1 reply \ @0xbitcoiner OP 13 Oct

Yeah, I agree. Specialized LLMs are obviously gonna have some bias, that’s kind of the whole point, right? But general ones should stay neutral (or 'natural' bias), especially on key topics. I don’t really know how you’re supposed to properly check an LLM for bias, but I get that it’s not easy, and it’s definitely not fair or accurate to just say “it’s biased this way or that way.” It’s way more complicated than that, for sure.

44 sats \ 0 replies \ @optimism 13 Oct

It's very hard and a guessing game, though "unlearning bias" has been done to an extent. See for example https://erichartford.com/uncensored-models

Footnotes