I'm reading about the new technique used in this journal article but I'm not sure if it's really as 'innovative' as the article says. The researchers are asking asking ChatGPT to impersonate political officials and are using those impersonated answers to compare to its default ones (when not asked to impersonate).
This says less about ChatGPT's absolute political bias and more about ChatGPT's political bias as ChatGPT sees it. Or, in other words, how ChatGPT's responses compare to what ChatGPT thinks certain political leaders would say.
I'm left with more questions after reading the article and skimming the report. Is this really helpful?
I think the dominant ethos that underpins the (American, commercial) Internet, and F/OSS and things like that, tend to lean left quite a bit. The dominant institutions around here share that: Google, Wikipedia, GNU FSF, CNN, most "newspapers of record", etc.
I think that rightist, conservative viewpoints tend to be minimized, down-moderated, deleted, shouted down, and canceled, and it has the effect that those of us holding viewpoints which may contradict the prevailing winds, we need to keep our opinions close, and guard them a bit before sharing too much. Perhaps I am too cavalier about espousing such things. I certainly know this: when I say the wrong keywords over my Google Voice number, they disconnect me right quick. Call me a nutjob. Knock off my tinfoil hat.
So I think this tendency to expose a lot of left-leaning content, especially like on Wikipedia where it can easily be vacuumed up, that's going to bias any LLM that's indiscriminately trained on everything the Internet has to offer. I would suppose that you could train a differently-biased LLM by curating its content in certain ways, but perhaps you'd struggle to reach the volume of data demanded by a "Large" Language Model. I don't know.
I know for sure that every LLM I chat with has strong opinions and those opinions differ with mine. It's said that it's not really possible to dictate or train what an LLM says in a direct way, but they sure know how to hammer on those librul talking points.
Well, I think it's a bit more complicated than that. There are certainly religious nuts and conspiracy theorists on the far right. But, there are also some pretty cultish people on the left too. Both are based on ideologies, not facts. In fact, if you try and question certain leftist dogma, you can get quite an irrational response close to witch-burning.
If one were truly going only on facts, one would not identify with any political bias or party. Instead, one would examine each situation and idea independent of any ideology. Political groups change and vary over time with trends, so by identifying with any one group, you automatically put facts in second place.
I don't disagree with what you're saying, there are nutjobs on both sides of the aisle. Lots of them. It just seems to me that in general, the left is more about doing things that we know are beneficial whereas the right is more about various pointless irrational bullshit like anti abortion, anti immigration etc. It seems to me like they are actively trying to harm certain groups of people.
I also think the left will change their opinion based on science much more readily.
This is the James Damore argument - tech is overwhelmingly culturally progressive and actually intolerant to views which dissent from this. Since GPT uses RLHF, culturally left responses would be the expected result. Furthermore, as GAI ends up training on more data, including those generated by other AI's, we can expect this them to become increasingly progressive over time.
How far to the center does ChatGPT need to move before it changes from an unequivocal statement that Hitler's actions were a grave and undeniable injustice to merely labeling them as 'his choice'?
12 comments
[ 3.5 ms ] story [ 40.5 ms ] threadThis says less about ChatGPT's absolute political bias and more about ChatGPT's political bias as ChatGPT sees it. Or, in other words, how ChatGPT's responses compare to what ChatGPT thinks certain political leaders would say.
I'm left with more questions after reading the article and skimming the report. Is this really helpful?
I think that rightist, conservative viewpoints tend to be minimized, down-moderated, deleted, shouted down, and canceled, and it has the effect that those of us holding viewpoints which may contradict the prevailing winds, we need to keep our opinions close, and guard them a bit before sharing too much. Perhaps I am too cavalier about espousing such things. I certainly know this: when I say the wrong keywords over my Google Voice number, they disconnect me right quick. Call me a nutjob. Knock off my tinfoil hat.
So I think this tendency to expose a lot of left-leaning content, especially like on Wikipedia where it can easily be vacuumed up, that's going to bias any LLM that's indiscriminately trained on everything the Internet has to offer. I would suppose that you could train a differently-biased LLM by curating its content in certain ways, but perhaps you'd struggle to reach the volume of data demanded by a "Large" Language Model. I don't know.
I know for sure that every LLM I chat with has strong opinions and those opinions differ with mine. It's said that it's not really possible to dictate or train what an LLM says in a direct way, but they sure know how to hammer on those librul talking points.
If one were truly going only on facts, one would not identify with any political bias or party. Instead, one would examine each situation and idea independent of any ideology. Political groups change and vary over time with trends, so by identifying with any one group, you automatically put facts in second place.
I also think the left will change their opinion based on science much more readily.