Mhm, but with the way LLMs work, it’s not possible to actually remove bias since it’s baked into the training data. Any adjustment towards “neutral” would be biased by what the adjuster considers neutral.
That’s how we make them work today. It is possible to stay politically neutral in a language though. And therefore your generalized statement is incorrect.
It will never be objective if its dataset is something like the internet. It will always be prone to bias because that’s the double-edged sword of LLMs, they have to have vast quantities of data and the only place they can get that is the internet which is biased opinions everywhere.
Weekly reminder that scholars since at least 500 B. C thought the earth was a sphere, so Galileo’s LLM would tell him that, and even tell its size with a reasonable accuracy.
Far right media considers “Fact Based” to be a “Liberal Bias”.
It actually is biased though. UpperEchelon did a video exposing this. I swing to the left myself, but I would prefer if the LLMs were objective.
Political objectivity is impossible.
I would argue that asking a machine to list known information is not impossible.
Here’s a very clear example where chatGPT refused to answer a question regarding Biden but happily answered the exact same question for Trump.
https://youtu.be/_Klkr6PtYzI?t=520
And before anyone starts, NO! I’m not a supporter of the oompaloompa king.
Mhm, but with the way LLMs work, it’s not possible to actually remove bias since it’s baked into the training data. Any adjustment towards “neutral” would be biased by what the adjuster considers neutral.
Here is an alternative Piped link(s): https://piped.video/_Klkr6PtYzI?t=520
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source, check me out at GitHub.
Only if emotions are involved. Of course it’s not possible as long as we train our AI with flawed human-generated data though.
LLM’s require human generated data, that’s how they work.
That’s how we make them work today. It is possible to stay politically neutral in a language though. And therefore your generalized statement is incorrect.
What does politically neutral even mean though? Because it’s definitely not centrism.
How would you train AI without biased data anyways? All data carries some bias from whoever generated it.
A conversation being void of any kind of politics is politically neutral. And that’s most conversations I have.
Most people don’t agree on what subjects are and aren’t political.
It will never be objective if its dataset is something like the internet. It will always be prone to bias because that’s the double-edged sword of LLMs, they have to have vast quantities of data and the only place they can get that is the internet which is biased opinions everywhere.
LLMs are only as “fact based” as the data they source. An LLM in the time of Galileo would have told you the earth is flat.
Weekly reminder that scholars since at least 500 B. C thought the earth was a sphere, so Galileo’s LLM would tell him that, and even tell its size with a reasonable accuracy.
Assuming the sources used for training were the correct ones.
today’s fact is tomorrows hate crime saying it - beware!