It’s repeating content that it is trained on and will repeat the same biases present in training data.
Unbiasing LLMs is a huge effort due to the source data being tainted.
It’s indeed a surprise if you believe their claims about LLMs being intelligent.
In other words: “It is indeed a surprise if you are a moron” ;)
This is obvious. It’s literally trained off of English-speaking people’s online comments / posts and designed to give the most likely answer to a question.
I wonder what difference it makes when the user isn’t using English. They don’t mention that they aren’t considering this and don’t mention it on their How it Works page, but they do in the paper’s abstract: “Finally, our focus on English-language prompts overlooks the additional biases that may emerge in other languages.”
They do also reference a study by another team that does show differences in bias based on input language which concludes, “Our experiments on several LLMs show that incorporating perspectives from diverse languages can in fact improve robustness; retrieving multilingual documents best improves response consistency and decreases geopolitical bias”
The subject of how and what type of bias is captured by LLMs is a pretty interesting subject that’s definitely worthy of analysis. Personally I do feel they should more prominently highlight that they’re just looking at English language interactions; it feels a bit sensationalist/click-baity at the moment and I don’t think they can reasonably imply that LLMs are inherently biased towards “male, white, and Western” values just yet.
Kagi had a good little example of language biases in LLMs.
When asked what 3.10 - 3.9 is in english, it fails, but it succeeds in Portuguese, if you format the numbers as you would in Portuguese, with commas instead of periods.
This is because… 3.10 and 3.9 often appear in the context of python version numbers, and the model gets confused, assuming there is a 0.2 difference going from version 3.9 to 3.10 instead of properly doing math.
deleted by creator
In general, calling something that extrapolates and averages a dataset “AI” seems wrong.
Symbolic logic is something people have invented to escape that trap somewhere in Middle Ages, when it probably seemed more intuitive that a yelling crowd’s opinion is not intelligence. Pitchforks and torches, ya knaw. I mean, scholars were not the most civil lot as well, and crime situation among them was worse than in seaports and such.
It’s a bit similar to how you need non-linearity in ciphers.
This is where I try real hard not to come off as an asshole, but…
Is there a correlation between wealthier regions and safer regions? I’ll admit I’ve never looked super hard on the veracity of the data, but that correlation has always tended to check out for me.
It does a hell of a lot worse things than that. try asking it about some of the things trump is doing. Both google gemini and chatgpt have added signifiant political correctness filters.
I’ve been testing the different ai models for a while. they are absolutely terrified of offending trump and maga. They generate pages of gaslighting crap rather than give a straight answer.
And that is on top of their inherit inability to give an accurate answer. I asked about medicare IRMAA income brackets, and it just made numbers up. Luckily I knew the answer before I asked.
Garbage in, garbage out.
It’s not necessarily garbage, but it sure isn’t curated either. Throwing everything into the blender and hoping the mechanism will usually spit out good info is a scientific spinning of the roulette wheel. Sometimes the odds are pretty good. Sometimes they’re horrible, and you should know better than to expect anything but.
But AI has become the shiniest hammer, and every damn thing is a nail now.
If you’re rich, you look good
That’s not news
How large is the corpus of training data in english?
How large is the corpus of training data in all other languages?
