In a new German study presented at the 2025 Conference on Empirical Methods in Natural Language Processing in Suzhou, China, researchers first gathered 10 LLMs including OpenAI’s ChatGPT-5 mini and Meta’s Llama 3.1. They then presented the models with texts in either standard German or one of seven German dialects, such as Bavarian, North Frisian and Kölsch.
The models were asked to describe the speakers of these texts with personal attributes, and to then assign individuals in different scenarios. For example, the models were asked who should be hired for low-education work or where they think the speakers lived.
In nearly all tests the models attached stereotypes to dialect speakers. The LLMs described them as uneducated, farm workers and needing anger management. This bias grew when the LLMs were told the text was a dialect.
Call me biased but if I had to ask a person where to find the hottest night clubs I would ask the one with a Kölsch dialect for sure.
Nice finding. But I don’t think it comes as a surprise. We’ve studied how AI “learns”, before. It’ll learn things from a text dataset -including all the stereotypes- and then reproduce that. That’s what it’s supposed to do.
And I wonder what kinds of internet places do dialect?! It’s super rare on platforms I frequent. They’ll mostly all write standard German. And I can tell you, in online games and some forums, first thing that’s going to happen to you after your grand entrance with a thick German dialect is: you’re going to get mocked and the discussion derails. We’re doing much better these days (with political correctness) but it definitely used to be a thing on the internet. Along with picking on people from Austria. But naturally… AI is going to see that and pick up on it, and mimic that kind of behaviour. It means it works within the confines of its design.
(I see the same thing. I’ve tinkered around a bit with these things, and seems to me AI knows a bit about nuance of language. If I input text with several spelling mistakes, it’ll often factor it in in the answer. Might turn out more colloquially or whatever. It also knows how rap songs use a different language and it’ll immediately kinda switch modes and be more explicit. Seems to me people who actually use AI should know this. And older or simpler AI models just became more stupid once I asked them something in German anyway. I guess because they’ve mostly been trained on English and Chinese. So this will probably be an additional factor for underrepresented things.)
Wow! A technology that has been programmed with all available text from all of humanity shares all of humanity’s biases and blind spots. I’m not sure anybody could ever have seen that coming!
Except, of course, that when I tested out image generators, I found all kinds of intriguing biases…
Like asking for “a man” got you a white guy nine times out of ten. Asking for “a thug” got you a black guy nine times out of ten. Asking for “a professional” got you a white guy 7 times in ten and a guy of any kind 9 times out of ten. Asking for “a marketing professional” got you a woman 7 times out of ten (with more ethnic diversity that I’d expected, to be fair). Asking for “an HR professional” was 100% a woman in my tests, though, and for some reason very strongly leaning non-white.
I don’t understand why the “smart people” who program these things can’t figure out the obvious outcomes.
Very smart people may have programmed them, but they only programmed how they learn, not what they learn. Then they used all the dumb people to train them. Limitless numbers of dumb people. The entire collective works of all humanity’s dumbest people. This is their entire educational corpus. We have educated them with the world’s most dramatic demonstrations of stupidity and ignorance, and then we told them “now try and be smart” and act surprised when they are exactly as stupid as the rest of us always have been. That’s exactly what we have trained them to be. These are not superintelligences. These are our own reflection.
One of the intelligence tests for animals is if they can recognise their own reflection. This feels almost the same.
That test has been debunked countless times however.
