AI language models exhibit bias against German dialects
A study by researchers at Johannes Gutenberg University Mainz, University of Hamburg, and University of Washington found that ten large language models systematically demonstrate bias against speakers of German dialects, rating them less favorably than Standard German speakers and assigning stereotypical negative attributes in hiring decisions and other contexts.
What happened
Researchers from Johannes Gutenberg University Mainz, University of Hamburg, and University of Washington tested ten large language models including GPT-5, Llama, Gemma, and Qwen using texts written in Standard German and seven German dialects.
In nearly all tests, the models systematically rated speakers of German dialects less favorably than Standard German speakers, assigning dialect speakers stereotypical labels such as "rural," "traditional," and "uneducated," while Standard German speakers were more often described as "educated," "professional," or "trustworthy."
Reported impact
The bias affected hiring decisions, workshop invitations, and residential choices, with dialect-written texts systematically disadvantaged and linked to farm work or anger-management workshops.
Researchers noted that language models associate dialects with negative traits and perpetuate problematic social biases, reflecting stereotypes embedded in training data, even though sociolinguistic research traditionally links dialects to positive traits like friendliness.
Index analysis
All ten tested AI systems reproduced social stereotypes against dialect speakers despite identical content being presented in different language varieties.
The bias was consistent across open-source and commercial models and grew stronger in larger language models.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.