Study finds large language models mirror human attitudes toward Dutch multiethnolect Straattaal
Researchers report that LLMs, like human participants, associate the Dutch multiethnolect Straattaal with people of migrant background, though interpretation questions remain.
In plain words
Straattaal is a Dutch street-language variety linked in public perception to migrant communities. The study's authors tested whether large language models carry the same associations as human speakers. They found the models do, and that the pattern is stronger in larger and newer models.
What was announced
The arXiv abstract describes matched study designs for humans and models. Human participants completed an Implicit Association Test plus an explicit rating task, while the models were assessed with an LLM Word Association Test and the same explicit rating task. Human participants strongly linked Straattaal with people with a migration background at both implicit and explicit levels, and the models showed consistent associations of the same kind, with stronger links in larger and newer models. When given the identical explicit rating task, the models produced response patterns comparable to the human ones.
The authors conclude that many LLMs encode language attitudes resembling those observed in humans, but they also raise questions about how concepts and measures built for studying human attitudes should be interpreted when applied to LLMs. They argue the results support bringing sociolinguistic perspectives and language variation into NLP research to help identify and mitigate socially and linguistically embedded biases.
What it does not show
- The evidence is the paper's abstract only; full methods, model lists and detailed results are not available in the supplied passages.
- The abstract does not name which specific models were tested, so the reported model-size pattern cannot be verified in detail.
- The authors themselves note that applying human attitude concepts and measures to LLMs requires careful interpretation, so the findings do not establish that models hold attitudes in the human sense.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.