“I think we’re seeing truly shocking epithets being assigned to dialect speakers,” said one of the study’s lead authors, Minh Duc Bui, in comments to the outlet.

An analysis by Johannes Gutenberg University found that ten tested models — including ChatGPT-5 mini and Llama 3.1 — described speakers of German dialects (such as Bavarian and Kölsch) using labels like “uneducated,” “farm workers,” and “prone to anger.”

The bias became stronger when the AI was explicitly told that the text was written in a dialect.

Other cases

Researchers have observed similar issues globally. A 2024 study by the University of California, Berkeley compared ChatGPT’s responses to different English dialects (including Indian, Irish, and Nigerian English).

It found that the chatbot responded with more pronounced stereotypes, more degrading content, and a more patronizing tone than it used for standard American or British English.

Cornell University computer science graduate student Emma Harvey called dialect bias “significant and alarming.”

In summer 2025, she and her colleagues also reported that Amazon’s shopping assistant Rufus produced vague or even incorrect answers for users writing in African American Vernacular English. When queries contained errors, the model sometimes replied rudely.

Another illustrative example involved a job applicant from India who asked ChatGPT to proofread an English resume. The chatbot reportedly changed the person’s surname to one associated with a higher caste.

“The mass deployment of language models risks not just preserving entrenched biases, but amplifying them at scale. Instead of mitigating harm, the technology could make it systemic,” Harvey said.

However, the problem is not limited to bias — some models simply fail to recognize dialects. In July, an AI assistant used by Derby City Council (England) reportedly could not understand a radio host’s dialect when she used words like mardy (“whiner”) and duck (“dear/love”) live on air.

What can be done?

The issue is less about the models themselves and more about how they are trained. Chatbots ingest vast amounts of internet text, which then shapes their outputs.

“The key question is who writes that text. If it contains biases against dialect speakers, the AI will copy them,” explained Caroline Holtermann of the University of Hamburg.

She also noted a potential advantage of the technology:

“Unlike humans, bias in AI systems can be detected and ‘switched off.’ We can actively combat these manifestations.”

Some researchers argue that one solution is to build customized models for specific dialects. In August 2024, Acree AI introduced Arcee-Meraj, a model designed to work with multiple Arabic dialects.

According to Holtermann, the emergence of newer and more dialect-aware LLMs makes it possible to view AI “not as an enemy of dialects, but as an imperfect tool that can be improved.”