The language models behind search engines, translation tools, and voice assistants don't speak equally for everyone. A University of Copenhagen study found systematic bias in how well these models match different demographic groups' language. Young white speakers get the best alignment; young non-white men get the worst. The researchers say this affects information access for everyone.
At a glance
- Language models align best with young white speakers under 40 with shorter education
- Young non-white men show the worst alignment—up to one in ten predictions differ significantly
- Models are trained on data mostly from social media and Wikipedia
- Young whites contribute most to training data through online forums
- Larger models trained on more data show stronger biases
Why this matters for information access
When you search online, language models help predict what results to show. If your phrasing doesn't match what the model was trained on, you might get different or fewer results than someone whose language patterns align better. Professor Anders Søgaard, the study's lead author, says this means "information available to others may not be available to you."
The problem extends beyond search. Insurance companies use language models to assess risk. Legal systems use them to find similar cases. In these contexts, small accuracy differences can have major consequences.
Where the bias comes from
Language models learn by processing huge amounts of text to predict word probabilities. The texts come mostly from online sources—social media and Wikipedia. These sources aren't representative of all people using technology. Wikipedia content is primarily written by young white users. Social media contributions also skew toward this demographic.
The team hasn't pinpointed exactly why young white speakers align best with these systems, but Søgaard offers a hypothesis: the demographic most represented in training data—drawn heavily from social media—contributes disproportionately to the text pools these models learn from.
The growing problem
As computing power increases and more data becomes available, models get larger. Søgaard says larger models show stronger biases, not weaker ones. "Unless something is done, the gap between certain social groups will widen."
What might fix it
Simply adding more data won't help. The solution is better training. Algorithms could be changed to weight data differently rather than equally—being especially cautious with data from less-represented populations to balance outcomes.
Practical example
Say you search for medical information using terms common in your community. A language model trained mostly on different communities' phrasing might not recognize your query as relevant, returning fewer or different results than someone using "standard" phrasing. This could delay diagnosis or treatment for conditions described differently across demographics.
What to watch next
Watch for algorithmic fairness research to influence how companies train language models. Also watch for regulation requiring transparency in how these systems affect different groups. The researchers' findings suggest technical solutions exist—implementation depends on industry and policy priorities.
Key takeaways
- Language models work better for some demographics than others
- Training data from social media and Wikipedia skews toward specific groups
- Larger models show stronger biases, not weaker ones
- Better training methods—not more data—could reduce disparities
Frequently asked questions
Which groups' language do language models match best?
Young white speakers under 40 with shorter education show the best alignment with language models.
Which groups have the worst alignment?
Young non-white men show the worst alignment, with up to one in ten predictions significantly different compared to young white men.
Why do these biases exist?
Models are trained on data mostly from social media and Wikipedia, where young white users contribute the most content.
How can language model bias be fixed?
Researchers suggest changing algorithms to weight data differently rather than equally, being especially cautious with data from less-represented populations.