The study reported in Poland on 12 November is a reason to audit language models beyond their polished final answers. In almost all of five tested reasoning models, pairings that ran against familiar stereotypes used more reasoning tokens. That does not establish that every such system makes discriminatory decisions. It does indicate that a neutral-looking final response can conceal a different processing pattern.
Researchers Calvin K. Lai and Messi H.J. Lee adapted the Implicit Association Test for LLMs. Five models were given male and female names and words associated with careers or family. The researchers then compared the reasoning-token counts for conventional associations and for stereotype-incongruent pairings, such as a female name with an executive role. In the report, more tokens represent greater computational effort.
Almost every model tested used more effort for the stereotype-incongruent tasks. Claude 3.7 Sonnet was the exception, showing the reverse pattern. The finding does not mean language models possess prejudice in the human sense, nor does it settle the result of a particular pricing, hiring or healthcare decision. The report does not provide the full numerical results, statistical uncertainty, replication evidence or whether the test versions match current deployments.
That is why the useful conclusion is specific rather than sweeping. Organisations using LLMs in consequential settings should test both outputs and model behaviour on matched stereotype-reversing prompts. Filters applied to final text alone may miss relevant patterns. As the report notes, many models remain closed systems, so independent scrutiny will often have to work through indirect tests of this kind until transparency improves.
Sources
- Nauka w Polsce · Fundacja PAP — Źródło: Nauka w Polsce (Fundacja PAP) (https://naukawpolsce.pl/) · Source date: 2026-11-12
Independent editorial content. WUA is not affiliated with any political party. 2026-11-12