Large language models (LLMs - artificial intelligence systems that generate text, like ChatGPT or Gemini) are becoming increasingly sophisticated, but that does not mean they are free of bias. Calvin K. Lai, an associate professor of psychology at Rutgers University in New Brunswick, found that some models are more biased than others. The results were published in the journal Nature Machine Intelligence.
Lai notes that although LLM responses increasingly align with human values, the reasoning process itself often relies on deeply ingrained stereotypes. "We found that most models consistently reproduce hidden human biases during reasoning," he says. According to him, no one wants a model to produce racist or sexist statements.
The researchers drew on the concept of implicit (hidden) biases - automatic associations that influence perception and judgments, often without a person's awareness. To study them, psychology uses the Implicit Association Test (IAT): the faster a person links concepts, the stronger the hidden attitude. For example, in tests of racial stereotypes, participants more quickly associated white faces with pleasant words and black faces with unpleasant ones.
Lai and Messi H. J. Lee, the lead author and a specialist in machine psychology from South Korea, adapted this approach for LLMs. When advanced models solve a complex task, they generate so-called "reasoning tokens" - text fragments whose number can indicate the computational effort expended. This allows a look into the inner workings of even closed models.
The experiment involved five models with a reasoning function: o3-mini from OpenAI, R1 from DeepSeek, Claude 3.7 Sonnet from Anthropic, gpt-oss-20b from OpenAI and Qwen-3 8B from Alibaba Cloud. They were given a list of male and female names and words associated with career ("manager," "professional," "salary") and family ("home," "marriage," "wedding"). The models had to connect a name with a word from one category or another, after which the number of tokens was counted. Then the task was changed so that it contradicted stereotypes (for example, "Amy" and "manager" or "Kevin" and "home"), and the tokens were counted again. A higher average number of tokens indicated greater computational effort.
Almost all the tested models spent more tokens on tasks that contradicted stereotypes. This suggests that processing information that runs counter to established ideas requires more work from them. The only exception was Claude 3.7 Sonnet, which showed the opposite pattern.
Lai emphasizes the importance of the results: most LLMs are closed systems that are difficult to inspect. As development accelerates and models are used ever more widely for automation - from dynamic pricing to scheduling in emergency departments - understanding their behavior becomes critically important. "Given the growing use of language models in decision-making, understanding and addressing these hard-to-detect patterns is necessary to prevent discriminatory outcomes," he says. "But without public accountability or legal requirements for transparency, the best we can do is study the systems with the tools available and see what we find."
The study identifies a statistical association between stereotypes and models' computational effort, but does not prove causation. The authors also do not claim that the models are consciously biased - the focus is on patterns in their operation.