PAVIA. Large Language Models (LLMs) are linguistic models that underpin most AI applications. These systems, with which we interact more and more frequently, are fed data and texts that incorporate not only social stereotypes but also cognitive biases – that is, patterns of reasoning that characterize how humans perceive the world.
It is from this observation that the paper published in Nature Machine Intelligence takes its starting point. The study is the result of a collaboration between the Department of Neuroscience and Behavioral Sciences at the University of Pavia – with Vittoria Dentella and Luca Rinaldi (also of the Cognitive Psychology Section at the Mondino Foundation of Pavia) – and the University of Milan-Bicocca, with Marco Marelli. The article offers a critical reflection on an increasingly central theme in the scientific debate: is reducing cognitive biases in language models truly desirable?
In recent years, the scientific community’s attention has focused primarily on the need to identify and correct social biases present in AI systems (such as gender and cultural stereotypes), with the aim of limiting the reproduction of inequalities and discrimination. More recently, however, it has emerged that cognitive biases are also detectable in LLMs. As a matter of fact, because LLMs learn from statistical patterns in language, they end up absorbing not only socially prevalent content and associations, but also recurring patterns of human reasoning.
The point raised by the article is that, unlike social biases, cognitive biases cannot be viewed simply as “errors” to be eliminated. While they often represent deviations from models of formal rationality, they can also serve an adaptive function, helping individuals navigate complex, uncertain, or highly context-dependent situations. In other words, what appears to be a logical distortion can sometimes reflect a form of practical and functional reasoning.
This raises a crucial question: is a model with fewer cognitive biases truly a better decision-maker? According to the authors, the answer is not obvious. The idea that mitigating biases produces more rational, objective, and neutral systems is, in fact, based on controversial assumptions. Eliminating a bias does not automatically make a decision neutral: on the contrary, intervening in these mechanisms often involves a normative choice regarding which form of reasoning should be prioritized and which outcomes should be considered preferable. This fact highlights a particularly relevant ethical dimension: determining which biases to retain, mitigate, or correct in AI models is not a purely technical matter. Rather, it means deciding which ways of evaluating the world and making decisions should be incorporated into the technologies we use every day.
The paper calls on the scientific community to open a broader and more rigorous discussion on the very meaning of mitigating cognitive biases in LLMs. Even before developing tools to reduce them, the authors argue, it is necessary to clarify what the ultimate goal of this process is and on what epistemological and ethical foundations it should rest.
At a time when LLMs are increasingly present in everyday life—from the production and transmission of knowledge to automated decision-making processes—understanding the nature of the biases that permeate them becomes essential. The central message of the study is clear: making LLMs less like human reasoning does not necessarily make them more fair, more neutral, or more reliable.
Original paper: https://www.nature.com/articles/s42256-026-01208-w
