How Human Cognitive Biases Influence AI Behavior and How to Detect and Mitigate Them

What is it about?

This research explores how large language models (LLMs) can be influenced by cognitive biases—systematic patterns of flawed reasoning that also affect human thinking. While most studies focus on social biases (such as gender or racial bias), this work looks at cognitive biases like agreeing too easily, following the majority, or relying on recent information. We design a three-step experimental approach to study how these biases affect AI responses, how they can be detected using reasoning techniques and external knowledge, and how their impact can be reduced. Our findings show that these biases can significantly affect the accuracy of AI systems, but also that interventions—such as making the model aware of the bias—can improve its reasoning and performance.

Featured Image

Photo by Merrilee Schultz on Unsplash

Photo by Merrilee Schultz on Unsplash

Why is it important?

As AI systems are increasingly used in areas like healthcare, education and decision-making, it is essential that they provide reliable and unbiased outputs. Cognitive biases can lead AI systems to produce misleading or incorrect answers, even when they appear confident. This research highlights a less explored but critical type of bias in AI and shows that it is possible not only to detect it, but also to reduce its impact. By improving how AI systems reason, this work contributes to making them more trustworthy, transparent and safe to use in real-world applications.