Cognitive bias are thought patterns that can lead to incorrect decisions and judgments. Recently, Large Language Models (LLMs) have been shown to be capable of processing and understanding language well. However, because they learn from human data, they may also pick up human bias. While several scholars have analyzed social bias in LLMs, cognitive bias have not been studied as thoroughly. The goal of this research is to analyze how Artificial Intelligence (AI) may be susceptible to a particular cognitive bias, the attractiveness bias, highlighting the consequences of using LLMs. This study adopts an exploratory, simulation-based approach: rather than analyzing real-world court data, it leverages a Large Language Model to generate a synthetic dataset of 500 defendant profiles and corresponding sentences under controlled conditions. This is a bias according to which people considered more attractive are viewed more positively than those considered less attractive. The use of AI in sensitive fields, such as the legal sector, where neutrality is essential, requires regulations that limit the risk of bias. However, the results show that AI technologies are susceptible to cognitive bias, highlighting a lack of clarity and robustness at key decision-making moments.
Attractiveness bias in Artificial Intelligence systems
Maria Grazia Olivieri
;
2026-01-01
Abstract
Cognitive bias are thought patterns that can lead to incorrect decisions and judgments. Recently, Large Language Models (LLMs) have been shown to be capable of processing and understanding language well. However, because they learn from human data, they may also pick up human bias. While several scholars have analyzed social bias in LLMs, cognitive bias have not been studied as thoroughly. The goal of this research is to analyze how Artificial Intelligence (AI) may be susceptible to a particular cognitive bias, the attractiveness bias, highlighting the consequences of using LLMs. This study adopts an exploratory, simulation-based approach: rather than analyzing real-world court data, it leverages a Large Language Model to generate a synthetic dataset of 500 defendant profiles and corresponding sentences under controlled conditions. This is a bias according to which people considered more attractive are viewed more positively than those considered less attractive. The use of AI in sensitive fields, such as the legal sector, where neutrality is essential, requires regulations that limit the risk of bias. However, the results show that AI technologies are susceptible to cognitive bias, highlighting a lack of clarity and robustness at key decision-making moments.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
