LLM e Bias. L'illusione della neutralità

This article challenges that illusion through a critical review of the main forms of bias affecting such systems, distinguishing four layers: social bias inherited from training corpora, cognitive-like bias in reasoning tasks, bias introduced or amplified by alignment procedures, and sycophancy, the tendency of chatbots to validate the user. Drawing on work in word embeddings, cognitive psychology and reinforcement learning from human feedback, it argues that LLMs are not less fallible than humans but fallible in different ways. Their fluent, confident and seemingly impartial responses make these biases harder to detect than human ones. Linking sycophancy to the bias blind spot – users perceive a chatbot that agrees with them as more impartial than one that challenges them – the article argues that the illusion of technological neutrality amplifies epistemic harm: the more objective the machine appears, the more its confirmations distort human judgment.