Nagy nyelvi modellek igazságossági érvelésének vizsgálata a „statisztikai alapú érvelés” mítosza és egy büntetőeljárási kísérlet

Absztrakt

This study examines whether the fairness and moral reasoning of large language models can truly be regarded as value-neutral, with particular attention to their use in legal decision support. Its starting point is that generative artificial intelligence has become embedded in social and institutional practices more rapidly than the development of reliable knowledge about its limitations, risks, and regulatory conditions. The article therefore reviews the relevant literature on the implicit value structures, moral preferences, political and cultural biases, and legal, moral, and fairness-related reasoning capacities of large language models. It argues that model outputs should not be understood merely as products of statistical text generation, since they often display coherent normative patterns that may create risks in decision-support contexts. The paper also presents an ongoing empirical research project based on anonymized Hungarian criminal cases (human smuggling), designed to measure how models weigh aggravating and mitigating circumstances and whether implicit normative alignment or discriminatory bias can be detected in sentencing-related reasoning. The study concludes that the legal use of generative models can only be justified if their normative operation becomes auditable, transparent, and controllable.