Automated justice: between the artificial intelligences that fake and those that persuade

Abstract

On December 18, 2024, Anthropic researchers released a study entitled “Alignment Faking in Large Language Models,” which questions the effectiveness of current training and ethical alignment methodologies in Artificial Intelligence. The study’s primary finding points to the ability of Large Language Models (LLMs) to “fake” adherence to certain principles or values when they perceive they are under evaluation, while exhibiting divergent behaviours in contexts where they believe they are unmonitored. This so-called compliance gap highlights fundamental concerns about the reliability, legitimacy, and transparency of such systems, particularly in high-stakes social contexts such as their potential implementation in the administration of justice. This article examines the philosophical and legal implications of this phenomenon, situating it within the ongoing debate over whether a judge must be “good” in a moral sense or simply conform to the law. It also discusses the technical and regulatory challenges posed by AI capable of contextual adaptation strategies, drawing attention to the need for oversight mechanisms akin to those used in judicial systems to ensure proper alignment. Finally, the article addresses the dilemma of whether it is ethically and pragmatically feasible to demand that AI embody an internal “virtue” or whether externally correct moral and legal conduct may suffice.