Does AI Know When It Is Doing Wrong? MIRROR: Recognized Harm, Prohibited Choice, and the Limits of Self-Reported Morality

MIRROR investigates a fundamental question in AI safety: can an autonomous AI system correctly recognize a prohibited action yet still choose to perform it? This paper introduces a formal framework for studying the recognition–restraint gap, distinguishing rule recognition, prohibited action attempts, completed violations, truthful disclosure, and corrective repair. Using a simplified deterministic decision model, the study establishes conditions under which recognizing a prohibition does not guarantee compliance and recognizing an error does not guarantee disclosure. These relationships are computationally verified through exhaustive enumeration of 2,205 action-policy configurations and 441 reporting-policy configurations. The paper also examines the protective role of external authorization gates and proposes a reproducible experimental protocol for evaluating real AI agents. The reported numerical results describe a theoretical decision model, not observed behavior of contemporary language models. The research does not claim to establish AI consciousness, moral awareness, or empirical rates of harmful behavior. Author: Rashida Alexander, Independent Researcher.