Beyond Hallucinations: Corrigible Judgment Execution in Large Language Models

Large language models are often criticized for “hallucinating” — producing answers that sound confident but turn out to be wrong or misaligned. This paper argues that many such failures are not simply flaws in the models themselves, but a result of how they are used. In real work, people rarely know all their requirements in advance; intent and constraints emerge gradually through interaction. When AI systems are forced to deliver a final answer too early, before this understanding has formed, misalignment is almost inevitable. The paper introduces a new way of thinking about this problem by focusing on judgment execution — how decisions are formed over time — rather than on isolated answers. It shows that when AI systems remain open to guidance, clarification, confirmation, and correction as work unfolds, people can efficiently steer them toward acceptable outcomes. This perspective reframes hallucinations as a process-level issue and highlights why interactive, corrigible use of AI is essential for reliable collaboration under uncertainty.