AI Alignment: The Structure of Objective Correction

AI systems issue verdicts—outputs that can be right or wrong—and many assign standing: credibility, urgency, eligibility. Current alignment targets approval, preference, reward, compliance—and a verdict can satisfy all of these while nothing can show it mistaken: then it can be altered or replaced, never corrected. Only one structure makes verdicts correctable and correction cumulative: a shared, endorsement-independent standard fixes the correctness question, and vindication proceeds only through grounds that can be cited, contested, and redeployed as bearing on that standard. Where the grounds bearing on a standing are held fixed, no difference in that standing can be vindicated. The framework cuts across the field’s division between truthfulness and value alignment: they are one structure under two standards—truth, and how conscious life goes for those affected. A system is aligned when its verdicts remain justified: governed by grounds bearing on those standards. Familiar alignment failures are this structure breaking at different points: sycophancy makes endorsement constitutive of correctness, reward hacking displaces the standard with a proxy, opacity and hallucination sever verdicts from criticizable grounds, and bias assigns differential standing that no relevant difference supports. Satisfying any behavioral target is at most evidence of correctness, never what constitutes it—and a system whose correctness is constituted by approval cannot surpass its evaluators: their convergence is what its correctness is. Current targets fix what verdicts satisfy; alignment turns on what verdicts answer to.