The Conditions of Agreement: AI Alignment and the Authority to Judge It

An AI assistant may be accurate, accept corrections and earn its user’s trust while becoming the main route through which that user encounters information, remembers past decisions and considers alternatives. Under these conditions, approval alone cannot establish that the relationship remains open to challenge. This essay distinguishes the ability to judge a reason once it is presented from the opportunity to encounter it in the first place. It develops non-appropriation as a requirement for continuing assistance: people must retain practical means of challenging its framing, mandate and continuation, including the ability to seek help elsewhere without losing access to the reasons for earlier decisions. The argument connects non-domination, contestable AI and agency preservation to concrete design requirements: independent routes to sources, transferable records distinguishing human authorisations from assistant inferences, and competent support after disagreement. A proposed experiment separates access to neglected reasons from judgment after disclosure. The essay offers a conceptual argument and a testable research proposal; it reports no new experimental results.