When AI judges AI: Who defines quality in responsible NLP?

Abstract What gives an automated judgment the authority to define a good answer? Large language models increasingly evaluate the outputs of other models, helping researchers compare systems, select training data, and inform deployment. However, agreement with human judgments cannot settle whose understanding of quality should guide these decisions. Automated evaluation can influence how quality itself is defined. When a rubric becomes a benchmark, a selection rule, or a training signal, its assumptions can become standards that others inherit. This piece develops bounded evaluative authority: reliance on an automated judgment should extend only to interpretations and decisions supported by evidence and accountable reasons within a stated scope. Drawing together measurement theory, evaluator research, and philosophical work on epistemic injustice, it distinguishes the ability to apply a criterion from the authority to choose it. Reliability, validity, and legitimacy address connected questions about consistency, meaning, and justified reliance. The discussion considers whose knowledge is recognised, who may resolve conflicting values, and when reasonable disagreement should remain visible rather than disappear into a single score. An Evaluative Authority Envelope makes the grounds of delegated judgment open to scrutiny. The argument extends earlier pieces in Responsible NLP on the limits of performance and the institutional distribution of responsibility. Its central claim is that responsible evaluation requires more than accurate scoring: it requires an account of why particular judgments should count, for whom, and with what possibility of challenge. A score can inform a decision; it cannot supply its own warrant to govern it.