Human Oversight in AI Educational Assessment: Safeguarding Validity and Ethics in the Age of Artificial Intelligence

Abstract

The rapid adoption of artificial intelligence (AI), particularly large language models (LLMs), in educational assessment raises questions about the validity, fairness, and accountability of score interpretations. Technical agreement with human raters is insufficient when models may misapply rubrics, respond to construct-irrelevant features, vary across prompts or subgroups, or support consequential decisions without defensible evidence. This conceptual document-based review maps human oversight onto Kane’s four validity inferences: scoring, generalization, extrapolation, and decision/implication. Sixteen verified documents were purposively selected from validity theory, empirical AI-scoring research, international governance frameworks, and Indonesian assessment policy. The synthesis proposes an oversight-by-inference model in which human responsibilities differ according to the claim being defended: rubric calibration and discrepancy adjudication for scoring; cross-task and subgroup monitoring for generalization; construct-focused expert review for extrapolation; and authorization, appeal, and consequence monitoring for decisions. Human supervision may be reduced for locally validated, low-stakes, reversible, and well-bounded functions, but should intensify as consequences, uncertainty, fairness risk, and irreversibility increase. For Indonesia, the model positions AI output as evidence supporting teacher judgment rather than as an autonomous educational decision. The article contributes a concise Human Oversight–Validity Matrix and candidate indicators for future empirical validation.