When AI Draws the Map, the Territory Changes: A Study on Classifying Public Administration Research

Ask five different systems to define what counts as "public administration" scholarship, and you will get five different disciplines. That is the core finding of a new paper, "When AI Classifies: What Counts as Public Administration?", which tests how AI-assisted and citation-based classification tools carve up the academic literature on public administration (PA) and AI-in-PA research.

The study, submitted to arXiv on June 26, 2026, draws on two major bibliometric databases, Web of Science and OpenAlex, to compare five distinct methods of grouping scholarly output. Some of these methods rely on how authors themselves label their work. Others use citation patterns to infer disciplinary membership. A third category leans on AI-assisted tools that classify papers algorithmically, without direct human tagging. The researchers set out to see whether these approaches, despite their different mechanics, would converge on a roughly similar picture of the field.

They did not. The five approaches produced corpora that varied sharply in size, and diverged in the types of publications they captured, the journals and outlets where that work appeared, how the field's output changed over time, and the thematic clusters that emerged from the text. More strikingly, the paper reports that these were not overlapping subsets of one shared literature. The classification methods effectively surfaced separate bodies of knowledge, with no shared publications and no shared publishing venues across the five representations.

Why the divergence matters

That result reframes what classification systems actually do. Rather than acting as neutral filters that simply sort existing scholarship into the right bins, the paper argues that AI-enabled and citation-driven classifications actively interpret the field, and in doing so, help construct it. A tool trained on certain citation signals or textual patterns will surface a version of "public administration" shaped by its own assumptions, not by some fixed, objective boundary that all methods would eventually find if run long enough.

The authors flag a further risk: these classification systems can become self-reinforcing. If a database's AI-assisted tagging shapes which papers get grouped together, cited together, and indexed together, future scholarship built on that foundation may simply entrench the original classification choices rather than challenge them. Over time, that dynamic could narrow or freeze how a discipline's boundaries are understood, even as the underlying research evolves.

The paper's conclusion is measured rather than alarmist. It does not call for abandoning AI-assisted classification tools, but insists that human disciplinary judgment remains indispensable alongside them. Algorithmic sorting, in this framing, is a complement to expert judgment about what belongs in a field, not a replacement for it.

The study is filed under Digital Libraries, Artificial Intelligence, Computers and Society, and Databases on arXiv, reflecting its position at the intersection of bibliometrics, AI methodology, and the sociology of academic disciplines. For research administrators, university rankings bodies, and anyone using AI-assisted tools like OpenAlex to map scholarly fields, the findings are a caution against treating any single classification output as ground truth.