Academic research on government AI has a basic problem: it rarely says what kind of AI it is talking about.
A paper published July 1 reviewed 91 highly-cited public administration studies from 2019 to 2025 and found the field riddled with imprecision. Fifty-five percent of papers left the AI system they studied technically unspecified. Thirty-one percent cited one type of AI as motivation but then studied a different type entirely. Forty-one percent drew conclusions too broad for the system they actually examined. The authors propose a five-category typology — hand-coded, glass-box, black-box, general-purpose, and agentic systems — to give researchers a shared vocabulary.
The distinction matters more than it might sound. A rule-based, hand-coded system that flags benefit applications for review is auditable in ways a black-box model is not. Collapsing both into the bucket of "AI" obscures which accountability and fairness questions even apply. The authors argue that technical precision is not a niche concern but a prerequisite for useful policy research.
The paper also offers a practical diagnostic guide — a short set of questions a researcher can answer from public information, without specialist knowledge, to place any system in the typology. That is a low bar to clear, which makes the 55% underspecification rate look worse, not better.