A Technical Typology of AI Systems in Public Administration

The Problem with Treating AI as a Monolith in Government

A new preprint from arXiv (2606.31755v1) tackles a persistent blind spot in public administration research: the tendency to treat "AI" as a single, undifferentiated technology. The authors argue that this lumping obscures critical technical distinctions—such as whether a system uses rule-based logic, supervised machine learning, or generative models—which in turn shape how those systems affect core public values like transparency, accountability, and procedural fairness.

This is not an abstract academic quibble. When a city deploys a predictive model for welfare fraud detection, it is fundamentally different from using a large language model to draft public notices. The first relies on statistical correlations and may perpetuate historical biases; the second introduces issues of factual reliability and authorship. Both are "AI," but their failure modes, regulatory needs, and oversight mechanisms diverge sharply.

Why This Matters for Public Trust

The paper’s core insight is that policy and governance frameworks cannot be effective if they address a phantom category. Current regulatory efforts—from the EU AI Act to local government AI guidelines—often categorize systems by risk level, but risk itself depends on technical architecture. A high-risk classification for a neural network does not translate cleanly to a decision tree or a retrieval-augmented generation pipeline.

For public administrators, this means that blanket procurement rules or one-size-fits-all transparency requirements will either be too restrictive for low-risk systems or too lax for high-risk ones. The authors provide a typology that could help governments match oversight to actual technical behavior—for example, requiring explainability reports for models that learn from data, but not for deterministic rule-based systems.

Implications for AI Practitioners

For those building or deploying AI in government, this research offers a practical lens. First, it demands that technical teams articulate which type of AI they are using and why that distinction matters for compliance and ethics reviews. Second, it suggests that impact assessments should be tailored to the system’s learning paradigm—supervised, unsupervised, reinforcement, or rule-based—rather than applied generically.

Third, the paper implicitly warns against the "AI washing" of legacy automation. If a system is essentially a set of if-then rules, calling it "AI" may trigger unnecessary regulatory burdens or, worse, create false expectations of adaptability. Conversely, a genuinely adaptive system requires ongoing monitoring that static rules do not.

Key Takeaways

• AI is not a single technology in government contexts. Technical distinctions between rule-based, machine learning, and generative systems have direct implications for public values like fairness and accountability.
• Generic AI governance frameworks risk being ineffective. Oversight must be tailored to the specific learning paradigm and failure modes of each system.
• Practitioners must clearly classify their systems. Teams should document whether their AI is deterministic, statistical, or generative, and adjust compliance and monitoring accordingly.
• Avoid conflating automation with AI. Legacy rule-based systems should not be rebranded as AI, as this can distort risk assessments and regulatory obligations.