A Global Comparison of Schemas, Transparency, and I... | AI Research

A Global Comparison of Schemas, Transparency, and Interoperability in Public-Sector AI Registers and Inventories
This research examines how governments around the world track and disclose their use of artificial intelligence. While many countries have created "AI registers" or inventories to make their use of technology more transparent, these tools vary significantly in how they are built and what information they collect. By analyzing 8,368 records across 72 countries, the authors investigate whether these registers are becoming more standardized over time and how different reporting practices shape what the public actually sees regarding governmental AI. The same ai evaluation question is explored in LimiX-2, which adds a research perspective.

How AI Registers Are Built

The study treats AI registers as "policy instruments"—tools that do more than just list technology; they define what is considered important to report. The researchers harmonized 23 key fields across different registers to see how they compare. They found that while most registers share a "descriptive core" (such as the name of a system), they rarely ask for critical details like legal justifications, potential risks, external evaluation results, or how citizens can appeal decisions made by these systems. This suggests that even when registers exist, they may not be capturing the information most necessary for true public accountability.

Patterns in Transparency

One of the study’s central goals was to determine if registers are becoming more similar as they spread globally, a process known as policy diffusion. The results showed no significant evidence of this. There was no clear pattern of registers becoming more alike based on when they were created or where they are located. Furthermore, the researchers found that having a "broader" schema—one that asks for more types of information—did not guarantee that the register would be more complete. Many registers with extensive lists of required fields still suffered from high levels of missing information, indicating that the design of a register is only half the battle; the actual practice of reporting is equally important. The same ai evaluation question is explored in Navigating Sparse Evidence, which adds a research perspective.

The Layered Visibility Framework

To help policymakers and researchers understand these discrepancies, the authors developed a "layered visibility framework." This framework argues that the visibility of governmental AI is not a simple "yes or no" property. Instead, it is the result of four layers:

• Institutional scope: Which agencies and systems are required to report.

• Schema affordances: What specific questions the register is designed to ask.

• Reporting realization: How much of that information is actually filled out by agencies.

• Public representation: The final picture of AI use that is presented to the public.

Key Takeaways for Interoperability

The study highlights that when different organizations or countries try to compare their AI inventories, they often struggle because the data is not interoperable. The authors conclude that for these registers to be useful on a global scale, they require more than just technical compatibility. They need shared concepts, clear and consistent definitions for what is being reported, and preserved "provenance"—a clear record of where the information came from and how it was collected. Without these, registers may continue to provide only a fragmented and inconsistent view of how AI is being used in the public sector. The same ai evaluation question is explored in LLM-Generated Feature Pools for Time Series..., which adds a research perspective. as detailed in the full paper on Arxiv