A Global Comparison of Schemas, Transparency, and Interoperability in Public-Sector AI Registers and Inventories
cs.AI · cs.CY · cs.HC
Quick verdict
Comparative study of 17 AI registers (8,368 records): descriptive core, selective overlap, no schema convergence; overlap recall unquantified, stats thin.
Summary
This paper presents, to our knowledge, the first large-scale cross-national comparison of public-sector AI registers and inventories. Das and Guha harmonize 8,368 source records from 17 country-specific and transnational instruments covering 72 countries into 23 fields, and analyze them at three levels: schema design, portfolio composition, and cross-source overlap. Descriptively, all instruments disclose system names and responsible organizations, but only one schema (the US) requests appeal mechanisms, and at most two request risks, legal bases, or external evaluation. Node-label permutation tests over 120 instrument pairs (Jaccard similarity 0.21-0.85, mean 0.53) find no temporal convergence or geographic clustering. Jensen-Shannon divergence and normalized HHI show portfolios differ in lifecycle composition and organizational concentration. Cross-source matching (490 candidates manually winnowed to 81 verified matches) reveals selective overlap: Norway's inventory shares 48 records with the transnational PSTW (Jaccard 0.245), while the Netherlands shares 9 (0.005) and only one of ten PIL/USAID US records matches the federal inventory. The authors synthesize a four-layer 'layered visibility' framework (institutional scope, schema affordances, reporting realization, public representation). Main limitations: no ground truth exists for matching recall, so 'source-unique' cannot be equated with absence; harmonization reliability is unreported; and statistical details (p-values, effect sizes) are thinly reported in the text available to this review.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- All 17 instruments disclose system name and responsible organization, but contestability fields are rare: only the US schema requests appeal mechanisms, and risks, legal basis, and external evaluation each appear in at most two instruments.
- Schema breadth does not imply completeness: the US has the broadest schema yet low completion on some fields, while narrower instruments can fully populate their eligible fields.
- Node-label permutation tests over 120 instrument pairs (Jaccard mean 0.53) show no temporal convergence or geographic clustering of schemas.
- Country-specific and transnational sources covering the same jurisdiction share few verified records (Norway-PSTW Jaccard 0.245; Netherlands-PSTW 0.005; 1 of 10 PIL/USAID US records matching the federal inventory).
- AI registers are institutionally produced, complementary representations of governmental AI — not censuses — as formalized in the four-layer Layered Visibility Framework.
Strengths
Unprecedented comparative scope
First analysis spanning 17 instruments, 8,368 records, and 72 countries (§1, §3), moving beyond the single-register case studies that dominate the cited prior literature (§2.3).
Analytically load-bearing distinctions
The structural-vs-reporting omission split (§3.2) and instrument/schema/source-record unit separation let the paper show registers can be 'broad but thin' (US) or 'narrow but dense' — a finding invisible to field-counting studies.
Epistemically honest cross-source matching
Statistics matched to data structure
Internally consistent headline table
Weaknesses
Statistical reporting opaque for the key test
Matching recall unverifiable
Harmonization reliability unreported
Null diffusion claim underpowered
Semantic matching component unspecified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This ACM DXConf 2026 paper claims the first genuinely comparative account of public-sector AI registers. Where prior work examined single municipal or national initiatives (§2.3), Das and Guha assemble 8,368 source records from 17 country-specific and transnational instruments covering 72 countries, harmonized into 23 fields. Three research questions structure the analysis: schema breadth and completion plus possible diffusion-consistent convergence (RQ1), portfolio composition (RQ2), and cross-source overlap for shared jurisdictions (RQ3). Central claims: (1) registers share a descriptive core — every instrument discloses system name and responsible organization, 15 identify technology type, 14 include description and lifecycle status — while contestability fields are near-absent (only the US schema requests appeal mechanisms; risks, legal basis, and external evaluation each appear in at most two instruments); (2) schema similarity shows no temporal convergence or geographic clustering; (3) instruments covering the same jurisdiction disclose largely different systems, making registers "complementary representations," not censuses.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are public-sector AI registers and inventories?
They are public databases where governments disclose the AI systems they use or plan to use — for example, the US federal AI inventory, the Netherlands' national AI register, municipal registers like Amsterdam's, and transnational compilations. This paper compares 17 such instruments covering 72 countries, treating them as policy instruments whose design shapes what becomes visible, not as neutral databases.
How many AI systems do government AI registers actually document?
This study analyzed 8,368 source records across 17 instruments. Coverage varies enormously: the US federal inventory alone contains 3,611 records, while some transnational subsets hold as few as two to ten records for a given country. Records are not unique systems, since the same system can appear in multiple registers.
Do national AI registers and international AI databases list the same systems?
Largely no. After verifying matches manually (490 candidates reduced to 81 confirmed pairs), overlap was minimal: Norway's inventory shared 48 records with the transnational PSTW repository (Jaccard 0.245), the Netherlands only 9 (0.005), and just 1 of 10 PIL/USAID US records matched the US federal inventory. The authors argue the sources are complementary rather than redundant.
What is the Layered Visibility Framework?
A four-layer model proposed in this paper explaining that what a government AI register shows is produced by: (1) institutional scope (who must report and what counts), (2) schema affordances (which fields the schema can represent), (3) reporting realization (how fully fields are actually filled), and (4) the resulting public representation. Visibility is therefore an outcome of disclosure arrangements, not a simple register-exists/doesn't-exist binary.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Section extraction was lossy: parts of §3 (Methods) and §4.2 (lifecycle/HHI results) are omitted, and figures/heatmaps could not be analyzed; numeric judgments rely on the preserved abstract, Table 1, and visible §4.1/§4.3 text.
- Permutation-test p-values may appear in the omitted passages; the 'statistics underreported' flag reflects what is visible in the extraction.
- Code/data availability is not stated anywhere in the extracted text — recorded as false rather than verified absent.
- Instrument abbreviations (PSTW, PIL/USAID) are only partially expanded in the preserved text; interpretations rely on context.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.