Mapping U.S. Federal AI Governance Against Sector Vulnerability

What We Did

• 684 federal AI governance documents were assessed for sector and risk coverage and compared with vulnerability ratings from 272 experts.Coverage captured both how frequently a sector or risk appeared and how much detail documents provided, including specific governance measures.Source: What We Did · p. 4

What We Found

• Public administration and several other sectors receive comparatively high coverage, while finance, healthcare, and social assistance receive less attention despite relatively high vulnerability ratings.Risks related to system security, robustness, and governance are addressed more frequently than most socioeconomic, environmental, and emerging risks; most documents provide limited detail.Source: What We Found · p. 4
• The observed patterns describe AI-specific federal documents and do not establish that the sectors or risks identified are inadequately governed.Relevant protections may come from generally applicable laws, sector-specific regulation, state and local measures, private standards, or other mechanisms outside the analysis.Source: What We Found · p. 4

Why This Matters

• The findings offer a starting point for examining whether existing laws and other governance mechanisms adequately address risks.The paper presents the patterns as useful for policy research and high-level decision-making, especially in public administration.Source: Why This Matters · p. 4, What We Found · p. 4

1 Introduction

• The paper examines concern that AI governance is not keeping pace and that differing understandings of risk contribute to variation in oversight.U.S. governance uses instruments including executive actions, legislation, regulations, agency guidance, and standards.Source: §1 · p. 5
• The analysis measures how frequently and substantively 684 federal AI governance documents address sectors and risk subdomains, then compares those patterns with expert assessments of sector vulnerability.The documents had major activity between January 2020 and January 20, 2026.Source: §1 · p. 5
• The paper provides a descriptive account of AI-specific federal coverage and uses comparisons with expert-rated vulnerability to identify areas warranting further investigation.The comparison is intended as a diagnostic tool, complementing prior work focused on policy trajectories, institutional arrangements, and governance philosophies (Davtyan, 2025; Kalnin, a et al., 2025; Mallinson et al., 2025).Source: §1 · p. 5
• Coverage is uneven: public administration, national security, information, and scientific services receive comparatively high attention, while finance and healthcare receive less despite relatively high expert-rated vulnerability.System security, robustness, and governance-related risks are addressed more frequently than most socioeconomic, environmental, and emerging risks, and most documents provide limited detail.Source: §1 · p. 5

2.1.1 MIT AI Governance Mapping project

• Coverage scores from 1–3 classify how extensively documents address 24 risk subdomains and 14 sectors; good coverage requires specific references and at least multiple sentences.General governance structures do not count as coverage of governance failure unless a document explicitly describes how those structures might fail or become ineffective.Source: §2.1.1 · p. 6

2.1.2 MIT AI Risk Prioritization Delphi Study

• A three-round Delphi study of 272 experts rated vulnerability across the same 24 AI risk categories, using a five-point sector-vulnerability scale.The paper uses those ratings to contextualize possible gaps between expert concern and federal documentary coverage.Source: §2.1.2 · p. 6

2.2 Analysis

• The corpus includes U.S. federal-authority proposals with major activity from January 2020 through January 20, 2026, so coverage primarily measures documentary attention.The analysis filters AGORA documents to U.S. jurisdictions and a single federal policy period.Source: §2.1.1 · p. 6, §2.2 · p. 7

2.2.1 Breadth and Depth of Coverage

• Breadth is the share of 684 documents mentioning a sector or risk, while depth is the share of mentioning documents that provide good coverage.Comparing the measures distinguishes how often an item is discussed from how substantively it is addressed.Source: §2.2.1 · p. 7

2.2.2 Comparing Coverage Against Sector Vulnerability

• Attention divergence compares normalized expert vulnerability with a sector’s share of documents mentioning a risk; positive values indicate vulnerability exceeds coverage.Integrated governance divergence also accounts for depth, and both composite scores represent relative mismatch rather than demonstrated regulatory absence or failure.Source: §2.2.2 · pp. 7–8
• The integrated governance divergence combines sector share with adjusted depth, then compares that coverage value with normalized vulnerability and aggregates the 24 risk-specific divergences by RMS.The depth adjustment uses a floor of 0.1 so coverage remains nonzero when documents mention a sector-risk pair, while still lowering the value when good coverage is absent.Source: §2.2.2 · pp. 7–8

3.1 Degree of Coverage of Sectors and Risk Subdomains

• Public Administration appears in more than half of documents across every risk domain, while Management, Arts and Recreation, and especially Accommodation and Food Services are scarcely covered.Information, Scientific Services, and National Security are also frequently cited across risk subdomains.Source: §3.1 · p. 9, Figure 2 · p. 9

3.1.1 Comparison with Expert Evaluations of Sector Vulnerability

• Finance and Insurance is rated extremely vulnerable in 17 of 24 subdomains but appears in under 30% of documents for any risk domain, with a 20% share for fraud, scams, and targeted manipulation.It has the largest divergence between vulnerability and coverage breadth; Public Administration, by contrast, has one of the smallest divergences despite being widely mentioned and rated less vulnerable in most subdomains.Source: §3.1.1 · p. 10
• Federal coverage breadth tracks expert-rated sector vulnerability unevenly: highly vulnerable sectors can receive limited attention, while the most-covered sectors are not always the most vulnerable.Some national-security risks also diverge from expert ratings; for example, only 44% of documents mentioning false or misleading information address national security, despite expert consensus that the sector is among the most vulnerable to that subdomain.Source: §3.1.1 · p. 10

3.2 Breadth and Depth of Coverage in U.S. Federal AI Governance Documents

• Breadth measures the share of 684 documents mentioning a sector or risk subdomain, while depth measures the share of those documents providing good coverage.Depth scores for low-breadth topics are sensitive to which documents are included; sector shares for sparsely mentioned subdomains are unstable, and the unmentioned AI welfare and rights subdomain affects composite divergence scores.Source: §3.2 · p. 10

3.2.1 Risk Subdomains

• No risk subdomain has a depth share above 50%, and fewer than one-third of documents provide good coverage for most of the four subdomains with breadth above 30%.The four higher-breadth subdomains are system security vulnerabilities and attacks, competitive dynamics, governance failure, and lack of capability or robustness; governance failure is the exception to the low-depth pattern.Source: §3.2.1 · p. 11, Table 2 · p. 12
• Socioeconomic and environmental risks are often in the lowest breadth bands, and Subdomain 6.3 has no document providing good coverage; multi-agent risks are also thinly covered.The text links low coverage to the difficulty of reaching consensus on normative or distributive issues and to the technical uncertainty around emerging multi-agent risks.Source: §3.2 · p. 10, §3.2.1 · pp. 11, 13

3.2.2 Sectors

• Public Administration and National Security are the only sectors with depth above 60%; all other sectors fall below 50%, with Management and Accommodation and Food Services particularly thinly covered.Scientific Services, Public Administration, and National Security are mentioned in more than half of documents addressing them.Source: §3.2.2 · p. 13

3.3 Integrated Governance Divergence

• Finance and insurance and healthcare receive the highest integrated governance divergence scores, while Information and National Security show wider vulnerability-coverage divergence once depth is included.The four sectors are collectively rated as more vulnerable to multiple risks than other sectors; a mention alone does not ensure detailed treatment.Source: §3.3 · p. 13, Figure 4 · p. 14

4 Discussion

• Finance and healthcare receive comparatively limited coverage despite high expert-rated vulnerability, while public administration and several other sectors receive extensive attention.Finance and insurance is mentioned in under 30% of documents addressing any risk domain, despite being rated extremely vulnerable in 17 of 24 subdomains.Source: §4 · p. 15
• Most documents provide only a few sentences of detail, while socioeconomic, environmental, and multi-agent risks are often weakly specified or rarely mentioned.Established categories such as privacy and system security tend to be referenced more widely and receive operational detail, particularly for the public sector.Source: §4 · p. 15
• Coverage gaps cluster around harms that are technically or politically complex, indirect or cumulative, or lack a clear locus of responsibility.The authors suggest that avoiding regulatory overlap or limited technical expertise could also contribute to thin coverage.Source: §4 · p. 15
• Thin guidance may leave accountability gaps when documents name risks to monitor but omit assessment criteria or enforcement mechanisms.The discussion gives examples of security roadmaps without criteria and coordination guidance without assigned responsibilities.Source: §4 · p. 15
• Where federal coverage is thin, finance and healthcare may rely on non-federal initiatives, or state and local governments may need to act.The authors identify these sectors as the clearest cases where attention may not match rated vulnerability.Source: §4 · p. 15
• Proposal-to-enactment delays can extend beyond a year, while emerging risks such as multi-agent interactions may develop faster than governance responses.The discussion reports delays of up to 664 days and notes that the gap between GPT-5 and 5.1 was less than 100 days.Source: §4 · pp. 15–16
• Coverage-vulnerability comparisons are a starting point for understanding governance gaps, not a uniform call for more coverage across every sector and risk.The authors emphasize that sectors differ in exposure and some risks require technical rather than legislative management.Source: §4 · p. 16

5 Conclusion

• Coverage patterns differ from expert vulnerability assessments, and most documents mentioning a sector or risk provide limited detail.Finance and healthcare receive comparatively limited AI-specific coverage despite relatively high expert-rated vulnerability.Source: §5 · p. 17

5.1 Limitations and Opportunities for Future Research

• The AGORA dataset excludes generally applicable laws that may govern AI-related activity without explicitly referring to AI.Thus, the broader governance environment may differ, and AI-specific rules may overlap, complement, or conflict with those laws.Source: §5.1 · p. 17
• The analysis relies on LLM-based classifications, and future work could validate them more extensively and incorporate classification uncertainty..Source: §5.1 · p. 17
• The study covers U.S. federal documents but does not systematically capture state or local governance or interactions across levels.The authors recommend examining coverage and overlap or contradiction across those governance levels.Source: §5.1 · p. 17
• The depth measure records detail about governance measures, not whether they are technically feasible or practically effective.Source: §5.1 · p. 17
• Documents count equally regardless of quality, so narrow documents and major executive actions contribute identically to document counts.The authors note this may misrepresent coverage.Source: §5.1 · p. 17

A Appendix: Document Counts

• Table A.1 presents document counts and corpus proportions by risk subdomain and coverage score.Source: Table A.1 · p. 21, A Appendix: Document Counts · p. 21
• The supplied risk-subdomain entries include governance failure, environmental harm, and multi-agent risks.Source: A Appendix: Document Counts · p. 22
• Table A.2 presents document counts and corpus proportions by sector and coverage score.Source: A Appendix: Document Counts · p. 21, Table A.2 · p. 23

B Appendix: Sector Share

• Figure B.1 uses a conditional percentage: documents mentioning a sector, among those mentioning a subdomain.Source: B Appendix: Sector Share · p. 24, Figure B.1 · p. 24
• The figure’s measure can be read across subdomains to compare sector representation within each subdomain’s documents.Source: Figure B.1 · p. 24, B Appendix: Sector Share · p. 24
• The caption specifies a percentage of documents, not a count of all documents in the corpus.Source: Figure B.1 · p. 24

• Tick marks indicate consensus: at least 90% of responses fall within ±1 of the median and at least 60% are exactly on it.Source: C Appendix: Sector Vulnerability Scores · p. 25
• Figure C.1 displays sector vulnerability scores organized by risk subdomain.Source: Figure C.1 · p. 25
• Consensus indicators are available only at the subdomain level.Source: C Appendix: Sector Vulnerability Scores · p. 25

• Subdomain 7.5, AI welfare and rights, is not mentioned by any documents.Source: D Appendix: Composite Index · p. 26
• Despite no document mentions, Subdomain 7.5 contributes its full normalized vulnerability to every sector’s RMS divergence used to derive the composite scores.Source: D Appendix: Composite Index · p. 26
• Figure D.1 presents the Composite Attention Divergence Index, and Figure D.2 presents the Composite Integrated Governance Divergence Index.Source: Figure D.1 · p. 26, Figure D.2 · p. 27

• Sector 10 is Health Care and Social Assistance, and Sector 14 is National Security.Source: §D.1 · p. 28
• With d0 = 0.05, the positions of Sectors 10 and 14 switch in the robustness check for the IGD score, while the overall ranking pattern remains robust.Source: Figure D.3 · p. 28, §D.1 · p. 28

• Across the AD and IGD results, the major conclusions remain valid after Subdomain 7.5 is excluded.Source: §D.2 · p. 29
• Excluding Subdomain 7.5 leaves the AD ranking largely robust; adjacent Sectors 02 and 03 swap, while finance and insurance retains the largest divergence between vulnerability and coverage breadth.Source: Figure D.4 · p. 29, §D.2 · p. 29
• In the IGD result, adjacent Sectors 03 and 14 swap, but their relatively wider coverage-vulnerability divergence remains.Source: §D.2 · p. 29, Figure D.5 · p. 30