Jonathon's AI Wiki

Stanford HAI AI Index 2026 — Chapter 8 Policy and Governance Deep-Dive

17 min read

Source: raw/ai_index_report_2026.pdf (Chapter 8 “Policy and Governance,” pp. 323-359); Epoch AI methodology documentation (ai-research/, see below)

Deep-dive companion to Stanford HAI AI Index Report 2026, which covers only the chapter-map summary. This article extracts the full Chapter 8 content — the 2025 global policy timeline, national AI strategies, the report’s new five-dimension AI sovereignty framework, and legislative/congressional-hearing data — plus a methodology deep-dive on Epoch AI (the data partner behind this chapter’s model-count figures and Chapter 1’s notable-models dataset) that resolves a standing question about non-English-source coverage gaps.

Key Takeaways

  • 2025 was a bifurcation year for AI policy: the EU AI Act’s first prohibitions took effect (Feb 2, 2025) in the same year the US shifted decisively toward deregulation (executive orders in January and July 2025; the Senate stripped a proposed 10-year federal moratorium on state AI regulation on July 1).
  • National AI strategies are expanding fastest among countries that had none five years ago — in 2024, more than half of newly adopted strategies came from emerging economies (sub-Saharan Africa, South/Central Asia, Latin America).
  • “AI sovereignty” is a new analytical framework this edition, decomposed into five dimensions (infrastructure, data, model, application, talent) — each with different leaders. China leads infrastructure sovereignty by raw supercomputer count (85 vs. Europe/Central Asia’s 44 and North America’s 41); East Asia/Pacific leads data-localization measures (77); the US leads model releases (1,618 cumulative) but China is scaling fast (151→849, 2022-2025).
  • US AI-related congressional witness testimony grew 20x since 2017 (5 witnesses in 2017 → 102 in 2025); industry’s share of witnesses nearly tripled (13%→37%) while academia’s fell (to 15%).
  • **US public AI investment (285.9B in 2025 alone) — public investment is not the lever driving the US AI economy.
  • Epoch AI — the data partner behind this chapter’s sovereignty figures and Chapter 1’s “notable models” dataset — has no formal published policy on language/geographic coverage, but has repeatedly, publicly acknowledged (in its own blog, not under external pressure) that non-English-language models, especially Chinese-language ones, are harder for its team to discover and “may have flown under the radar.”
  • The OECD.AI Index is a real, distinct publication (Feb 2026, 56pp, DOI-registered) that complements rather than duplicates HAI — a composite policy-implementation score across 38 OECD countries, vs. HAI’s narrative chapter and separate 36-country Global AI Vibrancy Tool. Stanford HAI sits on the OECD.AI Index’s own Expert Group; the two are collaborators.

8.1 Major Global AI Policy News in 2025 — Timeline

Selected events (full source: Digital Policy Alert, congressional/executive records):

DateEvent
Jan 23US Executive Order “Removing Barriers to American Leadership in AI” — rescinds earlier AI directives, establishes a pro-innovation, deregulatory federal posture
Feb 1UK positions itself as first country to criminalize AI tools used to generate CSAM
Feb 2First measures of the EU AI Act take effect — bans high-risk uses (predictive policing, emotion recognition), sets stage for stricter rules
Feb 11Paris AI Action Summit: US and UK decline to endorse a 60-country inclusive/ethical-AI declaration
Mar 14China finalizes mandatory labeling rules for AI-generated content
Mar 24Zimbabwe/Cassava Technologies partners with Nvidia on Africa’s first “AI factory”
Apr 3Kigali Global AI Summit on Africa — AI opportunity and labor-market risk
Apr 16Montana enacts the Right to Compute Act (SB 212)
May 17African Union declares AI a strategic priority
Jun 17California AI Policy Report (Newsom-commissioned) warns of “irreversible harms”; same day, G7 issues joint statement on AI governance coordination
Jul 1US Senate strikes the proposed 10-year federal moratorium on state AI regulation from a major spending bill, leaving states free to proceed with their own AI oversight laws
Jul 10EU releases voluntary Code of Practice for general-purpose AI (transparency, copyright, safety)
Jul 23US launches “America’s AI Action Plan” — broad strategy covering innovation, infrastructure, diplomacy, plus 3 executive orders (data centers, exports, government procurement)
Jul 26China announces a 13-point action plan for global AI governance coordination (World AI Conference)
Aug 2EU general-purpose AI obligations take effect (risk assessments, transparency, systemic-risk mitigation)
Aug 26UN General Assembly approves an Independent International Scientific Panel on AI + Global Dialogue on AI Governance
Sep 17Italy becomes the first EU member state to pass its own national AI law (complementing EU-level regulation)
Sep 29California enacts SB 53 (AI Safety/Transparency Law) — mandates safety-protocol disclosure and whistleblower protections for large model developers
Oct 8European Commission launches its European Strategy for AI in Science, alongside the “Apply AI Strategy”
Nov 24US Executive Order launches the “Genesis Mission” — a Manhattan-Project-scale initiative (DOE-led) to accelerate AI-driven scientific discovery
Dec 11”Pax Silica” declaration — US-led AI/tech supply-chain cooperation initiative signed by multiple countries
Dec 12US Executive Order seeks to curb/preempt state-level AI laws — the deregulatory posture continuing into year-end, in direct tension with the July 1 Senate action

Reading the arc: the year opens and closes with the US executive branch pushing to limit state-level AI regulation (Jan 23 EO, Dec 12 EO), while in between, the Senate explicitly preserved states’ ability to regulate (Jul 1) and California/Texas/Montana/Utah all passed substantive state AI laws. This is the “shift toward deregulation” the drain question asked about — it’s a federal-executive-branch posture, actively contested by both the legislative branch and individual states, not a uniform national trend.

8.2 National AI Strategies

Oxford Insights data (tracks published strategies, not implementation quality — “results should be viewed as policy intent, not actual progress”). More countries adopted national AI strategies in 2024-2025 than in prior years, with the growth concentrated in emerging economies: new frameworks surfaced in sub-Saharan Africa (Ethiopia, Ghana, Nigeria), South/Central Asia (Sri Lanka, Nepal), and Latin America/Caribbean (Costa Rica, Jamaica), with Mexico and South Africa in development. High-income economies continue adding strategies too, but at a slower pace, often to align with EU AI Act requirements (e.g., Malta). The report frames the open challenge as implementation and regulatory capacity — “particularly in Africa, where many countries still lack formal strategies and risk falling behind in AI governance and readiness.”

8.3 AI Sovereignty — the New Five-Dimension Framework

Defined as a state’s capacity to “act deliberately and make independent decisions over the development, deployment, and governance of AI systems within its jurisdiction.” Drawing on Epoch AI, Zeki, and Brookings data across five dimensions:

1. Infrastructure sovereignty (state-owned/state-backed AI supercomputing, Epoch AI data): cumulative public/public-private AI supercomputers, 2010-2025 — China 85, Europe and Central Asia 44 (up from 3 in 2018, driven by the EuroHPC Joint Undertaking), North America 41 (nearly 7x growth, driven by the US NAIRR initiative), East Asia/Pacific excl. China 27, Latin America/Caribbean 8, Middle East/North Africa 3, South Asia 2. Caveat: this figure captures frontier-training-capacity clusters specifically, not total national compute capacity, and most large-scale compute globally remains privately owned — public-private lines are blurry (OpenAI’s Stargate project and Nvidia’s “AI Factory” model both extend sovereign compute buildouts through private-firm partnerships across many countries).

2. Data sovereignty (data-localization measures, cumulative 2000-2024, Ferracane et al. 2026): East Asia/Pacific leads with 77 measures, sub-Saharan Africa 71, Europe/Central Asia 66 — a “high-localization” cluster. Middle East/North Africa 44, Latin America/Caribbean 36, South Asia 24 — “moderate.” North America is the outlier at just 3, reflecting a long-standing “flow-first” policy orientation (the report notes US diplomats have recently pushed back against other countries’ data-sovereignty initiatives even as the US itself has added some restrictions on sensitive-data transfers to “countries of concern”). The steepest adoption rise across all regions began around 2016, coinciding with GDPR and the subsequent “Brussels Effect.”

3. Model sovereignty (cumulative AI model releases by region, 2018-2025, Epoch AI — the broader “all publicly reported models” dataset, not the narrower “notable models” dataset used in Chapter 1): United States 1,618 (up from 237 in 2018), China 849 (up from 151 in 2022 — more than quintupled in 3 years), Europe/Central Asia 666 (UK 229, France 141 as leading contributors), East Asia/Pacific excl. China 330, North America excl. US (i.e., mainly Canada) 125, Middle East/North Africa 74, South Asia 21 (largely India-driven), Latin America/Caribbean 2. The report explicitly flags these regional figures as conservative estimates, because “model documentation and reporting are less systematic in these regions” — and names a specific, concrete gap: “the growing ecosystem of smaller and language-specific models in sub-Saharan Africa, for example, are not represented at all,” citing AfriBERTa (Ogueji et al. 2021), AfriTeVa (Ogundepo et al. 2022), AfroLM (Dossou et al. 2022), EthioLLM and EthioMT (Tonja et al. 2024), and AfroXLMR (Alabi et al. 2022) as concrete named examples of models excluded from the count. See the Epoch AI Methodology section below for the mechanism behind this gap.

4. Application sovereignty (Brookings data): a heatmap of AI-application funding by country × domain (agriculture, business services, education, financial services, healthcare, security/defense, etc.). A small set of countries — US, China, UK, Germany, France — show high-intensity investment across nearly all categories; most other countries show concentrated, selective investment (e.g., Germany strong in industrial/manufacturing applications, Estonia in education technology, Israel in security/defense, South Africa in financial applications). The report frames this as the “least concentrated” sovereignty layer, giving smaller countries more room for niche specialization than the model or compute layers allow.

5. Talent sovereignty (workforce capacity + mobility): cross-border AI-talent circulation has slowed even where net flows remain stable — both inflows and outflows are declining across most regions (Zeki data), suggesting talent is increasingly staying within national/regional systems rather than circulating globally. The US remains the top global attractor of AI talent, though its lead is narrowing; India is transitioning from a net exporter to a net absorber of talent (a near-mirror-image of the US’s own inflow/outflow curve, reflecting the well-documented US-India talent pipeline). The Middle East/North Africa region is making incremental gains as new talent hubs emerge with targeted policy and investment.

On the “36-country comparison”: the drain question referenced a 36-country comparison in connection with AI sovereignty. This is a mis-association worth correcting: the 36-country panel is Stanford’s own Global AI Vibrancy Tool (an interactive companion product, not a static chapter table, ranking 36 countries across 42 indicators with user-adjustable weights) — noted on report page 5: “Compare the AI ecosystems of 36 countries. The Global AI Vibrancy tool will be updated by the end of 2026.” It is not tied to the sovereignty framework specifically; Chapter 8 treats AI sovereignty as a qualitative regional theme (the five dimensions above), with no fixed 36-country panel of its own.

OECD AI Index — the International Policy/Metrics Complement

This directly resolves the drain question on the OECD AI Index as a complement to HAI’s US-centric framing.

An OECD.AI Index literally exists as a distinct, citable publication: OECD, The OECD.AI Index (OECD Publishing, 2026), DOI 10.1787/32c01014-en, 56 pages, CC BY 4.0, published February 19, 2026 — five months before this drain pass, making it current. It should not be conflated with two adjacent OECD products:

  • OECD.AI Policy Observatory (oecd.ai, launched 2020) — the umbrella platform hosting the AI Principles, a live Policy Navigator, and an AI Incidents Monitor. The Index is built on top of this platform, not a synonym for it.
  • OECD AI Capability Indicators (beta, June 2025) — an entirely different product from a different OECD unit, scoring AI systems’ capabilities against human benchmarks (closer in spirit to HAI’s Chapter 2 than to HAI’s Chapter 8).

What the OECD.AI Index measures: a composite score tracking countries’ implementation of the OECD AI Recommendation (2019, updated 2024) across five national policy areas — R&D, Enabling Infrastructure, Policy Environment, Jobs and Skills, and International Co-operation — broken into 15 sub-components and 28 underlying indicators, reported on a normalized 0-1 scale (2023/2024 scores ranged 0.17-0.66, with the US, UK, and Switzerland among the top performers). It currently covers only the 38 OECD member countries — explicitly not the full 47 countries that have formally adhered to the OECD AI Recommendation, let alone all GPAI (Global Partnership on AI) members; broader coverage is stated as future work. Updated annually, with full scope reviewed every two years.

How it compares to HAI, in the OECD’s own words: the OECD.AI Index’s own technical paper (Annex A) explicitly benchmarks itself against the Stanford AI Index — describing HAI’s report as covering “over 30 countries” with no single composite score, versus its own approach (a genuine composite score across a fixed 38-country OECD panel) and Stanford’s separate Global AI Vibrancy Tool (36 countries, 42 indicators, user-adjustable weights, no OECD-Recommendation-alignment scoring). The two organizations are not competitors but collaborators: Stanford HAI is a named member of the OECD.AI Index’s own Expert Group.

Practical read: OECD.AI Index is the better source for government policy-implementation scoring against a formal international framework (28 indicators mapped to a treaty-like Recommendation); Stanford HAI Chapter 8 is the better source for qualitative sovereignty narrative and event timeline; the Global AI Vibrancy Tool is the better source for broad ecosystem comparison (talent, research, infrastructure, commercial activity together) without a policy-compliance framing. None of the three replaces the other two.

8.4 AI and Policymaking

Global legislative records (G20 countries, Digital Policy Alert data, enacted legislation only — not proposed/pending bills, and large omnibus bills with multiple AI provisions count as a single law, so volume likely understates actual policymaking activity): zero AI-related G20 laws on record in 2016; growth to a peak of 16 in 2024, then 11 in 2025 (cumulative 2016-2025: US 25, South Korea 17, France 10, Japan 10, Italy 9, UK 6, Germany 5, Russia 5, Brazil 3, Australia 2, Argentina/Canada/China/India 1 each). Note volume ≠ significance — “a single major law can carry more impact and enforcement weight than dozens of narrower ones.”

US Congressional hearings: AI-related witnesses in US congressional hearings grew from 5 (2017) to 102 (2025) — a 20x increase. Industry’s share of those witnesses nearly tripled, from 13% to 37% (now the largest single witness category), while academia’s share fell to 15%.

Public investment (contracts and grants, not private investment): US public AI investment totaled approximately **285.9B in US private AI investment in 2025 alone (Chapter 4). European public AI commitments reached approximately 1.6B, Germany 320M); European public spending is accelerating — in 2024 alone the UK committed 206.6M (40% of its decade total).

Epoch AI Methodology — Resolving the Non-English Coverage Question

This directly resolves the drain question on whether Epoch AI’s “notable models” list under-covers non-English-language sources.

Epoch AI (epochai.org) is the data partner behind this chapter’s model-sovereignty figures and Chapter 1’s “notable models” dataset. Two distinct questions, both now answerable from primary sources:

How to query Epoch AI’s data directly (not just wait for HAI’s annual snapshot)

There is no traditional REST API. Access surfaces, all live-verified:

  • Interactive dashboard: epoch.ai/trends — headline compute/growth-rate charts, no direct download.
  • AI Models explorer: epoch.ai/data/ai-models — filterable by compute, parameters, cost, and date, colorable by country/organization/domain. Append ?subset=notable&view=table to filter to the same “notable models” subset HAI Chapter 1 uses.
  • Adjacent trackers: epoch.ai/data/ai-data-centers, /ai-chip-sales, /machine-learning-hardware, /gpu-clusters.
  • Direct CSV downloads (CC-BY licensed, synced daily): epoch.ai/data/notable_ai_models.csv, frontier_ai_models.csv, large_scale_ai_models.csv, all_ai_models.csv.
  • Programmatic client: pip install epochai (github.com/epoch-research/epochai-python) — reads Epoch’s underlying Airtable base directly; requires copying their Airtable base and generating an Airtable personal access token. This is the closest thing to a formal API.
  • Full methodology documentation hub: epoch.ai/data/ai-models-documentation (sub-pages: Inclusion criteria, Records, Database updates, Estimation methods, Changelog, Downloads, Acknowledgements).

Non-English / geographic coverage — what Epoch AI says about its own gap

Epoch AI’s formal inclusion-criteria page (epoch.ai/data/ai-models-documentation/inclusion) defines “notable” as meeting at least one of: >5,000 citations, >$1M training cost (or ≥1% of the priciest model to date), >1M monthly active users, state-of-the-art benchmark performance, equivalent historical significance, or Epoch-staff discretion. This formal page is silent on geography or language — its “coverage” caveats only address temporal (pre/post-2010) and domain (audio vs. language/vision) gaps, not linguistic ones.

However, Epoch AI itself has explicitly and repeatedly acknowledged the language/geography gap in its own blog — first-party, not extracted under pressure from a critic:

  • “Tracking large-scale AI models” (Apr 2024): “systems developed in regions where languages other than English are predominant may not receive enough coverage in English-language media to come to our attention unless we seek them out.” Epoch describes actively searching across 15 non-English languages to mitigate this, names Chinese-language sources (via Baidu search) as their single most fruitful non-English discovery channel, and concludes candidly: “we don’t know how many such models have flown under the radar” — citing lab secrecy as a compounding factor.
  • “Chinese language models have scaled up more slowly…” (Jan 2025): “it is more difficult for the team to discover models through Chinese language documents and websites… coverage may still be worse for these models,” though Epoch believes “the issue is limited when tracking frontier models” specifically (i.e., the gap is more likely to affect mid-tier/regional models than the highest-compute frontier systems, which tend to get international press coverage regardless of source language).

No substantive third-party critique specifically targeting Epoch AI’s language-coverage methodology was found (a search of AI-safety/AI-bias commentary venues turned up nothing on-point) — the acknowledgment currently comes entirely from Epoch AI itself. Combined with the Stanford HAI report’s own concrete example (sub-Saharan African language models “not represented at all” in the Model Sovereignty section above), the honest summary is: this is a known, self-acknowledged limitation of the primary compute/model-tracking data source used across the entire HAI report, most pronounced for regional and non-frontier models, and not yet independently quantified by anyone (including Epoch AI itself).

Try It

  • For “is the US deregulating AI” claims: cite the specific tension, not a blanket claim — federal executive branch (Jan 23 + Dec 12 EOs) pushed to limit state regulation, while the Senate (Jul 1) explicitly preserved states’ regulatory authority and California/Texas/Montana/Utah all passed substantive state laws in the same year.
  • For sovereignty-strategy conversations: use the five-dimension framework (infrastructure/data/model/application/talent) to diagnose which layer a client or region is actually strong or weak in — “AI sovereignty” as a single word flattens five very different policy levers.
  • For any claim built on Epoch AI model counts (frontier-model leaderboards, “who’s ahead” narratives): caveat explicitly that non-English, non-frontier, and especially African-language models are under-counted — this is Epoch’s own admission, not a hostile read.
  • For querying Epoch data directly instead of waiting for the next annual report: start with epoch.ai/data/ai-models?subset=notable&view=table for a browsable notable-models view, or pull notable_ai_models.csv directly for analysis.
  • For EU-sovereignty or international-policy-benchmarking work: pull the OECD.AI Index’s 28-indicator, 38-country composite score (10.1787/32c01014-en) rather than trying to force HAI’s narrative sovereignty chapter into a scoring framework it doesn’t provide.

Open Questions

  • Epoch AI has not published a quantified estimate of how many non-English-language models are missing from its notable-models dataset — only qualitative acknowledgment. (researched 2026-07-02: still open — no primary or third-party source quantifies the gap; Epoch AI’s own language is “we don’t know how many.“)
  • The Global AI Vibrancy Tool’s full 36-country dataset was not independently queried for this article (report page 5 confirms it exists but says it “will be updated by the end of 2026” — the current live state was not verified).
  • The OECD.AI Index’s coverage is currently capped at the 38 OECD member countries, explicitly excluding the broader 47-country set that has formally adhered to the OECD AI Recommendation. Whether/when that gap closes is stated as “future work” in the OECD’s own technical paper, with no committed date found.

Graph View

Backlinks