The Think Tank’s working file on artificial intelligence and sustainable development: an annotated register of the instruments that matter, a searchable catalogue of all 171, and a searchable concordance of 4.07 million words of report text, and a macro reading of how the conversation moved from catastrophic risk to compute, capability and sovereignty.
This page merges the two working documents the Think Tank has been using: the annotated Register, which explains what each instrument is and why it carries weight, and the Observatory, which lets you search the full catalogue and read the macro analysis. The library holds 171 documents: a curated core of 104 strategic instruments, the complete OECD Artificial Intelligence Papers series (64), and three Hiroshima-Process transparency filings. The macro analysis is computed on the core 104 — the two reference shelves are searchable but would swamp the frame and concept measures.
Eight thematic sections written to be read in order, each entry with two or three sentences on what changed and who uses it, then a chronological timeline of how the architecture was built.
All 171 documents with live search and filters — collection, instrument status, document type, source entity typology, sector and year — each linked to its official source.
152 report covers, filterable by document type, issuing institution, country or bloc of origin, sector and year.
The main ideas behind the executive summaries of the twenty most consequential reports, each with its figure and its unresolved question.
Creates legal obligations. Six of the 171 qualify: the EU AI Act and its 2026 amendment, the Council of Europe convention, and three national statutes.
Politically binding or standard-setting — resolutions, recommendations, declarations, codes. Enforced by peer pressure.
Scientific or expert assessment commissioned to inform governments. Non-prescriptive by design.
Analytical reports, indices and research that supply the numbers policymakers argue with.
If you read nothing else, these twelve give you the architecture, the law, the science and the development case. They are sequenced so each one makes sense of the next.
UN High-Level Advisory Body, Sept 2024. The blueprint everything else was built from.
A/RES/79/1, Sept 2024. Turned that blueprint into an intergovernmental mandate.
Aug 2025. The operative constitution of UN AI governance.
July 2026. The newest and most consequential document on this page.
2021. Still the only text all 194 states adopted.
Reg. (EU) 2024/1689. The world's only comprehensive binding AI law — with extraterritorial reach.
CETS 225, 2024. The first binding international AI treaty.
2025 and 2026 editions, chaired by Yoshua Bengio. The shared technical evidence base.
Aug 2026. The development argument, with the adopt–adapt–advance frame.
UNDP. "A matter of choice" — the human-agency counter-narrative.
The equity and concentration case, with the numbers the Global South cites.
2025, updated April 2026. Sets the terms of the entire AI-and-climate debate.
For a foundation working on sustainable development, the whole file resolves into three usable positions.
Sixty-five instruments in eight thematic sections, written to be read in order. Each entry says what the document is, why it carries weight, and where to read it; items that could not be verified to publication standard carry a flag rather than being silently included.
Note. The register covers the sixty-five documents of the original reading brief in depth. The Library tab holds all 171 — the core corpus plus the full OECD AI Papers series and the Hiroshima transparency filings — with shorter notes.
In four years the UN went from having no AI mandate at all to running a standing scientific panel and an annual intergovernmental dialogue. Nothing in this section is legally binding — that is the defining feature, not an omission.
UN General Assembly · Summit of the Future
The mandating instrument for everything that followed. Objective 5 committed Member States to establish an Independent International Scientific Panel on AI and to launch a Global Dialogue on AI Governance. Every subsequent UN AI text cross-references it.
UN General Assembly, 79th session
Fixes the Panel at 40 members on three-year terms, confines it to the non-military domain, and requires one annual evidence-based report plus thematic briefs. It sets the Dialogue up as an annual non-negotiating forum concluding in a co-chairs' summary rather than a negotiated outcome — the single most important design choice to understand about its leverage.
UN Independent International Scientific Panel on AI · co-chairs Yoshua Bengio and Maria Ressa
The first universal-membership scientific assessment of AI — the IPCC analogue the Advisory Body argued for. Headline finding: current safeguards cannot keep pace with capability growth, and the evidence policymakers need may arrive too late to act on. Structured across seven domains including economics, human rights and information integrity, and cultural flourishing.
This is the designated common evidence base for the Global Dialogue. Anything your foundation publishes on AI governance from here should engage with it.
UN General Assembly · hosted at Palexpo, Geneva, back-to-back with AI for Good
Roughly 170 Member States and 4,200+ registered participants, co-chaired by El Salvador and Estonia. Four thematic clusters: opportunities and implications; bridging AI divides through capacity-building and access; safe, secure and trustworthy AI; and human rights, transparency and accountability. It is now the only annual venue where all Member States, industry and civil society sit at the same AI table.
Check before citing: an official co-chairs' summary had not appeared on the resources page at time of compilation. Next session: 3–4 May 2027, New York.UNESCO General Conference
Predates the whole window and still the only universally adopted global normative framework on AI — which is why the 2024–26 UN architecture defers to it rather than duplicating it. It carries a monitoring and evaluation obligation, and that clause is what produced the implementation toolkit below.
UNESCO · with the Global AI Ethics and Governance Observatory
The operational bridge from principle to national policy, and UNESCO's most-used AI product. It scores a country's legal, social, economic, educational, scientific and technological readiness and recommends policy. Published country cases include Ghana, Ukraine, Colombia, the Philippines, Nigeria and Thailand.
For comparative work on developing-country AI governance capacity, this is the best available baseline. The Observatory also hosts the Global Network of AI Supervisory Authorities — the closest thing the UN system has to a regulator-to-regulator channel.
UN General Assembly, 78th session · US-led, 120+ co-sponsors
The first-ever UNGA resolution on AI, and the moment "safe, secure and trustworthy AI" became standard UN language. It frames AI through the 2030 Agenda and the SDGs rather than through security, and deliberately excludes military AI.
UN General Assembly, 78th session · China-led, 140+ co-sponsors
The counterweight to A/RES/78/265. It reframes the agenda around the capacity gap — compute, data, talent, infrastructure — rather than around risk. Together the two resolutions define the axis every subsequent UN AI negotiation runs along, and the Geneva Dialogue's four clusters are a direct compromise between them.
UN Secretary-General, report to the General Assembly
The money question behind the capacity-building resolution. Proposes a Global Fund for AI capitalised at US$1–3 billion over 2–4 years, financed partly through a 0.01–0.05% contribution on relevant technology transactions, and disbursed via performance-based grants, AI development bonds and debt-for-AI swaps.
Status: recommended, not operational. No evidence of capitalisation as of September 2026 — do not describe "the UN AI fund" as existing.UN Secretary-General's High-Level Advisory Body on AI (39 members)
Seven recommendations for a "globally inclusive and distributed architecture" built from light institutional mechanisms rather than a new treaty agency. Two were adopted almost verbatim — the scientific panel and the policy dialogue. Five were not: a capacity development network, a standards exchange, a global fund, a global AI data framework, and a UN AI office.
Tracking which recommendations were dropped is the cleanest way to measure the gap between expert ambition and intergovernmental appetite.
ITU, on behalf of the UN system
The only comprehensive map of what the UN system is actually doing on AI: 53 entities and 729 projects, indexed against the SDGs. The practical starting point for identifying implementing partners in a given SDG domain. See also the AI for Good Impact Report, which maps AI contributions goal by goal, and the AI Standards Exchange Database indexing 700+ AI standards.
Bletchley → Seoul → Paris → New Delhi. In three years the track changed purpose entirely: from catastrophic risk, to innovation and inclusion, to action and sustainability, to development and diffusion. Geneva 2027 is where it either merges with the UN process or formally diverges from it.
Independent expert panel chaired by Yoshua Bengio · 30+ countries, EU, OECD, UN · secretariat at the UK AI Security Institute
The only artefact that has survived every change of summit host and political mood since Bletchley, and the most authoritative shared evidence base in the field. Its "evidence dilemma" framing — act without evidence and risk ineffectiveness, wait for evidence and risk being too late — is now standard in national policy documents.
The 2026 edition deliberately narrowed to frontier capability risks, dropping the 2025 treatment of bias, privacy and environmental impact to avoid duplicating the new UN Panel. That narrowing is itself a governance signal about how the division of labour is settling.
India AI Impact Summit, Bharat Mandapam, New Delhi · chaired by India (MeitY)
The decisive inflection point. Delhi completed the pivot from risk mitigation to diffusion and development, with the broadest coalition the track has assembled — and, critically, a single text both Washington and Beijing signed. Seven pillars under a People / Planet / Progress frame, including democratising AI resources and human capital development.
Breadth was bought at the cost of depth: everything is voluntary, and frontier-risk language is largely displaced into subsidiary workstreams. Accompanying instruments include the New Delhi Frontier AI Impact Commitments (usage-data sharing and underrepresented-language datasets), a Charter for the Democratic Diffusion of AI (22 countries) and reskilling principles (23 countries).
Endorsement count varies by source: 88 states at adoption, ~92 including international organisations. State the basis when you cite it.AI Action Summit participants, Paris
The first visible fracture in the track: the United States and United Kingdom declined to sign. The rebranding from "Safety" to "Action" signalled the shift toward opportunity framing that Delhi then completed. The companion Paris Charter on AI in the Public Interest drew only ten signatories but launched Current AI, a public-interest foundation with US$400m committed against a US$2.5bn five-year target.
UK DSIT and Republic of Korea, co-chairs
Where the track's now-standard three baskets — safety, innovation, inclusivity — were fixed. The annexed Statement of Intent toward International Cooperation on AI Safety Science is the more consequential half: it founded the network of national technical institutes, the only durable state capacity the summit series has produced.
UK Government, on behalf of attending states
The founding text, and the first document in which the US, China and the EU jointly acknowledged frontier-AI catastrophic risk. Substantively thin, but it created the venue everything since has run through and mandated the state-of-the-science report that became the International AI Safety Report.
UK DSIT, agreed by frontier AI developers
The first international commitment requiring frontier developers to publish safety frameworks with defined risk thresholds and stated responses when thresholds are crossed. Twelve companies published or updated such frameworks during 2025 — arguably the most materially consequential output of the whole summit series, despite being unverified by any third party.
OECD · 47 adherents
The only intergovernmental standard, as distinct from declaration, in this landscape, and the substrate for the G7 Hiroshima Process, the EU AI Act's terminology and the UN's own definitional work. The 2023 amendment to the definition of "AI system" is the reason that term now means roughly the same thing across the EU, US and Council of Europe instruments — an under-appreciated piece of quiet harmonisation. The May 2024 update added privacy, IP, safety and information integrity in light of generative AI.
Formerly the International Network of AI Safety Institutes (launched San Francisco, Nov 2024)
The removal of "safety" from the network's name — following the same excision at the UK institute (now AI Security Institute) and the US one (now CAISI) — is the clearest single indicator of how far the political framing has moved since Bletchley. Members include Australia, Canada, the EU, France, Japan, Kenya, Korea, Singapore, the UK and the US.
OECD Publishing
The Capability Indicators score AI against nine human ability domains on graduated scales — the missing measurement layer beneath every "AI and jobs" debate, designed to plug into education and skills policy. The OECD.AI Index (Feb 2026) is the first instrument that measures whether adherents actually implement the AI Principles rather than merely endorse them.
This is where obligations actually live. Three structural shifts define the period and all three crystallised in 2026: the EU delayed its high-risk tier but kept its architecture; the US turned federal power against state AI law; and the forum landscape split, with a China-headquartered intergovernmental body now in existence.
European Parliament and Council
Still the only comprehensive binding horizontal AI law in force, and extraterritorial by design: Article 2 binds non-EU providers placing systems on the EU market and providers whose system output is used in the EU. Penalties reach €35m or 7% of worldwide turnover. That reach is why it functions as the global compliance baseline.
Phasing: prohibited practices and AI literacy from 2 Feb 2025; general-purpose AI model obligations from 2 Aug 2025; general application including transparency duties from 2 Aug 2026; high-risk obligations now deferred to 2 Dec 2027 (Annex III) and 2 Aug 2028 (Annex I products).
European Parliament and Council · from the Commission's November 2025 simplification package
The most consequential regulatory development of 2026. It defers the high-risk tier by roughly 16 months to two years, narrows the "safety component" definition, widens a GDPR basis for bias-detection processing, adds SME simplifications, centralises supervisory powers in the AI Office, and adds new prohibitions on non-consensual intimate imagery and CSAM. Read it as partial deregulation under competitiveness pressure — the prohibitions, GPAI rules and transparency duties were not rolled back. The Brussels Effect is delayed, not reversed.
Council of Europe
The first legally binding international AI treaty, and the only cross-bloc instrument the United States signed. Around 19 signatories plus the EU, which ratified in May 2026. Its reach runs through State obligations over public-sector AI with a softer regime for private actors, so it sets a floor rather than a market-access barrier. Its strategic value now is as the normative bridge between the EU regime and non-EU democracies.
Verify the live ratification count and entry-into-force status on the CoE Treaty Office chart before citing — entry into force requires five ratifications including three CoE member states.The White House
The United States has no binding horizontal AI statute; federal policy is now aimed at suppressing subnational regulation. The order creates a DOJ AI Litigation Task Force to challenge state AI laws, a Commerce review of "onerous" state rules, funding conditionality, and an FCC proceeding on preemptive federal reporting standards. A National AI Legislative Framework followed in March 2026.
Preceding steps: EO 14179 (Jan 2025) rescinding the Biden-era EO 14110, and Winning the AI Race: America's AI Action Plan (July 2025). Transatlantic divergence is now structural rather than cyclical.
US National Institute of Standards and Technology
Voluntary, but the most widely adopted technical risk vocabulary in the world — Govern / Map / Measure / Manage now appears in contracts, insurance terms and non-US regimes. It survived the change of administration intact, which makes it the most durable US contribution to global AI governance.
Founding agreement signed in Shanghai; headquartered there
China's proposed body, floated in the July 2025 Global AI Governance Action Plan, became an actual intergovernmental organisation in July 2026 with 29 founding signatories and the United States absent. It gives the Global South an alternative venue to OECD- and G7-centred processes. Watch for membership growth and any standard-setting mandate overlapping ITU, ISO and IEC.
China's binding domestic rules are also worth tracking: the Interim Measures for Generative AI Services (effective Aug 2023, the first such national rules anywhere), the AI Safety Governance Framework (v2.0, Sept 2025) and labelling requirements for AI-generated synthetic content (effective Sept 2025).
G7 Leaders · reporting framework hosted by OECD
The main voluntary bridge between EU and US/Japanese approaches, and a direct input into the structure of the EU's GPAI Code of Practice. The OECD reporting framework (Feb 2025) is the first standing international transparency mechanism for frontier developers — self-reported and unverified, but the emerging locus of soft accountability.
African Union Commission
Development-first rather than risk-first, and deliberately positioned against importing the EU model wholesale. Binding continental rules remain absent; in practice governance leverage runs through the Malabo Convention on data protection and national DPAs. Strategically the swing bloc courted by both the EU and the new Shanghai organisation. Follow-on: the Africa Declaration on Artificial Intelligence adopted at the Global AI Summit on Africa, Kigali, April 2025.
Government of the Republic of Korea
The world's second comprehensive binding AI law. Risk-based like the EU but lighter-touch, with "high-impact AI" duties, generative-AI labelling, and a domestic representative requirement for large foreign providers that gives it real extraterritorial bite in Asia. Its timing matters: in force January 2026, while the EU deferred its high-risk tier to December 2027, which briefly makes Korea the most demanding operative high-risk regime anywhere.
ASEAN · Japan · India · Latin America · Gulf
Expanded ASEAN Guide on AI Governance and Ethics — Generative AI (Jan 2025), voluntary and business-facing. Japan's AI Promotion Act (May 2025) — binding but with no penalties or prohibitions, the clearest innovation-first counter-model. India AI Governance Guidelines (Nov 2025), voluntary and techno-legal, relying on existing law. In Latin America, the UNESCO ministerial series (Santiago 2023, Montevideo 2024, Santo Domingo) plus the separate Cartagena de Indias Declaration of 17 countries (Aug 2024). In the Gulf, Saudi Arabia's NSDAI and the UAE's National AI Strategy 2031 are capability- and investment-led rather than rules-led.
There is no single "CELAC AI declaration" — two distinct Latin American tracks are routinely conflated.The evidence base policymakers actually cite. Note the tension running through it: the IMF finds rich countries most exposed, the World Bank finds them most at risk of automation and best placed to gain, and the ILO–World Bank work finds developing countries can get the disruption without the dividend.
World Bank Group
The newest and most important item in this section. Its adopt–adapt–advance sequencing is the argument that developing countries need not build frontier models to capture AI gains. Jobs in high-income countries are over three times more likely to face automation risk (14.2%) than in low- and middle-income countries (4.5%) — but 16.2% of jobs in developing economies could see meaningful AI productivity gains versus 18.7% in wealthy ones. Low exposure to displacement also means low exposure to upside.
UNDP, Human Development Report Office
UNDP's flagship devoted entirely to AI, published against the finding that human development progress has slowed to a 35-year low, with the high–low HDI gap widening for a fourth consecutive year. Its argument — that AI outcomes are a matter of political choice, not technological destiny — is the most authoritative development-side counter to both techno-optimism and doom framing, and it carries weight with finance ministries.
The accompanying Global Survey on AI and Human Development is the most-quoted attitudinal dataset in the field: 60% of respondents expect AI to create job opportunities; in lower-income countries 70% expect productivity gains.
HDR 2026 is Towards Peace with Nature — not an AI report. HDR 2025 remains UNDP's standing AI reference.UN Trade and Development (UNCTAD)
Supplies the two numbers most used in Global South AI advocacy: an AI market reaching US$4.8 trillion by 2033, and up to 40% of global jobs affected. Its equity finding is the sharper one — 118 countries, overwhelmingly from the Global South, are absent from international AI governance discussions altogether, and 100 firms, mostly US and Chinese, account for 40% of world private R&D. The shared global compute facility proposal originates here.
There is no UNCTAD Digital Economy Report 2025 or 2026. The most recent DER is 2024, and it is about environmental sustainability of digitalisation, not AI.IMF Staff Discussion Note · Cazzaniga, Jaumotte, Li et al.
Source of the single most quoted statistic in the field: almost 40% of global employment is exposed to AI — around 60% in advanced economies, 40% in emerging markets and 26% in low-income countries. Its distributional finding, that women and college-educated workers face higher exposure but are better placed to benefit, underpins most labour-transition policy language. The companion AI Preparedness Index covers 174 economies and appears constantly in country diagnostics.
ILO · Gmyrek, Berg et al.
The current authoritative occupational-exposure index, built on around 30,000 tasks at six-digit occupational level. One in four workers globally is in an occupation with some generative-AI exposure; the highest-exposure band covers 3.3% of global employment, ranging from 11% of total employment in low-income countries to 34% in high-income ones. The gender asymmetry is the most-cited element — 9.6% of female versus 3.5% of male employment in the top band in high-income countries, concentrated in clerical work.
ILO's "transformation, not destruction" framing is the standard counterweight to alarmist displacement estimates.
ILO and World Bank joint working paper · background to WDR 2026
The sharpest statement of the equity problem and, for a development foundation, probably the most useful single paper here. In developing countries the workers most vulnerable to displacement often already have digital access, while those positioned to gain productivity frequently lack reliable connectivity — so disruption can arrive before the dividend. The strongest evidence-based rebuttal to "developing countries are insulated by low exposure."
World Bank
The best single source of hard divide statistics, structured around connectivity, compute, context and competency. High-income countries, with 17% of world population, hold 87% of notable AI models, 86% of AI start-ups and 91% of venture funding. Data-centre capacity splits 77% high-income, 18% upper-middle, 5% lower-middle and under 0.1% low-income. Fewer than 5% of people in low-income countries have basic digital skills, against 66% in high-income.
Counterweight worth quoting alongside: over 40% of ChatGPT traffic in mid-2025 came from middle-income countries, and generative-AI job postings rose ninefold from 2021 to 2024.
World Economic Forum · 1,000+ employers, 14m workers, 55 economies
The employer-survey counterpart to the modelling studies. By 2030: 170 million new jobs created (14% of current employment) and 92 million displaced (8%), for net growth of 78 million. 39% of workers' existing skill sets will be transformed or outdated between 2025 and 2030. Cited mainly for skills and reskilling policy rather than macro effects.
There is no Future of Jobs Report 2026 — the series is biennial and the next edition is due 2027. Blog posts circulating under that title recycle the 2025 figures.Oxford Internet Institute · PI Vili Lehdonvirta
Origin of the term "compute deserts", now standard in Global South AI advocacy. GPUs are concentrated in roughly 30 countries; Amazon, Microsoft and Google hold about 70% of the global cloud infrastructure market; Amazon's high-performance GPU instances run from only around seven locations worldwide, forcing most national research communities to move data across borders to train models. The empirical backbone for sovereign-compute and shared-facility proposals.
Pair with the Tony Blair Institute's Powering AI in the Global South (Dec 2024), which links the compute divide to the energy divide most compute analyses omit.
Current AI · UAE · India AI Impact Summit
Three vehicles worth distinguishing carefully. Current AI — US$400m committed, US$2.5bn five-year target, repositioned in mid-2026 from operator to movement infrastructure. The UAE AI for Development Initiative — US$1bn dedicated exclusively to African countries, announced November 2025 and the largest single sovereign AI-for-development commitment to date. And the US$200bn+ announced around the Delhi summit.
The Delhi US$200bn is overwhelmingly private Indian and US corporate capital expenditure — Reliance, Adani, Google — not concessional development finance and not for the Global South broadly. Do not present it as such.The fastest-moving section, and the one where the numbers have shifted most. In April 2025 the IEA put data centres at 415 TWh; in April 2026 it recorded 485 TWh for 2025 and projected roughly 950 TWh by 2030 — about 3% of global electricity.
International Energy Agency
The reference document that set the terms of the entire 2025–26 debate. Data centres consumed about 415 TWh in 2024, roughly 1.5% of global electricity, growing 12% a year since 2017 — over four times the rate of total electricity demand. Concentration: US 45%, China 25%, Europe 15%. Data centres account for about a tenth of global electricity demand growth to 2030, and over 20% in advanced economies. Emissions reach 300–500 Mt by 2035, still under 1.5% of energy-sector emissions.
The April 2026 update raises the near-term trajectory without moving the 2030 endpoint much: 485 TWh actual for 2025, +17% growth in one year (+50% for AI-focused facilities), ~950 TWh by 2030. Its framing of AI as both demand driver and grid-optimisation tool is the structure most official reports now adopt.
ITU Telecommunication Standardization Sector, Study Group 5
Strategically the most consequential item in this section after the IEA reports, and the one most likely to be missed. The first international standard providing lifecycle-assessment guidelines for AI systems, mapping eight AI lifecycle phases onto four LCA stages, mandating climate-impact assessment, requiring PUE and WUE data-centre reporting, and specifying how training impacts are allocated. It converts the recurring "we lack standardised metrics" recommendation into an actual normative instrument regulators can cite in disclosure mandates.
ITU Telecommunication Development Bureau with the World Benchmarking Alliance
The only annual, company-level, UN-issued emissions series for the digital sector — the natural evidence base for a disclosure-mandate argument. The 2026 edition puts operational emissions of 200 digital companies at 301 Mt CO₂e in 2024, about 0.8% of global energy-related emissions, with the top ten firms accounting for 54% of reported electricity use. Some AI and cloud providers reported 2024 emissions at up to 239% of 2020 levels while telecom operators collectively cut emissions 11%. Only 25 of 200 companies reached 100% renewable electricity.
Two cautions. The widely circulated "150% increase" figure is from the 2025 edition and covers 2020–2023 (Amazon +182%, Microsoft +155%, Meta +145%, Alphabet +138%) — it is routinely misattributed to the 2026 report. And the 2025 and 2026 electricity totals (581 TWh vs 494 TWh) are not comparable as reported; check the methodology before using them as a time series.UN Environment Programme
The document that put AI's footprint on the UN environmental agenda and seeded the Paris coalition. Four recommendations: standardised measurement, mandatory disclosure, efficiency plus water and component recycling, and renewable sourcing. Its public-facing companion carries the campaign figures now in wide circulation — data centres growing from 500,000 in 2012 to 8 million, and 800 kg of raw materials to produce a 2 kg computer.
The "a ChatGPT request uses ten times the electricity of a Google search" estimate originates in this campaign material and is materially undercut by Google's own 2025 measurement below. Attribute it carefully.Convened by France with UNEP and ITU
37 technology companies, 11 countries and 5 international organisations. Workstreams on standardised measurement methods, lifecycle analysis frameworks for disclosure, research prioritisation, and steering AI toward climate, pollution, biodiversity and ocean objectives. The institutional bridge between UNEP's 2024 issue note and ITU-T L.1801, and the main vehicle carrying the sustainability agenda between summits. Voluntary, which is the obvious critique.
Sasha Luccioni (Hugging Face), Boris Gamazaychikov (Salesforce), Scott Chamberlin (Neuralwatt)
An appliance-style efficiency rating for AI models — the most policy-legible instrument in this space. The v2 headline is strategically important because inference-time compute scaling is the dominant industry direction: reasoning-enabled models use on average around 30 times more energy than non-reasoning variants, with 150–700× increases in specific same-model comparisons. And 8 of 14 comparable newer models used more energy than similarly sized models from ten months earlier — newer is not automatically more efficient.
Google Cloud / Google DeepMind
The first per-prompt footprint disclosure by a named frontier lab: a median Gemini text prompt at 0.24 Wh, 0.03 gCO₂e and 0.26 mL of water, with a claimed 33× energy and 44× carbon reduction per prompt over twelve months. A genuine transparency first — but the boundary conditions are the company's own, the figures are medians rather than means, and independent commentary questions whether upstream impacts are in scope. Cite as a disclosed corporate figure, not an independently verified one.
Li, Yang, Islam and Ren
The paper that created water as a distinct AI policy issue, introducing the operational/off-site water distinction and the methodology now used across the literature, including by UNEP. Two companions worth having: Luccioni, Jernite and Strubell's Power Hungry Processing (FAccT 2024), the standard citation for inference dominating lifetime energy at scale; and Health-Informed Computing, which projects the annual US public-health burden of data centres above US$20bn in 2028 and reframes the footprint as an environmental-justice and siting issue.
Where AI policy meets specific SDGs. Note the common diagnosis across WHO, UNESCO and FAO: the binding constraint in low-resource settings is institutional capacity, data governance and literacy — not algorithmic capability.
World Health Organization
The first global normative guidance specifically on generative systems in health, issued roughly a year after ChatGPT's release. Over 40 recommendations across governments, developers and providers, covering five application areas from diagnosis to scientific research, and risks including automation bias and the concentration of development in a handful of firms. The most operationally cited AI-in-health instrument globally and the template for national health-AI guidance.
It builds on rather than replaces WHO's 2021 six guiding principles, which remain the reference frame.
WHO Executive Board, 158th session
Worth knowing chiefly for what it establishes about the state of play: as of September 2026 there is no dedicated World Health Assembly resolution on AI. WHA78 extended the Global Strategy on Digital Health to 2027; the successor strategy for 2028–2033 — the vehicle that will carry AI governance — goes to WHA80 in 2027. That is the date to watch. See also the recent WHO/Europe Knowledge Community report (Sept 2026), which documents implementation failure modes rather than restating principles.
UNESCO · Miao and Holmes
The first global guidance on generative AI in education, and still the reference instrument. Its central diagnostic — that iterative model releases are outpacing the adaptation of national regulatory frameworks — has aged well. Headline recommendations: mandatory data-privacy protection and a minimum age for independent generative-AI use, which UNESCO proposed at 13.
No revised second edition could be confirmed despite databases citing "UNESCO (2025)". Cite the 2023 original.UNESCO
This — not a revised guidance document — is the substantive post-2023 UNESCO education update. The first global frameworks defining the AI knowledge, skills and values students should acquire and teachers need to teach ethically, structured around a human-centred mindset, ethics of AI, techniques and applications, and system design, across three progression levels. It shifts UNESCO from cautioning about generative AI to specifying curriculum content, and it is what education ministries are actually operationalising.
UNESCO
The landmark sceptical assessment: evidence of educational technology's learning benefits is thin, often industry-generated, and much deployment is driven by commercial rather than pedagogical logic. The most important counterweight in the UN system to techno-optimist AI-in-education framing, and it visibly shaped the tone of the 2023 generative-AI guidance.
FAO with Wageningen Food Safety Research
The most substantive FAO AI output identified, and strategic in a specific way: it finds food-safety authorities, particularly in resource-limited countries, are constrained by data scarcity and institutional capacity rather than by algorithmic capability — the same diagnosis WHO/Europe reaches for health. Include if the brief needs SDG 2 coverage.
Weather, climate services and disaster risk
WMO maintains an active AI programme covering forecasting and Early Warnings for All, but no flagship WMO AI report comparable to the IEA or WHO items exists in this window. Its work is programmatic rather than report-based. Flagged here as a deliberate omission rather than an oversight — and as an open niche if your foundation is looking for underserved terrain.
Two findings recur across independent corpora and are the most useful things this literature offers a development organisation: the research skews heavily toward environmental goals over social ones, and the authorship skews heavily away from the Global South.
Vinuesa, Azizpour, Leite, Dignum, Tegmark, Fuso Nerini et al.
Predates the window but is the single most-cited paper in this literature and the origin of the enable/inhibit framing everything since has adopted. Expert elicitation across all 169 SDG targets: AI may enable 134 targets (79%) and inhibit 59 (35%), with enabling effects concentrated in environmental goals and inhibiting effects where AI can entrench inequality or undermine governance.
Whenever you cite it, state two things: it predates the transformer era entirely, and its method is structured expert judgement, not measured outcomes.Gohr et al. · Nature Sustainability
The strongest recent empirical evidence base in this section. Systematic analysis of 792 peer-reviewed articles applying AI to the SDGs finds that very few studies combine advanced AI capability with deep sustainability expertise, and that applications cluster heavily on environmental goals while poverty reduction and gender equality are markedly underserved. Its conclusion — that AI's potential in sustainable development "remains to be realised" — is directly usable as an argument for redirecting research funding.
Independently corroborated on a different corpus by a 2024 systematic review of 191 articles: 74% address environment and climate action, with Africa, Oceania and South America markedly underrepresented among authors.
Nasir, Javed, Gupta, Vinuesa and Qadir · Societal Impacts
Co-authored by Vinuesa himself and explicitly a critical revisiting of the 2020 assessment in light of generative AI. The natural anchor citation for "what has changed since 2020".
Newer, low-profile journal; obtain the full text before relying on its specific conclusions.Adams, Adeleke, Junck, Alayande et al. · Nature Computational Science
The most authoritative Global-South-authored contribution in this literature, drawing on researchers based in low- and middle-income countries. It finds real potential to accelerate progress on some SDGs but identifies substantial barriers to building locally appropriate tools, and argues rights-based safeguards must be grounded in the lived experience of local projects. Strategically valuable because it supplies evidence from exactly the settings the Gohr review identifies as the literature's biggest gap.
Feng, Li, Fu, Guo, Liu and Bengio · Nature Machine Intelligence
Short-format but weighty, both for its author list and because it explicitly integrates the environmental-footprint literature into the AI-for-SDGs literature — the two strands usually treated separately. Identifies three domains where AI's expansion directly conflicts with the SDGs: energy and e-waste (with generative-AI e-waste potentially reaching five million tonnes between 2020 and 2030 against a formal recycling rate below a quarter), misinformation, and algorithmic opacity undermining accountability under SDG 16.
Recurring publications you can build a monitoring calendar around. Each has a known release month, which makes them useful as scheduled inputs rather than one-off citations.
Stanford Institute for Human-Centered AI
The comprehensive annual barometer across R&D, technical performance, responsible AI, economy, science, medicine, education, policy and public opinion. Reported 2026 headlines: the US–China frontier performance gap has essentially closed; US private AI investment of US$285.9bn in 2025, some 23 times China's; generative AI reaching 53% population adoption within three years, faster than PCs or the internet; and a striking trust divide, with 23% of the general public viewing AI's labour-market impact positively against 73% of experts.
Verify 2026 figures against the report PDF — the landing page did not render its takeaways at time of compilation and several statistics above come from secondary reporting.Global Center on AI Governance / Research ICT Africa, with 11 regional hubs
The only responsible-AI index with genuine Global South coverage and methodology — 135 countries, 68,000+ assessed data points, 396 responsible-AI frameworks, 38 indicators. Its headline is the commitment–implementation gap in one number: average scores around 35 out of 100, with evidence of implementation in only 55% of cases where frameworks are nominally active. The 2024 first edition covered 41 African countries and is the baseline for change-over-time claims.
Oxford Insights
The longest-running government-readiness ranking, now covering 195 governments. The 2025 edition's framing themes — "from strategy to reality" and rising AI sovereignty, citing OpenEuroLLM and Latam-GPT — are useful for arguing that national AI strategies are moving into implementation. Persistent gaps in Sub-Saharan Africa and South and Central Asia. Read the methodological critique before relying on rank positions.
World Economic Forum
The perception dataset that shapes boardroom and ministerial risk registers. The AI finding is striking and useful: adverse outcomes of AI rank 30th over a two-year horizon but 5th over ten years — the largest rise in ranking of any risk assessed. Misinformation and disinformation ranks second at two years. Evidence that AI risk is perceived as structural and deferred rather than immediate, which is itself a governance problem.
UN Commission on Science and Technology for Development, serviced by UNCTAD
Useful when you need a Member State-endorsed reference rather than a secretariat flagship. Covers AI across conceptualisation, research, development and deployment, flagging black-box opacity, researcher skill erosion and scientific fraud alongside capacity-building recommendations for developing countries.
Read down this column and the logic is visible: an expert body proposes, the General Assembly mandates, a resolution constitutes, and a panel reports. The summit track runs alongside it and is now converging on Geneva.
Filter by instrument type, issuer typology, year and theme, or search titles, issuers and summaries. Every title links to the official source.
Four measures of the core corpus of 104 documents: what kind of instruments they are, who issues them, which policy concepts they engage, and which words recur in their official titles.
Method. Computed over the coded catalogue — official titles, issuer, year, instrument type, plus theme and frame codes assigned by the analyst. Concept counts are the number of documents whose title, summary or codes match a concept's term set; a document can match several. These are not full-text word counts. The reports themselves could not be downloaded from this session (see Cautions & method); once the PDFs are in the folder the same charts can be rebuilt on the real text.
Documents published per year, split by what kind of authority they carry
Documents by source entity typology
Number of documents engaging each policy concept, of 104
What form the documents take
Terms recurring across the 104 core catalogued titles and summaries — size and weight scale with the number of documents using the term
Documents per theme per year — where the publishing energy went
The same measurement applied institution by institution. Each cell is the share of that institution's documents that engage the concept substantively. Read down a column to see which bodies own a vocabulary; read across a row to see what a given institution talks about when it writes about AI.
Share of each institution's documents engaging the concept. Darker is higher.
Share of documents naming the principle. The ethics vocabulary of 2021 is still carried by UNESCO and the UN organs; the OECD and the summit track have largely dropped it.
What changed inside each body's own output between its documents to 2024 and its documents from 2025 onward.
Method. A document “engages” a concept when it uses the matching vocabulary at least three times, or at least once per 20,000 words for short instruments. Institutions with fewer than four documents in either period are shown in the matrices but not given a movement panel. The OECD count is dominated by its working-paper series, which is a research shelf rather than a normative programme — read its row accordingly.
Each document is coded for the frames it advances — a document can carry more than one. Reading the mix year by year shows a field that changed subject twice in four years.
Method. Analyst coding of 103 documents against nine frames, then each year's frame mentions expressed as a share of that year's total. Shares, not counts, so a bigger publishing year does not distort the mix. Early years rest on few documents — 2021 and 2022 are indicative only.
Share of coded frame mentions, per cent
Frame share, 2023 compared with 2026
Institutions changed their own names — the least deniable signal of a framing shift
| Then | Now | When |
|---|---|---|
| AI Safety Summit (Bletchley) | AI Action Summit (Paris) → AI Impact Summit (Delhi) | 2023 → 2025 → 2026 |
| UK AI Safety Institute | UK AI Security Institute | Feb 2025 |
| US AI Safety Institute | Center for AI Standards and Innovation | Jun 2025 |
| International Network of AI Safety Institutes | International Network for Advanced AI Measurement, Evaluation and Science | Dec 2025 |
| International AI Safety Report 2025 — broad scope incl. bias, privacy, environment | 2026 edition narrowed to frontier capability risk | Feb 2026 |
2023 — the risk moment. Bletchley, EO 14110 and the G7 code arrive within three weeks of each other. Safety science and frontier risk together account for a quarter of coded frames. The organising question is whether the technology is dangerous.
2024–25 — the architecture moment. The Global Digital Compact, the EU AI Act, the Council of Europe convention and the first Scientific Report land. Institutional architecture peaks; the question becomes who decides, under what authority. This is also when the environment enters — the IEA and UNEP put electricity and emissions on the agenda for the first time.
2026 — the access moment. Capacity and the divide is now the single largest frame at 24% of coded mentions, nearly four times its 2023 share, while safety science has fallen from 19% to 7%. Delhi, the World Development Report and the UN Panel all ask the same new question: who gets the compute, the data and the gains.
Documents engaging each concept, per year — the same shift read through vocabulary rather than coding
Not a citation graph — a lineage graph. Each edge is a documented relationship: one text mandated, amended, superseded or was constituted by another. Time runs left to right; the four lanes are the institutional streams.
Method. 51 hand-verified lineage edges between catalogued documents. An edge means the relationship is stated in the documents themselves (a resolution mandating a body, a regulation amending another, a summit producing a commitment), not inferred from similarity.
Themes co-occurring within the same document; arc thickness is the number of documents
Source entity typology against theme; cell shade is the number of documents
The same corpus arranged as a map rather than a list: five streams, what each produces, and the instruments that carry real weight in each.
The binding branch is the thinnest. Of 103 documents, five create legal obligations: the EU AI Act and its 2026 amendment, the Council of Europe convention, Korea's AI Basic Act, and China's generative-AI measures. Everything else is normative, advisory or evidential — the architecture is wide and shallow by design.
The evidence branch is the fastest-growing and the most fragmented: two competing international scientific assessments now exist, chaired by the same person, and three separate readiness indices rank the same governments differently.
Every report in the folder, as its own cover. Filter by document type, issuing institution, country or bloc of origin, sector and year, or search the titles. Each cover links to the official source.
Method. Covers are the first page of each PDF as filed, rendered at low resolution. 152 of the 171 catalogued documents have a file on disk and therefore a cover; the rest are listed in the Library as not yet on file. Two files in the folder are corrupt and could not be rendered.
Sixty-eight quantitative findings pulled from the corpus, ordered by how much weight each carries for a policy argument rather than by how striking it looks. Filter by sector, by theme, or by rank. Every figure is attributed to the report it came from.
How the ranking works. Decisive means the finding changes what a policy conclusion can be — the compute concentration figures, the exposure estimates, the single binding fine. Important means it materially strengthens or qualifies an argument. Supporting means it is worth citing but would not alter a conclusion on its own. The judgement is the analyst's; the figures are the sources'.
The main ideas behind the executive summaries of the twenty documents that carry the most weight in this corpus — the binding instruments, the UN architecture, the scientific assessments, the development flagships and the measurement series. Each entry gives what the document does, four findings drawn from its own text, one figure worth quoting, and the thing it leaves unresolved.
Method. Every point below was read out of the document's own full text, held in this folder. Article, paragraph and section numbers are given where the source gives them. Nothing here is drawn from secondary commentary.
Eight sectors, each read across the whole corpus: how the framing changed between 2023 and 2026, the four strongest pieces of evidence, and what the policy question has become.
Documents engaging each sector substantively, of 152 with full text
Mentions per million words, 2024 to 2026, log scale. 2024 is the baseline rather than 2023 because the 2023 slice of the corpus is small enough for one long report to distort it.
Thirty-five actors named across the corpus, ranked by how many of the 152 full-text documents mention them. The count is a measure of presence in the written record, not of influence — but the two are not unrelated, and the gaps are instructive.
Method. Counted by matching each actor's name and acronym across 4.07 million words. Acronyms that collide with ordinary words (WHO, ISO, Meta) are matched case-sensitively. Reach is the number of documents mentioning the actor at least once; mentions is the raw total, which a handful of long reports can dominate.
For each country, the share of its mentions sitting within 250 characters of competition vocabulary (race, leadership, dominance, sovereignty, export control) against capacity vocabulary (access, divide, development, equity, left behind)
Documents mentioning each country, of 152 with full text
Documents using each phrase, of 152 — and total occurrences
Share of each year's documents engaging the term. The three lines that governance texts say matter most are the three that fall.
Type any term. The page reports how many of the 152 full-text reports use it, how often, how that changed by year, which reports use it most, and how it is actually phrased — with the sentence it came from. The index covers 4,297 terms and phrases drawn from 4.07 million words.
Four of these are contested; two are routinely misread. All six are defined as this corpus uses them, not as a dictionary would.
The same measure restricted to what each report puts at the front — the executive summary, key findings or, for short instruments, the whole text. 131 of the 152 documents have an identifiable summary section. This is the vocabulary institutions choose when they have one page to make their case.
Documents whose summary uses the phrase, of 131 — and total occurrences
Documents whose summary uses the word, of 131 — generic connectives removed
Method. A summary region is the text following the first heading matching “executive summary”, “key findings”, “key messages”, “main findings”, “key takeaways”, “summary for policymakers”, “at a glance” or “abstract”, capped at 14,000 characters. Documents under 4,000 words are treated as their own summary. Counts are case-insensitive; plurals of the listed phrases are matched.
Desk research, September 2026
Sources were identified by searching official issuers directly — UN document systems, EUR-Lex, OECD, IEA, World Bank, IMF, ILO, WHO, UNESCO, ITU, national gazettes — and by tracing the lineage backwards from the newest instruments. Each entry records its issuer, date, document symbol where one exists, instrument type and a link to the official text. Themes and narrative frames are analyst-coded.
Items that could not be verified to publication standard are flagged in the companion register rather than silently included. Where sources disagree — the New Delhi Declaration's endorsement count, the Council of Europe ratification tally — the register states the disagreement.
Two layers, measured differently
There are now two analytical layers on this page, and they answer different questions. The catalogue layer — instrument types, issuer typology, themes and narrative frames — is coded by the analyst over 104 curated strategic instruments. The full-text layer — the keyword trends, sector intensities, stakeholder reach, country framing and divide vocabulary — is computed over 4,068,059 words of actual report text from the 152 documents that are in the folder as readable PDFs.
Where a trend is shown as a share of each year’s documents rather than as raw frequency, that is deliberate: the corpus is not evenly sized across years, and a single 200-page report can otherwise move a line on its own. The 2023 slice is small enough (465,000 words) that it is used as a baseline only where the measure is a document share. Country and stakeholder mention counts are raw and are dominated by the long reports; the document-reach column is the steadier number.
Fifteen catalogued documents are not yet in the folder and so are absent from the full-text layer — four are journal articles behind a subscription and two are live dashboards rather than documents. Two files on disk are corrupt and could not be read or rendered: the ITU AI Standards for Global Impact report and the Luccioni Power Hungry Processing paper. Both should be downloaded again.
Regular-expression matching is a blunt instrument. It counts the word, not the argument: a report that mentions the digital divide once in a footnote and one that is about nothing else both register. The text is read for what it foregrounds, not for what it means, and the key-findings tab exists precisely because that distinction matters.
What the search tab can and cannot do
The index holds 4,297 terms and phrases: every word used at least twelve times across at least four documents, plus a curated list of 118 policy phrases. Phrase matching is case-insensitive and includes the plural, which is why a phrase count in the search tab can exceed the strict count quoted elsewhere on this page — “digital divide” appears in 39 documents as an exact singular phrase and 53 including “divides”. A term absent from the index is not necessarily absent from the corpus; it is below the frequency floor.
Example sentences are the first occurrence in the two documents that use the term most, trimmed to about 250 characters. They show how a term is phrased, not how representative that phrasing is.
Two judgements worth flagging
The rank on each quantitative finding — decisive, important, supporting — is the analyst's judgement about argumentative weight, not a property of the source. The figures are the sources'; the ordering is ours, and a reader who disagrees can filter it away.
The section on developments since mid-2025 is sourced to press and official pages rather than to catalogued reports, and it is dated: it is current to 16 September 2026 and will decay. Claims resting on a single outlet are marked in that section. The New Delhi signatory count is disputed across sources and is given as approximately 90.
Every item below is a claim that circulates widely and is either wrong, stale, or dependent on a status that may have changed. These are the ones most likely to cost credibility in a submission.
| Claim in circulation | What to do |
|---|---|
| "The UN AI capacity-building fund" | A US$1–3bn Global Fund is recommended in A/79/966. No evidence of capitalisation as of September 2026. Describe it as proposed. |
| "AI companies' emissions rose 150%" | From ITU's 2025 report, covering 2020–2023. Routinely misattributed to the 2026 edition, which reports different figures on a different sample. |
| "WEF Future of Jobs Report 2026" | Does not exist. The series is biennial: 2025 is current, 2027 is next. Pages using that title recycle 2025 numbers. |
| "AI will add $15.7 trillion to global GDP" | PwC, 2017 — pre-dates generative AI entirely. Still quoted in ministerial speeches. Retire it or caveat heavily. |
| The Council of Europe convention "is in force" | Requires five ratifications including three CoE member states. Check the live Treaty Office chart for CETS 225 before asserting status. |
| "$200 billion pledged at the Delhi summit" | Overwhelmingly private corporate capex in India, not development finance. Do not present it as Global South investment. |
| "A ChatGPT query uses 10× a Google search" | An estimate from UNEP campaign material, materially undercut by Google's own 2025 per-prompt measurement. Attribute the source and the boundary conditions. |
| "The CELAC AI declaration" | No such single instrument. Two separate Latin American tracks — the UNESCO ministerial series and the 2024 Cartagena de Indias Declaration of 17 countries. |
For a foundation working on sustainable development, the corpus above resolves into three usable positions.
HFSHD 2026 · Think Tank · AI Reports
| File | What it is |
|---|---|
| 00-INDEX.md | The full annotated register in markdown — every document, its link and why it matters. |
| 00-CATALOGUE.csv | The 103-document catalogue as a spreadsheet: title, issuer, typology, year, type, status, themes, links. |
| Download-AI-Reports.ps1 | PowerShell script. Right-click → Run with PowerShell. Creates the themed subfolders and downloads every document that has a direct PDF, logging anything it cannot reach. |
| 01–08 subfolders | UN architecture · Summits and safety · Regulation · Development and labour · Energy and environment · Sectors · Research · Indices. |