Governing AI for Humanity: Final Report

Claims that cite this source · 15

1. Source assertion· confidence high· passage checked

   International AI governance initiatives leave most countries out: seven countries participate in all of seven sampled initiatives while 118 participate in none.

   Used in: Civil society and independent researchers, International institution, Broadly shared benefit, Concentration of power

2. Synthesis· confidence medium· passage checked

   Concentration of power over AI's inputs and decisions is itself a governance problem, not only a side effect of unsafe systems.

   Used in: Competitive race dynamics, Compute controls, Concentration of power

3. Source assertion· confidence high· passage checked

   Access to compute, data and talent determines who can participate in AI; the UN advisory body proposes a floor under the AI divide rather than universal access to frontier compute.

   Used in: Evaluation capacity, Compute, Data, language resources, and data labor, Broadly shared benefit

4. Source assertion· confidence high· passage checked

   Representation and coordination in governance are not enough; commitments require implementation and capacity to produce tangible outcomes.

   Used in: Deploying organization, Procurement leverage, Regulator or public authority

5. Interpretation· confidence medium· passage checked

   Many governance failures begin before any rule is broken: competitive pressure, funding, procurement, liability exposure and performance metrics make some behavior rational regardless of stated principles.

   Used in: AI developer or provider, Deployment pressure, Competitive race dynamics, Procurement leverage

6. Source assertion· confidence high· passage checked

   No single level of governance can solve the whole problem: national and regional action is crucial, but the transboundary nature of AI requires global coordination.

   Used in: International institution

7. Source assertion· confidence high· passage checked

   Large parts of the world are data-poor, and failure to reflect linguistic and cultural diversity is linked both to bias and to missed benefits.

   Used in: Data, language resources, and data labor, Broadly shared benefit

8. Source assertion· confidence high· passage checked

   AI-enhanced creation and dissemination of disinformation is identified among AI's risks alongside bias, surveillance, and energy consumption.

   Used in: Systematically unreliable provenance

9. Interpretation· confidence medium· passage checked

   Where schools, clinics, community groups and small businesses adopt AI, governance is real but bounded: they can set how a tool is used, but usually cannot evaluate it independently or change how it was built, so their choices depend on decisions made by vendors and higher levels.

10. Interpretation· confidence medium· passage checked

    “How societies should” sets the scope wider than a single organisation, where several explicit definitions place governance (Mäntymäki et al., Schneider et al., IAPP, IBM, Japan’s business guidelines). It is closest to Chile’s multi-actor definition of governance and the UN advisory body’s multi-stakeholder framing, and it keeps the question normative: what should be done, not only what is done.

11. Interpretation· confidence medium· passage checked

    None of the definitions compared here lists incentives or infrastructure among what it governs; they govern AI itself, its lifecycle, or an organisation’s use of it. The idea is not new, though: the UN advisory body says governance frameworks should “shape incentives globally”, Japan’s guidelines run agile governance across systems that include infrastructure, and Sastry et al. treat compute as one tool of AI governance. Naming incentives, institutions, infrastructures and safeguards together as the objects of governance is AIGovSpec’s own synthesis.

12. Interpretation· confidence high· passage checked

    Organisational definitions aim governance at strategy, law and ethics principles. Broad sharing of benefits appears instead in intergovernmental and Global South framings: the UN advisory body’s “leave no one behind”, the African Union’s inclusion and African agendas, and the Montevideo declaration’s inclusion and equitable access.

13. Interpretation· confidence medium· passage checked

    Adaptability is common: the OECD, the UN advisory body, Japan, Kenya, the African Union and the World Bank all expect governance to be agile or regularly revised. Acting before arrangements set is rarer. Dafoe urges embedding principles while the stakes are relatively low, and Ojanen et al. draw on Guston’s anticipatory governance (“while effective intervention is still possible”), but none of the definitions compared here builds that timing into the definition of AI governance itself. This is the clause in which AIGovSpec departs most from them.

14. Interpretation· confidence high· passage checked

    “How societies should” leaves open who decides and how. Chile’s definition names the actors (government, the private sector, citizens and academia), and the UN advisory body calls for multi-stakeholder collaboration; the AIGovSpec definition names neither and relies on its legitimacy test instead.

15. Synthesis· confidence high· passage checked

    Of the 29 governmental, intergovernmental and standards documents compared here, only two, Japan’s AI Guidelines for Business and Kenya’s AI strategy, define AI governance in a sentence. Most use the term without defining it, and some laws, including the Council of Europe Convention and Korea’s framework act, do not use it at all.