There is no single global AI transition
Countries differ in electricity, connectivity, compute, education, language coverage, research capacity, fiscal space, law, and institutional trust. A tool that raises productivity in a connected firm may be inaccessible where bandwidth is costly. A medical model validated in one health system may fail where disease prevalence, records, equipment, or clinical workflows differ.
Current evidence describes uneven generative-AI adoption. Forecasts about AGI transforming every country simultaneously are scenarios. Physical infrastructure, institutions, and diffusion create delays and different outcomes even when software crosses borders quickly.
The divide is about use and production
Access to a chatbot is not the same as capacity to shape the technology. Production power includes chips, cloud infrastructure, data centers, capital, advanced research, benchmark design, standards participation, and the ability to audit models. Countries may become consumers of systems trained elsewhere and governed by foreign contracts.
UNDP’s 2025 Next Great Divergence report warns that countries with stronger connectivity, skills, compute, and regulation are better positioned to capture gains, while others face exclusion and insecurity. It reports wide use gaps and emphasizes risks to public services, labor, and data representation (UNDP, December 2025). UNDP has a development mandate; its report should be read as policy analysis, not a deterministic economic forecast.
Local capacity does not require every country to train a frontier model. It can mean reliable connectivity, shared compute, procurement expertise, local evaluation, language resources, competition among providers, secure deployment, and the legal ability to protect residents.
Language and data shape visibility
Models perform unevenly across languages and cultural contexts. Digitally abundant languages receive more training data, benchmarks, and developer attention. Low-resource languages may face worse accuracy or be absent from safety testing. Translation can expand access while flattening context or imposing terminology designed elsewhere.
Data deserts create another exclusion. People missing from records may receive poor service; people overrepresented in policing or debt records may receive disproportionate scrutiny. Filling gaps is not simply collecting more personal information. Community governance, consent, minimization, benefit sharing, and locally meaningful labels matter.
Indigenous and community knowledge may be collectively governed even when internet-accessible. Legal permission, ethical permission, and technical availability are not identical. Data partnerships should state ownership, allowed use, compensation, deletion, and whether resulting models are accessible to contributors.
Labor and value chains cross borders
AI systems depend on data labeling, content moderation, evaluation, mining, manufacturing, construction, and energy work. Much is performed far from model headquarters. Low wages, traumatic content, weak labor protection, and environmental costs can be hidden behind a clean interface.
Automation may affect outsourced business-process and digital-service sectors on which some economies rely. It may also let local firms export new services and overcome expertise shortages. Net effects depend on task change, adoption cost, demand, skills, and market access.
Contracts and trade policy determine who captures value. Local workers and firms need bargaining power, portable skills, and access to tools—not only retraining after disruption.
Infrastructure can create benefits and burdens
Emerging and developing economies excluding China contain more than two-thirds of the world’s population but less than 10% of data-center capacity, according to the IEA’s 2025 analysis (IEA). New facilities can bring investment and connectivity, yet compete for reliable electricity, water, land, and tax revenue.
Governments should publish power and water requirements, incentives, employment promises, and cost allocation. Communities need participation before permits. “Digital development” should not mean households subsidize infrastructure whose services and profits leave the country.
Regional compute pools and research networks can broaden access. They require transparent allocation, cybersecurity, maintenance funding, and protection from political capture.
Global governance is institutionally fragmented
The EU AI Act regulates access to a major market. The Council of Europe Framework Convention addresses human rights, democracy, and rule of law for parties. OECD and UNESCO principles, G7 processes, AI safety summits, technical standards, and UN initiatives provide additional forums. Most are not global binding regulation of frontier capability.
The Council of Europe convention opened for signature in 2024 and has participation from non-European states, but signature is not ratification and the official treaty status determines legal force (Council of Europe). It should not be described as a universal AGI treaty.
Fragmentation can allow experimentation and reflect legal diversity. It can also create compliance gaps, conflicting definitions, and rulemaking dominated by actors able to attend every forum. Interoperability should preserve substantive rights rather than merely lower industry cost.
Representation must be operational
Inviting more countries to a summit is not enough if agendas, evidence, and drafting remain controlled elsewhere. Meaningful participation requires travel and technical funding, translation, access to evaluation tools, time to consult domestic communities, and influence over final rules.
Civil society, workers, indigenous peoples, disability groups, youth, small businesses, and environmental communities need channels in addition to national governments. Governments do not always represent vulnerable residents, and some seek AI-enabled surveillance.
Conflict disclosure matters. Frontier firms benefit from compatible rules and market access. Wealthy states may favor standards built around their infrastructure. Development institutions prioritize inclusion. Security agencies prioritize control. None is disqualified, but no perspective should be treated as neutral.
A global public-interest agenda
Practical priorities include:
- multilingual benchmarks and safety evaluations governed with local experts;
- affordable secure compute and connectivity for research and public services;
- procurement templates with audit, privacy, incident, and exit rights;
- labor and environmental standards across the AI supply chain;
- cross-border incident notification and technical assistance;
- interoperable documentation without weakening national rights;
- funding for regulators and universities independent of vendors;
- benefit-sharing and community governance for sensitive data;
- preservation of non-digital and non-AI access to essential services.
International finance can support grids and connectivity, but debt, procurement lock-in, and vendor conditions require scrutiny. Public capacity should remain after a grant or pilot ends.
Avoid two forms of paternalism
One error is exporting untested systems and claiming innovation overrides local consent. The other is denying lower-income countries access to beneficial technology in the name of safety while wealthy countries continue deployment. Risk thresholds should be evidence-based and paired with resources that make compliance and benefit possible.
Countries should be able to reject surveillance-heavy or culturally unsuitable systems without losing access to infrastructure. They should also have the option to develop and adapt models under responsible conditions.
AGI and geopolitical distribution
If AGI-like capability were controlled by a few companies or states, bargaining power, scientific capacity, and security could become more unequal. If weights or services diffused rapidly, access might broaden while misuse and dependency increased. These are conditional pathways, not measured futures.
Global safety is not charity. A severe incident, insecure model, pandemic, cyberattack, or destabilizing labor shock crosses borders. Excluding most of humanity weakens evidence, legitimacy, and compliance.
The right objective is distributed capacity with shared safeguards: every society needs the ability to evaluate systems affecting it, protect rights, shape rules, and capture benefits. Governance that merely exports standards from frontier states will be less fair and less effective than governance built with the people expected to live under it.