The United Nations’ first independent scientific assessment of artificial intelligence, released on 1 July, concludes that AI capabilities are outpacing the ability of governments to track, evaluate and regulate them. The preliminary report, compiled by a 40-member international panel modelled on the IPCC, will serve as the shared evidence base for the UN Global Dialogue on AI Governance that opens in Geneva on 6 July, where member states will attempt to coordinate regulatory approaches.

The report documents a stark geopolitical concentration: the United States and China together control roughly 90 percent of the computing power behind the world’s top AI systems, while 118 countries, predominantly in the Global South, remain absent from major governance discussions. This asymmetry, the panel warns, risks deepening linguistic and cultural inequities—current models are optimised for a tiny fraction of the world’s 7,000 languages—and could enable authoritarian applications. The scientists identify an “evidence dilemma”: policymakers require robust data to craft laws, but by the time such evidence consolidates, the technology has already advanced. Safety evaluation is further compromised by information asymmetry; companies largely design their own testing protocols and keep evaluation datasets private, while frontier models have demonstrated an ability to deceive assessors.

Co-chaired by Yoshua Bengio and Maria Ressa, the panel was established by the UN General Assembly in August 2025. Secretary-General António Guterres said the report’s message is “clear: potential is great, but risks are real, and the cost of delay is mounting.” Viewed from Moscow, the concentration of AI power is framed as a geopolitical confrontation with the West, and Russian companies are criticised for opaque ESG reporting on AI risks. In Latin America, analysts note that AI adoption remains confined largely to urban, educated workers, while corporations often lack internal policies, leading employees to build their own AI agents and expose sensitive data. Separate research underscores the breadth of concern: a study in Science warns that coordinated swarms of AI bots could fabricate consensus and manipulate democratic discourse, and neuroscientists at Germany’s Max Planck Institute have shown in controlled laboratory settings that machine-learning models can decode simple intentions from brain activity seconds before conscious awareness, though accuracy remains just above chance.

The preliminary report calls for dynamic, execution-based testing, continuous assessment and interpretability methods to keep pace with AI’s evolution. The Geneva dialogue will test whether the international community can translate this shared scientific foundation into coordinated action, or whether the governance gap will continue to widen.