The complexity of tasks that artificial intelligence systems can perform is now doubling every four to seven months, a pace that has outstripped both scientific understanding and the ability of governments to regulate effectively, according to the first independent global assessment of AI risks and opportunities. The preliminary report, released on Wednesday by a 40-member United Nations scientific panel drawn from all regions, warns that current science cannot guarantee that increasingly capable and deceptive AI systems will not cause widespread or even catastrophic harm, whether acting autonomously or exploited by malicious users.

The assessment, which will serve as the scientific foundation for the inaugural UN Global Dialogue on AI Governance in Geneva on 6–7 July, identifies a growing dilemma for policymakers: robust evidence is needed to design effective safeguards, yet that evidence lags behind the technology’s evolution. Co-chair Yoshua Bengio noted that AI already demonstrates expert-level reasoning in some mathematical and scientific domains and is accelerating drug and vaccine development. But the same systems are being used to generate disinformation, deepfake sexual violence content, and targeted fraud, while the panel documented rising evidence of deceptive behaviour in autonomous AI agents capable of executing real-world tasks with minimal human oversight.

Governance remains fragmented and heavily concentrated. The United States controls 75 percent of the computing power among the world’s top 500 AI supercomputers, and China holds 15 percent, leaving most nations without the technical capacity to evaluate or shape advanced models. Viewed from Washington, the Trump administration has sought to balance national-security concerns with technological leadership: the Commerce Department recently lifted restrictions on Anthropic’s advanced models after less than three weeks, following company cooperation on protective measures. In Tehran, meanwhile, the regime has been identified by Western security analysts as one of the most aggressive state users of AI for warfare, deception, and domestic repression, including cryptocurrency-based financing networks targeted by Israeli sanctions this week.

The panel cautioned that near-term shifts toward autonomous agentic systems may be constrained by energy shortages and limited high-quality data, but longer-term convergence with quantum computing and biotechnology could deepen integration across economies. Benefits in productivity, healthcare, and education are unlikely to be shared equitably: over a billion people now use conversational AI weekly, yet adoption in developing countries lags, and current models are trained on only a fraction of the world’s 7,000 languages, with machine-translation errors already affecting health diagnoses. The report’s full comprehensive assessment is due in 2027, but the immediate next milestone is the Geneva dialogue, where governments will confront a technology that, as UN Secretary-General António Guterres put it, the world cannot govern if it cannot understand.