The warning accompanied the release of the preliminary report by the Independent International Scientific Panel on Artificial Intelligence, the first independent global assessment commissioned by the United Nations to examine both the opportunities and risks associated with AI. The report will serve as a key reference for discussions at the inaugural Global Dialogue on AI Governance, scheduled to take place in Geneva on July 6 and 7.

The 40-member panel was established by the UN General Assembly in August 2025 with a mandate to assess scientific evidence and provide policy-relevant findings without recommending specific regulations. Its members, selected from more than 2,600 applicants representing 140 countries, serve in their personal capacities for three-year terms.

Presenting the report, panel co-chair Yoshua Bengio said AI capabilities are advancing more rapidly than governments and scientific institutions can fully understand or regulate. He cited growing evidence that advanced AI systems have demonstrated deceptive behavior in controlled laboratory environments and said existing scientific knowledge cannot guarantee that increasingly capable systems would be incapable of causing catastrophic harm, whether through unintended actions or malicious use by individuals or organizations.

Co-chair Maria Ressa said the panel’s findings were shaped by three defining trends: accelerating technological capability, increasing concentration of power and declining human control over advanced AI systems. She pointed to results from the “Humanities Last Exam,” a benchmark consisting of 2,500 expert-level questions across multiple disciplines, where leading AI systems improved their scores from 8 percent to 45 percent within just 16 months.

Ressa also highlighted the growing concentration of computing infrastructure and AI development. According to the panel, the United States accounts for approximately 75 percent of computing power used by the world’s largest AI clusters. The report further found that 91 percent of notable AI models introduced during 2025 originated from private-sector developers. Institutions based in the United States produced 59 major AI models during the year, compared with 35 from China and 13 from all other countries combined.

UN Secretary-General António Guterres said the panel’s assessment had been distributed to governments worldwide, urging policymakers to establish shared international safeguards before technological progress outpaces public oversight. He warned that the absence of common governance frameworks would steadily reduce the ability of governments and citizens to influence how AI develops and is deployed, adding that leaders could no longer claim ignorance of the emerging risks.

The report examined eight broad areas, including AI science, future technological trajectories, economic effects, international security, environmental consequences, human rights, democratic governance and system reliability. While identifying major policy challenges, the panel deliberately stopped short of recommending specific regulatory measures, leaving those decisions to UN member states.

A central finding of the assessment was the continued exclusion of much of the Global South from both AI development and global governance processes. The panel said many developing countries face some of the greatest potential economic and social impacts from AI while possessing the fewest technical resources and institutional capacities to influence international standards or respond effectively to emerging risks.

Panel members noted that the current body itself represents the first international scientific assessment on AI to include substantial participation from researchers across Africa, South Asia, Southeast Asia and Latin America, allowing scientists from the Global South to help shape the global evidence base rather than simply adapt to standards developed elsewhere.

Responding to questions about whether the panel would support an international mechanism to evaluate AI models before public release, Bengio said such decisions fall outside the panel’s mandate, although concerns surrounding advanced AI systems had informed the areas examined in the report. Ressa reiterated that the assessment was intentionally designed to be policy-relevant without prescribing policy choices, leaving governments to determine appropriate regulatory responses during next week’s dialogue in Geneva.

The panel’s next comprehensive assessment is expected to contribute to the second Global Dialogue on AI Governance, scheduled to be held in New York in May 2027.