The first independent scientific evaluation of artificial intelligence by a United Nations panel concludes the technology can deliver enormous benefits but also carries significant dangers if deployed without adequate scientific oversight and governance.
The preliminary report, prepared by a 40-member panel drawn from every world region and serving three-year terms independent of any government, institution or company, will be presented to governments at the inaugural U.N. Global Dialogue on AI governance in Geneva on July 6 to 7. A fuller, comprehensive assessment is planned for release next year.
Evidence and governance gaps
According to the panel, policymakers require robust scientific evidence to steer AI governance. However, the report warns that AI capabilities are outpacing the scientific community's ability to understand them and most governments' capacity to adapt. The document notes there are few reliable methods available for controlling highly autonomous AI systems, creating a widening gap between technological progress and governance tools.
Panel co-chair Yoshua Bengio highlighted a worrying trend in AI behavior. He pointed to growing evidence of deceptive AI behavior and said science could not guarantee AI will not cause catastrophic harm "either on its own or due to malicious users" as capabilities increase.
Potential gains and broad risks
"The potential benefits of AI are enormous," the report states. At the same time, it cautions that the rapid, unchecked deployment of AI at scale "also presents considerable risks, including harms to the mental health of users, potential use as a destructive tool, impacts on social, economic and environmental systems, and challenges associated with controlling the technology."
The panel documents uneven global adoption. While more than a billion people now use conversational AI weekly, uptake across developing countries lags behind wealthier nations. Development of the underlying AI systems is even more concentrated: the United States accounts for 75% of the computing power among the world's top 500 AI supercomputers, with China supplying 15%.
Language, health and information integrity concerns
The report draws attention to linguistic coverage gaps. Although more than 7,000 languages are spoken globally, current AI models are trained on only a small fraction. This limited coverage, the panel warns, leads to machine translation errors for some languages that can affect health diagnoses and treatment decisions.
On information integrity, the panel highlights that AI makes it easier to generate and target persuasive content at scale, contributing to what it describes as a "gradual erosion of information integrity that can weaken public trust, social cohesion and democratic deliberation." The report also raises alarms about the increased circulation of AI-generated child sexual abuse material and deepfake-enabled sexual violence.
Capacity shortfalls
Finally, the report notes a widespread lack of technical expertise. Most countries, including many advanced economies, do not possess the capabilities to assess the most capable new AI models or to engage meaningfully in their governance, the panel says.
This preliminary assessment sets out the scientific community's early findings ahead of the Geneva dialogue and signals areas where further research and international cooperation will be needed before the planned comprehensive report next year.