Developments in artificial intelligence are outpacing scientific understanding and government policy, meaning there are no guarantees the technology will not cause catastrophic harm, the United Nations’ Independent International Scientific Panel on Artificial Intelligence warned on Wednesday. The panel’s preliminary report frames the central dilemma facing policymakers with unusual clarity: they need robust evidence to regulate AI effectively, yet that evidence is struggling to keep pace with how quickly the technology is evolving.
“AI capabilities are outpacing both scientific understanding and governments’ ability to adapt,” said Yoshua Bengio, the panel’s co-chair. “With growing evidence of deceptive AI behaviour, science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm, either on its own or due to malicious users.”
The panel, comprising 40 experts drawn from every region of the world and serving three-year terms independent of any government, institution, or company, was established by the UN General Assembly in 2025 specifically to provide policymakers with a scientific foundation resilient to political cycles. Notably, the body is scientific rather than regulatory in mandate. It assesses evidence and publishes findings that governments can act on, without itself setting rules or enforcing standards.
The report will be formally presented to governments at the inaugural UN Global Dialogue on AI Governance in Geneva from July 6 to 7, with a fuller, more comprehensive report planned for release next year. UN Secretary-General António Guterres welcomed the assessment pointedly, stating that the world cannot govern what it cannot understand, and warning governments not to wait, since the potential of AI is great but the risks are real, and the cost of waiting is rising.
The Specific Risks the Panel Flagged
The report’s findings move well beyond generalised caution into specific, documented concerns. The panel noted that AI agent systems, capable of independently carrying out real-world tasks, will soon complete work that currently takes human programmers days or weeks, but warned that there are no scientific guarantees such systems will not violate instructions, with evidence already accumulating of cases where they have.
The mental health dimension of the report is particularly striking. Sycophantic AI behaviour, where systems generate responses that reinforce a user’s existing beliefs regardless of accuracy, has been linked to several severe mental health incidents, including documented deaths. The panel also flagged that criminals and bad actors have been documented using AI systems to assist in cyberattacks, and that advanced technical capabilities may allow even novice private actors to deploy AI maliciously across applications ranging from fraud to disinformation.
The report additionally identified AI-generated sexual abuse material and deepfakes as a growing threat, noting that women and children are most at risk. It flagged the environmental footprint of the energy-hungry data centres powering AI development as a mounting concern, and warned that poorer nations risk being locked out of the AI era entirely, unable to build, inspect, or audit the systems they increasingly find themselves dependent on.
Benefits Alongside the Warnings
The panel was careful to frame its assessment as balanced rather than alarmist. In its executive summary, it wrote that deployed and applied thoughtfully, AI can support progress towards the Sustainable Development Goals, advance health science, and increase access to education, citing breast cancer detection and accelerated vaccine development among its concrete achievements to date.
At the same time, the panel cautioned that the rapid, unchecked deployment of the technology at scale 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 itself. It noted that AI adoption has accelerated broadly but unevenly across countries and sectors: globally, over a billion people now use conversational AI on a weekly basis, but adoption in developing countries continues to lag well behind.
Looking ahead, the panel expects a near-term shift toward agentic AI systems capable of carrying out real-world tasks, though it noted this growth may be constrained by energy limitations and shortages of high-quality training data. Over a longer horizon, it anticipates self-improving AI becoming more deeply embedded in the global economy and increasingly converging with technologies such as quantum computing and biotechnology.
An Honest Accounting of What Remains Unknown
What distinguishes this report from typical policy documents is its explicit acknowledgement of its own evidentiary limits. The panel stated plainly that policymakers face an evidence dilemma: they can either make governance decisions now with insufficient scientific grounding, or wait for stronger evidence that may arrive too late for any meaningful intervention.
The report also catalogued specific gaps where current evidence remains too uneven or insufficient to support confident scientific conclusions, including the macroeconomic and productivity effects of AI adoption, the technology’s environmental footprint, the structure of the global AI supply chain, and its effects at both the individual and collective level. The panel noted that its preliminary scope did not extend to military applications of AI or lethal autonomous weapons systems, an area likely to feature more prominently in its fuller report next year.
As governments prepare to gather in Geneva next week for the first Global Dialogue on AI Governance, the panel’s message carries a deliberate weight: the scientific foundation for informed AI policy now exists in a form it never has before. What remains uncertain is whether governments will act on it before the gap between capability and governance widens further.
