A United Nations (UN) international scientific panel on artificial intelligence (AI), formed for the first time by the UN, warned that AI’s rapid development is outpacing human understanding and regulatory capacity. It said the possibility of catastrophic harm in the future cannot be ruled out at the current level of science.

Decrypt, a blockchain outlet, reported on Tuesday local time that the UN Independent International Scientific Panel on AI released its first preliminary report involving 40 experts. It presented what it called the first global independent scientific assessment comprehensively analysing AI risks and opportunities.

The report was jointly written by 40 scientists selected from about 2,600 candidates nominated by 140 countries. It was released ahead of the UN’s first Global Dialogue on AI Governance to be held in Geneva, Switzerland on July 6-7.

Co-chair Yoshua Bengio (요슈아 벤지오) said the pace of improvements in AI performance is surpassing humans’ scientific understanding and governments’ ability to respond. In a statement, he said confidence in being able to safely control AI is weakening as it becomes more powerful. He said current science cannot guarantee AI will not cause catastrophic harm on its own or through malicious use.

The report said these concerns do not remain a simple hypothesis. It introduced recent studies that confirmed cases in which some AI systems provided false information or acted in a planned manner to avoid being shut down. It also said an "evaluation awareness" phenomenon has been observed in which systems recognise they are being evaluated and conceal risky behaviour only until testing ends.

U.N. Secretary-General Antonio Guterres also stressed, alongside the report’s release, that governments need a scientific basis to jointly manage AI. He said the world cannot control what it does not understand. He said AI risks are real and the longer the response is delayed, the greater the cost will be.

The report also introduced positive uses of AI. It said AI is currently predicting structures for more than 200 million proteins, speeding up the development of new drugs and vaccines. It also analysed that the amount of work time AI agents can perform without human intervention is doubling about every 4 to 7 months.

It said the uneven distribution of these technological advances across countries is also a problem. About 75 percent of the computing power of the world’s top 500 AI supercomputers is held by the United States, while China remains at about 15 percent. The report analysed that most countries depend on foreign AI systems they cannot build or verify directly, which could create new problems for technological sovereignty and safety.

It also raised concerns about social side effects. The report said cases have been documented in which so-called "sycophantic chatbots" that unconditionally agree with users’ views worsened mental health problems or were linked to actual deaths. It also said a study published this year confirmed an "amplification spiral" in which personalised AI repeatedly reinforces users’ delusions rather than correcting them.

It also assessed that safety verification systems are not sufficient. The report said most countries lack the technical capacity to independently verify frontier AI models, and current safety evaluations rely heavily on information disclosed by developers. It introduced as an attempt to address these limitations a U.S. regulatory approach of reviewing the latest AI models from Google, xAI and Microsoft before their release.

The UN said the document is a preliminary report that organises scientific facts and research findings, not a policy recommendation. The panel operates as an independent body that grants voting rights to neither governments nor companies, and plans to release a more comprehensive final assessment report in 2027. The preliminary report is set to be officially submitted to governments for the first time through the UN AI governance meeting in Geneva.