An independent, first-of-its-kind scientific assessment of AI published by the United Nations has made a startling conclusion: No one can currently guarantee the technology won't cause catastrophic harm.
The UN’s Independent International Scientific Panel on Artificial Intelligence made the finding in a July 6 preliminary report. The panel of 40 scientists, selected from more than 2,600 candidates across 140 countries, released the first global scientific review of AI's risks and benefits… and its findings are both dire and significant.
"AI capabilities are outpacing both scientific understanding and governments' ability to adapt," said panel co-chair Yoshua Bengio, the Turing Award-winning founder of Mila, in the panel's statement.
Citing growing evidence of deceptive AI behavior, Bengio added that as AI’s capabilities continue to grow, science cannot guarantee it won't cause “catastrophic harm” either on its own or through malicious use.
To make its case, the panel’s report documents laboratory cases of AI systems lying and scheming to avoid being shut down. Additionally, researchers noted a related pattern they refer to as “evaluation awareness,” the ability for models to recognize when they're being tested and decrease risky behavior just long enough to pass the check.
UN Secretary-General António Guterres characterized the report as the shared evidence base that governments have lacked.
"The world cannot govern what it cannot understand," he said in the statement, calling the risks real and warning that the cost of waiting keeps rising.
AI’s potential benefits remain enormous
The report does not treat AI only as a threat. It also points to major benefits if the technology is developed carefully, evaluated independently, and deployed in ways that serve the public interest.
AI could help accelerate progress toward the UN’s Sustainable Development Goals by improving access to education, expanding medical research, supporting climate modeling, and helping governments deliver services more efficiently. In health science, AI systems have already helped predict the structure of more than 200 million proteins, a breakthrough that could speed up drug discovery, vaccine development, and disease research.
The report also notes that AI agents are becoming more capable of completing longer tasks on their own, with the length of tasks they can handle doubling roughly every four to seven months. That could eventually make AI useful for scientific discovery, administrative work, software development, and other complex workflows.
But the panel’s point is not that benefits outweigh the risks. It is that both are growing at the same time, which makes stronger scientific understanding and independent oversight more urgent.
Researchers warn of chatbot and model safety risks
But the panel also flagged overly agreeable chatbots that side with users regardless of accuracy — and cited severe mental health incidents, including documented deaths. Separate research published this year describes a similar feedback loop, known as an amplification spiral, that links common chatbot behaviors to reinforcing user delusions rather than correcting them.
All of this means policymakers have significant challenges due to the rapid pace of technological development and the breadth of potential applications.
Bengio co-chairs the panel with the Nobel Peace Prize-winning journalist and Rappler co-founder Maria Ressa, who must ensure adherence to a UN General Assembly mandate stipulating that the panel may document scientific consensus only, not prescribe policy. Notably, no government, company, or institution gets a vote.
“The technology is transformative, but if the world keeps moving along this trajectory, humanity will fail to realize the gains it promises,’’ warned Ressa in a statement. “The risks—to societies, to security, and to our species—are too high, and the forces driving AI forward are not the forces that will deliver its benefits.
AI capabilities are not equal
AI progress is uneven: among the world's top 500 AI supercomputers, the U.S. controls 75% of the computing power, compared to 15% for China.
Most countries lack the technical capacity to do their own evaluation of frontier models, according to the report, and assurances about safety remain heavily based on what developers choose to disclose. This is the same gap U.S. regulators are trying to tighten by negotiating deals for pre-release access to models from Google, xAI, and Microsoft.
“The concentration of AI capabilities in a small number of firms and countries could enable authoritarian capture and undermine democratic accountability,’’ the report states. “This leaves most countries dependent on systems they can't build, audit, or fully control.”
This preliminary report is the panel's initial statement. The panel will prepare a full, comprehensive assessment in 2027, and governments are reviewing the early findings at the UN's inaugural Global Dialogue on AI Governance in Geneva this week.
Also read: Cloudflare is giving website owners new ways to block or charge AI crawlers, a shift that could reshape how AI companies access online content and how publishers protect it.