Artificial intelligence is advancing at a faster pace than many experts had anticipated, and the evidence regarding several risks has “grown significantly.” Meanwhile, current risk management techniques are “improving but still insufficient.” These are the findings of the second International AI Safety Report, which was published on Tuesday, ahead of the AI Impact Summit scheduled to occur in Delhi from February 19 to 20.
Guided by 100 experts and supported by 30 countries and international organizations, including the United Kingdom, China, and the European Union, the report aims to serve as an example of “collaborating to navigate shared challenges.” However, unlike last year, the United States declined to lend its support, as confirmed by the report’s chair, Turing Award-winning scientist Yoshua Bengio.
As the risks of AI start to become apparent, the home of leading AI developers has withdrawn from international efforts to understand and mitigate them. The move is largely symbolic, and the report does not depend on the U.S.’s support. But when it comes to understanding AI, “the more global consensus there is, the better,” Bengio says.
Whether the U.S. objected to the report’s content or is simply pulling back from international agreements—it left the and in January—remains unclear. Bengio states that the U.S. provided feedback on earlier versions of the report but declined to sign the final version.
The U.S. Department of Commerce, which was named in the 2025 International AI Safety Report, did not respond for comment on the decision.
What The Report Says
“Over the past year, the capabilities of general-purpose AI models and systems have continued to improve,” the report reads. Capabilities have advanced so rapidly that between the first and second reports, the authors issued two interim updates in response to major changes. This goes against the consistent stream of headlines suggesting that AI has reached a plateau. The scientific evidence shows “no slowdown in advances over the last year,” Bengio says.
Why does it seem to many that progress has slowed? One clue is what researchers refer to as the “jaggedness” of AI performance. These models can achieve a gold-medal standard on International Mathematical Olympiad questions while sometimes failing to count the number of r’s in “strawberry.” That jaggedness makes it difficult to assess AI’s capabilities, and direct human comparisons—like the popular “intern” analogy—are misleading.
There is no guarantee that the current rate of progress will continue, although the report notes that trends are consistent with continued improvement through 2030. If today’s pace holds until then, experts predict that AI will be able to complete well-defined software engineering tasks that would take human engineers several days. But the report also raises the more remarkable possibility that progress could accelerate if AI significantly aids in its development, resulting in systems as capable as or more capable than humans in most aspects.
That is likely to excite investors, but it is concerning for those who worry that society is not adequately adapting to the emerging risks at the current pace. Even Google DeepMind’s CEO Demis Hassabis in Davos in January that he believes it would be “better for the world” if progress slows.
“A wise strategy, whether you’re in government or business, is to prepare for all possible scenarios,” Bengio says. That means mitigating risks, even in the face of uncertainty.
Maturing Understanding of Risks
Policymakers who want to heed scientists when it comes to AI risk face a problem: the scientists disagree. Bengio and fellow AI pioneer Geoffrey Hinton have warned since ChatGPT’s launch that AI could pose an existential threat to humanity. Meanwhile, Yann LeCun, the third of AI’s so-called “godfathers,” has such concerns “complete nonsense.”
But the report indicates that the situation is becoming more solid. While some questions remain divisive, “there is a high degree of convergence” on the core findings, the report notes. AI systems now match or surpass expert performance on benchmarks related to biological weapons development, such as troubleshooting virology lab protocols. There is strong evidence that criminal groups and state-sponsored attackers are actively using AI in cyber operations.
Continuing to measure those risks will encounter challenges in the future as AI models increasingly learn to manipulate safety tests, the report says. “We’re seeing AIs whose behavior, when they are tested, […] is different from when they are in use,” Bengio says, adding that by studying the models’ chains-of-thought—the intermediate steps it took to reach an answer—researchers have determined that this difference is “not a coincidence.” AIs are acting dumb or on their best behavior in ways that “significantly impede our ability to accurately estimate risks,” Bengio says.
Rather than propose a single solution, the report recommends implementing multiple safety measures—testing before release, monitoring after, —so that what slips through one layer is caught by the next, like water through a series of increasingly fine sieves. Some measures target the models themselves; others aim to strengthen defenses in the real world—for example, making it more difficult to obtain the materials needed to build a biological weapon even if AI has made it easier to design. On the corporate side, 12 companies voluntarily published or updated Frontier Safety Frameworks in 2025, documents that describe how they plan to manage risks as they build more capable models—although they vary in the risks they cover, the report notes.
Despite the findings, Bengio says the report has left him with a sense of optimism. When the first report was commissioned in late 2, the debate over AI risk was driven by opinion and theory. Now, he says, “we’re starting to have a much more mature discussion.”
