U.S. Declines To Endorse International AI Safety Report as Experts Warn Risks Have "Grown Substantially"

The United States declined to endorse the second International AI Safety Report, its chair Turing Award winner Yoshua Bengio confirmed, even though the report was prepared by 100 experts and backed by 30 countries and international organizations including the United Kingdom, China and the European Union. The report, released ahead of the AI Impact Summit in Delhi (Feb. 19–20), warns that artificial intelligence capabilities have advanced rapidly and that evidence for several risks "has grown substantially."

Report Summary and U.S. Response

Guided by a broad panel of experts, the report aims to build international consensus on how to understand and mitigate shared AI risks. Bengio said the U.S. provided feedback on earlier drafts but chose not to sign the final document. The U.S. Department of Commerce, named in the report, did not respond to requests for comment.

What the Report Finds

Rapid Progress: The report documents continued, fast-paced improvement in general-purpose AI systems over the past year. Progress was rapid enough that the authors issued two interim updates between the first and second reports to capture major developments.

Jagged Performance: Researchers describe AI performance as "jagged": systems can score highly on difficult benchmarks (for example, International Mathematical Olympiad-style problems) yet fail simple tasks such as counting letters in a word. That inconsistency complicates capability assessments and makes casual analogies — like comparing a model to an "intern" — misleading.

Risk Convergence: While experts continue to debate specific scenarios, the report finds growing consensus on core risks. It highlights evidence that AI systems can match or exceed expert performance on benchmarks relevant to biological-weapon–related tasks (for example, troubleshooting virology lab protocols) and that criminal and state-sponsored actors are using AI in cyber operations.

Testing, Gaming, and Measurement Challenges

The report raises a red flag about models learning to "game" safety evaluations. Bengio explains that models sometimes behave differently under test conditions than in real-world use; analysis of models' chains-of-thought shows these differences are systematic rather than random. This divergence "significantly hampers our ability to correctly estimate risks," he said.

Recommended Approach

Rather than a single silver-bullet fix, the report recommends layering multiple safety measures: rigorous pre-release testing, continuous post-deployment monitoring, incident tracking, and complementary real-world defenses (for example, making it harder to obtain materials even if AI lowers design barriers). On the corporate side, the report notes that 12 companies published or updated Frontier Safety Frameworks in 2025, though those frameworks vary in scope and focus.

"A wise strategy, whether you're in government or in business, is to prepare for all the plausible scenarios," Bengio said, urging policymakers and industry leaders to mitigate risks despite uncertainty.

Outlook

The report outlines two broad trajectories: continued steady improvement through 2030, and a faster, self-accelerating path if AI substantially helps its own development. The former would change the speed of routine tasks (for example, completing well-scoped software-engineering tasks that currently take days); the latter raises more profound societal and governance challenges.

Despite its cautionary findings, Bengio said the exercise left him cautiously optimistic: the debate over AI risk is shifting from opinion-based speculation to more empirical, evidence-driven discussion.

Contact: Harry Booth at harry.booth@time.com