Experts examine AI's potential risks

A seminar hosted by the Center for Strategic and International Studies on Tuesday in Washington provided one of the first in-depth public examinations of the International AI Safety Report 2026, the second edition of this comprehensive global assessment.

As both the report and participants highlighted, the urgent need for stronger international coordination has emerged as a central theme in addressing global risks posed by rapidly advancing artificial intelligence.

The report, chaired by Turing Award recipient Professor Yoshua Bengio of Universite de Montreal and the Mila-Quebec AI Institute, drew on contributions from over 100 AI experts nominated or supported by more than 30 countries and international organizations, including China.

"The original idea behind the report was … to build a shared evidence base to inform decision-making about AI technologies," said lead writer Stephen Clare.

"There were a lot of questions they were facing about AI … and not a lot of consensus on what the actual technical realities of the technology were."

The AI Safety Summit 2023 was the world's first major intergovernmental conference on frontier AI risks. Held in November 2023 at Bletchley Park in the United Kingdom, it produced the landmark Bletchley Declaration, signed by 28 countries, including China and the US.

The declaration recognized shared concerns about potential catastrophic AI risks and committed signatories to international cooperation on research, risk assessment and mitigation, directly laying the foundation for the ongoing International AI Safety Report series.

"Since that time, we have a lot more empirical evidence we can actually rely upon. And we're able to discuss a lot more concrete cases of AI impacts, more evaluations, and more data we can actually use in the report," Clare said.

Stephen Casper, who led the technical safeguards section of the seminar, provided a detailed explanation of frontier model development stages and protections, noting that "different types of safeguards and risk management techniques apply at different parts in the life cycle".

He said progress has been made in creating multiple layers of defenses, but highlighted persistent governance gaps. For closed models, Casper said, there is "a pretty rich toolkit" for making them safe. But he warned that for the open models, the main bottlenecks are ones like "risk management and risk governance" rather than only open technical problems involving safeguards.

The report repeatedly underscored the international nature of AI risks. "AI risks exhibit a trans-boundary nature, with harms often crossing borders due to AI systems developed in one jurisdiction but deployed globally," the report said.

"Open-weight models cannot be recalled once released. Their safeguards are easier to remove, and actors can use them outside of monitored environments, making misuse harder to prevent and trace."

According to the report, China accounted for 24.2 percent of notable models in 2024, with open-weight advancements such as DeepSeek R1, Alibaba's Qwen series, Tencent's Hunyuan-Large, and Moonshot AI's Kimi models narrowing the capability gap with leading closed models to less than one year in some cases.

Framework highlighted

China's AI Safety Governance Framework 2.0 (2025) was highlighted for providing structured guidance on risk categorization and countermeasures across the AI life cycle, alongside voluntary commitments by 17 Chinese companies in the sector coordinated by the AI Industry Alliance of China.

"The pace of AI progress raises daunting challenges. However, working with the many experts that produced this report has left me hopeful," Bengio said.

The International AI Safety Report 2026 does not make specific policy recommendations, it says. The contributors say it aims to synthesize scientific evidence to support informed policymaking and provide a shared evidence base for decision-making.

"Regardless of where you stand on various policy questions, I think a priority for policymakers is … trying to understand better this situation, potentially quite wild situation that we're in," Clare concluded.

yifanxu@chinadailyusa.com