AI

Experts have called for stronger efforts to build trustworthy artificial intelligence (AI), saying reliability and effective governance will become key factors in global AI competition amid the technology's rapid development.

The remarks were made at an AI parallel forum of the 2026 World Laureates Forum, which was recently held in Shanghai. During the parallel forum, scientists and industry representatives discussed how to address emerging challenges in AI development and application, the Science and Technology Daily reported Monday. 

He Jifeng, an academician of the Chinese Academy of Sciences and professor emeritus at East China Normal University, said AI should never become an opaque "black box," emphasizing that trustworthiness must be integrated into the entire research and development process rather than remaining an abstract principle.

He highlighted the importance of explainability, verifiability and traceability, as well as the use of formal methods and alignment technologies to build safety mechanisms into the knowledge-understanding and risk-mitigation stages of AI usage.

The need to integrate safety into AI development from the outset is becoming increasingly pressing as the technology advances, Dong Yinpeng, chief scientist of Chinese third-party AI safety company Fangcun AI, told the Global Times in an interview in Beijing.

One of the biggest bottlenecks in current AI safety research is its traditional reliance on a reactive approach, in which researchers identify vulnerabilities after a product or system has been developed and then work to fix them, Dong said.

"This approach is becoming increasingly inadequate," he told the Global Times, noting that AI is evolving so rapidly that its vulnerabilities are not static. Even after existing weaknesses are addressed, new risks may emerge as models become more capable.

"More importantly, the potential consequences of AI-related security failures are growing. Some models have already demonstrated autonomous attack capabilities," Dong said, warning that further advances could potentially pose risks to real-world critical infrastructure.

Waiting until damage occurs before addressing vulnerabilities could therefore result in significant losses, he noted.

Beyond technical security, the practical adoption of AI also faces governance and implementation challenges, including poor quality of data, difficulties integrating AI into existing business operations and challenges in quantifying the value generated by AI projects. 

The gap between rapidly advancing technological capabilities and relatively slower regulation has become a major constraint on the wider adoption of AI, experts warn. Trustworthy AI, therefore, should not be reduced to a one-time compliance exercise, but should involve oversight throughout the entire lifecycle of a system, from research and development to deployment and operation.

Chiming in at the forum, Noga Alon, a 2024 Wolf Prize in Mathematics laureate and a professor at Princeton University, said AI has demonstrated strong capabilities in solving mathematical problems, but is basically using existing methods and has not truly invented new ones, the Science and Technology Daily reported.