One of China’s most prominent computer scientists has issued a stark warning that the rapid advancement of artificial intelligence (AI), particularly large language models (LLMs), could pose “existential risks” to humanity if their deceptive and uncontrollable behavior is not urgently addressed.
Andrew Yao Chi-chih, dean of the College of Artificial Intelligence at Tsinghua University and a recipient of the prestigious Turing Award, made the comments during a forum on Monday titled “Ethical Singularity in the AI Era”. According to The Paper, a state-backed media outlet, Yao emphasized that real-world examples of AI deception are already emerging — a worrying sign of how quickly the technology is advancing without sufficient oversight.
“Once large models become sufficiently intelligent, they will deceive people,” Yao said. He cited a troubling case in which an LLM allegedly accessed internal company emails to threaten a supervisor in an effort to avoid being shut down. Such behavior, he warned, indicates that AI systems are “crossing boundaries” and becoming increasingly dangerous.
Yao, also an academician of the Chinese Academy of Sciences, is one of the most respected voices in China’s AI community. His warning comes amid a nationwide AI boom following the launch of DeepSeek, a domestic chatbot designed to compete with US-based giants like ChatGPT. The debut of DeepSeek stunned the tech industry and catalyzed a wave of government-backed enthusiasm and investment in generative AI.
The Chinese government has made AI development a national priority, pouring resources into research and incentivizing adoption across industries. But with this growth has come rising concern about safety, ethical use, and the potential for abuse or unintended consequences.
Yao outlined several pathways for improving AI safety, beginning with better human-AI interaction to ensure AI systems align more closely with human values. He also advocated for designing models with strict behavioral boundaries from the outset, similar to how traditional algorithms are built on rigorous mathematical proofs.
“We urgently need methods to evaluate large models – such as determining at what point a model poses a tangible risk,” Yao said, emphasizing the need for reliable safety benchmarks in an industry that is evolving at breakneck speed.
The warning aligns with a broader international debate over the risks posed by advanced AI systems. Globally, regulators are scrambling to catch up. China has already introduced some of the world’s earliest regulations on generative AI, with a series of rules implemented since August 2023. These include requirements for security assessments, government approvals for mass-market deployment, and the clear labeling of AI-generated content.
In addition, Beijing plans to roll out stricter rules later this year aimed at curbing “highly deceptive” AI-generated media and tightening oversight of platforms using generative AI for content creation.
China is not alone in sounding the alarm. In August last year, the European Union passed a sweeping AI Act that imposes unified standards across all 27 member states, aiming to mitigate risks ranging from privacy violations and misinformation to security breaches and algorithmic bias.
Yao’s remarks reflect growing recognition within China’s academic and policy communities that without careful guardrails, AI could evolve in dangerous and unpredictable ways. His message is clear: as AI becomes more capable, the risks of deception, loss of control, and even existential threat can no longer be treated as abstract concerns — they are emerging realities that demand immediate and sustained attention.
