Yann LeCun (얀 르쿤), chairman of Ami Labs, pushed back against concerns that AI will escape human control. He said current ways of ensuring the safety of large language models (LLMs) have limits, but new AI designs such as world models and technical safety devices can raise the odds of keeping systems under control.

LeCun made the remarks on Oct. 12 at a press briefing titled “Industrial AI Global Cooperation” held at the Westin Josun Seoul in Jung-gu, Seoul, when asked about AI risks and whether control is possible.

"The scenario in which someone discovers the secret to implementing human-level intelligence and robots suddenly take over the world is closer to science fiction," he said. "Technological progress is a gradual process in which multiple innovations and breakthroughs accumulate."

He also pointed to system design issues, rather than AI itself, in recent AI agent security incidents.

"Recent incidents happened because companies that deployed AI agents were careless," he said. "It is a problem of not properly applying already known technologies such as sandboxes, and it can be sufficiently resolved," he added.

By contrast, he said mainstream LLM-based AI systems are structurally difficult to control.

"It is realistically impossible to train for every possible question and tune the system to give only non-dangerous answers," he said. "It is relatively easy to induce LLM-based AI to generate unintended answers."

As an alternative, he presented “objective-driven AI” using world models. The approach is designed so AI predicts the consequences of its actions and behaves according to pre-set goals and safety constraints.

LeCun co-founded AI startup Ami Labs at the end of 2025 and began developing world models. World models are AI technology that learns the real world and predicts the results of actions, aiming to overcome existing LLMs’ limitations in understanding physical environments and controlling actions.

"If we can design constraints that guarantee safety, we can make AI systems intrinsically safe," he said. "It does not mean it is an easy engineering problem, but in the end it is something to be solved through engineering."

He compared securing AI safety to designing aircraft engines. Even complex and high-risk technology can achieve a high level of safety through sophisticated design and verification, he said.

He acknowledged that world model-based AI is also not completely free from risk. "If AI predicts the real world incorrectly, an action judged to be safe could in fact be dangerous," he said.

In cybersecurity, he said AI is likely to work more in favor of defenders than attackers.

"Evidence so far shows that AI works more in favor of defenders than attackers," he said.

He said defenders have advantages in manpower, funding and information sharing, while attackers must operate covertly and face limits on cooperation. He added that there is uncertainty about long-term effects.