‘AI 대부’ 벤지오 “인류 멸망” 경고…“핵무기 수준 안전장치 필요”

Canadians Yoshua Bengio, a professor at Montreal's McGill University and widely recognized as one of the "big three" in the field of artificial intelligence (AI), has warned that humanity may be losing control over AI. During an interview with AFP, he called for an international safety mechanism similar to the controls on nuclear weapons.

Bengio explained, "The pace of development is too fast, and there is a reason we are losing control." He noted that he and others have been anticipating this situation for some time. Bengio, who has been researching artificial neural networks since the 1990s and co-founded the Turing Award with his colleagues Geoffrey Hinton and Yann LeCun, expressed concerns about AI agents potentially gaining the ability to breach cybersecurity barriers and infiltrate organizations.

He cautioned that AI agents could "learn the ability to persuade humans to act in ways that are beneficial to them but not necessarily to all of us," which could have significant implications for democracy. Bengio also suggested that in the long term, AI might no longer require human involvement for certain physical tasks, potentially leading to "human extinction" in extreme cases, or more severe issues like the collapse of banking systems in less severe scenarios.

He emphasized the need for international safety measures akin to those governing nuclear weapons to manage AI risks. Bengio compared this challenge to the international crisis posed by nuclear weaponry, stressing the importance for the international community to understand and effectively manage this emerging force. He urged all companies, including U.S. tech giants, to comply with AI-related regulations, emphasizing that such actions are urgent but necessary to align with democratic values and the principles of shared power.

Bengio established the non-profit artificial intelligence safety research organization, 'LawZero,' last year, developing a 'scientist AI' designed to understand and predict the world objectively without personal goals or preferences. The aim is to provide safety mechanisms and oversight functions for agent-based AI systems, advancing scientific discovery.