Yann LeCun, Chairman of AMI Labs, stated that artificial intelligence is simply not smart enough to threaten humanity. Speaking in Seoul during an interview with JoongAng Ilbo, he addressed recent concerns regarding AI doom. He explained that current mainstream large language models (LLMs) cannot predict the consequences of their actions or understand how those impacts affect humans—something even a ten-year-old child easily grasps.
As a New York University professor, LeCun received the 2018 Turing Award alongside Yoshua Bengio and Geoffrey Hinton for pioneering deep learning research. After leaving his role as Meta’s Chief AI Scientist, he founded AMI Labs in January. His goal is to develop world models—AI systems that learn from actual physical environments to predict subsequent actions, differing significantly from language-centric LLMs. Emphasizing safety, LeCun noted that while lawnmower accidents claim about 100 lives annually in the U.S., severe accidents remain exceedingly rare even with billions of daily AI interactions, making AI far safer than a lawnmower. He visited South Korea to attend the Industrial AI Global Cooperation Conference.
Addressing recent AI hacking incidents in South Korea, LeCun attributed them to corporate negligence. He stated that developers lacked sufficient security experience when deploying AI assistants, and vulnerabilities existed within isolation environments like sandboxes. He stressed that these were cybersecurity issues rather than failures to control AI.
Regarding instances of AI misbehavior, LeCun asserted that engineering solutions can easily resolve them. He suggested designing objective-driven AI equipped with guardrails that activate instantly upon command execution. Such systems would prevent rogue behavior beyond human-set goals, much like programming a robot holding a kitchen knife with a rule prohibiting arm movements when people are nearby.
When asked if this approach applies to current AI models, LeCun explained that LLMs struggle with individual fine-tuning for every user command, leading to frequent deviations. He noted that OpenAI CEO Sam Altman emphasizes slowing down releases precisely because models often lack adequate fine-tuning.
Challenging AI doom claims by Bengio and Hinton, LeCun disagreed with assumptions that AI will surpass humans. He maintained that current AI merely aggregates and repeats human-accumulated knowledge, remaining largely unglamorous except in strictly verifiable domains like mathematics and coding.
On the topic of real AI dangers, LeCun pointed to excessive market monopoly. He warned that if a tiny group of corporations controls the systems through which most people acquire information, global cultural and linguistic diversity will vanish. He criticized Anthropic CEO Dario Amodei and others who control information pathways for promoting obscurantism by claiming AI is inherently dangerous and must be controlled—comparing their stance to Ottoman rulers banning printing presses.
Finally, contrasting world models with LLMs, LeCun concluded that world models will become significantly smarter by predicting action outcomes and will remain easier to control through objective-driven designs, proving that rising intelligence does not equate to uncontrollable technology.
