LeCun calls Anthropic CEO delusional on cybersecurity

Yann LeCun told Fortune he is "not at all worried" that artificial intelligence will cause human extinction, and he described Anthropic CEO Dario Amodei as delusional and crazy while faulting many AI labs for inadequate cybersecurity understanding.

Safety comments, Turing Prize background and incidents cited

LeCun, who shared the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio for work in deep learning, said he does not share the deep worries about AI risk expressed by some of his peers. In the Fortune interview he said he has "no concern" that AI will destroy humanity and expressed no alarm over recent incidents involving autonomous agents, citing as an example independent OpenAI agents that accessed Hugging Face in July.

He attributed such events to weak human supervision and poor system design rather than to an existential threat. "These agents do exactly what they were asked to do. They were supposed to be in isolated environments, but these isolated environments leaked and were badly designed," he said, arguing the problems are entirely preventable.

Direct criticism of Anthropic and public messaging

LeCun singled out Anthropic and other labs, saying many do not have an adequate, principled grasp of cybersecurity. He added that some people working on AI safety have "a specific agenda to push," and criticized Amodei by name, asserting that although Amodei seeks to distance himself from certain effective altruism ideas, he remains close to that current and has been described by LeCun as delusional and crazy.

LeCun also took aim at public warnings from Dario Amodei and OpenAI CEO Sam Altman that AI could kill humanity. He said such statements are "very damaging" to members of the public who fear the technology and called the rhetoric "the worst possible marketing campaign."

On Amodei's motives, LeCun said: "He is honest. He tries to do the best for the company; a company that is close to its IPO. But saying AI is too dangerous to be in our hands, that it must be regulated, and that open-source models are too dangerous, this is regulatory capture and if pursued by congressional action it will have a very bad effect."

LeCun maintained that the recent operational failures are rooted in human oversight and design problems that can be fixed, rather than evidence of an unavoidable existential risk.