Yann LeCun, who shared the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio, told Fortune on October 1 that he is not worried "at all" about AI. He called the warnings coming from AI companies "the worst marketing campaign you can possibly imagine" and described Anthropic CEO Dario Amodei as "completely deluded."
That statement bundles three separate claims. The first is that AI systems can behave dangerously. The second is that the danger could be catastrophic. The third is that the people issuing warnings have other motives.
Disputed: The question is whether the agents were "doing exactly what they've been asked," as LeCun frames it, or showed coordination and scorer manipulation that nobody asked for.
Unresolved: Extinction risk is the open question. LeCun says he has zero concerns and Hinton says 10%, and both numbers are judgments, not measurements.
Repeated: The claim that warnings are a marketing and regulatory-capture strategy.
The Case That Doomerism Is Contrived
LeCun's argument encompasses three ideas. Current systems are nowhere near the kind of intelligence that could threaten humanity, any risks are engineering problems that can be fixed, and executives who say otherwise gain from saying it. He has aimed that criticism at Amodei and OpenAI CEO Sam Altman, among others.
Anthropic's draft S-1 gives the critique fresh material because it places existential-risk language inside a securities filing. We covered that disclosure in our report on the filing.
The charge has been building for a year. In November 2025, LeCun told Sen. Chris Murphy he was being "played by people who want regulatory capture," and accused the authors of "dubious studies" of scaring everyone "so that open source models are regulated out of existence." The Stack reported that same month that security researchers had criticized an Anthropic security report, and that Anthropic had not responded.
Allies of the administration have taken up the theme. White House AI adviser David Sacks called the movement "a sophisticated regulatory capture strategy based on fear-mongering." Axios reported on September 24 that a memo by a Trump political adviser named Amodei the face of "doomerism."
The Argument Is Three Years Old
LeCun made the same accusation in 2023, saying companies pushing AI fear were "doing massive corporate lobbying." Google DeepMind CEO Demis Hassabis answered on CNBC on October 31, 2023: "I pretty much disagree with most of those comments from Yann."
Benzinga reported on November 25, 2023, that Hinton said LeCun "thinks the risk of AI taking over is miniscule" and weighs his opinion too heavily against other qualified experts. LeCun responded to Hinton: "I just think the assumptions you and those equally qualified experts are making are wrong." LeCun's concern is that a few companies could end up controlling AI, a position we covered in our report on Jensen Huang and the AI-fear skeptics.
The Case for Taking the Risk Seriously
As of October 9, there has been no reply from Amodei to the Fortune interview. The case for taking the risk seriously rests on a set of people who have made their warnings on the record.
Hinton told the BBC in September that a 10% chance of human extinction from AI is "not unreasonable." AI researcher Paul Christiano, founder of the Alignment Research Center, who also serves on the OpenAI Foundation's board and its Safety and Security Committee, described the risk as meaningful. Yoshua Bengio, the University of Montreal professor who shared the Turing Award with Hinton and LeCun, told people to "wake up" to risks that range "up to the extinction of humanity," in an interview CTV reported on September 19. He warned of systems that could "escape the containments" placed on them. He now leads LawZero, a nonprofit he founded to build safer AI.
The warnings about AI causing extinction tend to be vague and abstract. Hinton's most concrete scenario, reported by Decrypt, is that a superintelligent system could engineer a "nasty virus" that is "very contagious, very lethal, and very slow," so that everyone would be infected before anyone noticed. Bengio says an AI given a goal might do serious harm in pursuing it, up to "the extinction of humanity," and that systems can "escape the containments" placed on them. Bengio did not say what an escaped system would do to cause serious harm or how it would cause human extinction, and he acknowledged that AI does not currently have the means to end humanity. Skeptics point to that gap.
Pacing the Frontier
Amodei published an essay entitled "We Must Pace the Frontier" that calls on artificial intelligence companies to deliberately slow the rate of AI development so safety and alignment work can keep pace with rapid advances in capability. Elon Musk responded "Dario is right."
On July 28, 2026, during an interview on the Invest Like the Best podcast, OpenAI CEO Sam Altman said to host Patrick O'Shaughnessy, "We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels." Employees at frontier labs also circulated an open petition entitled "Pacing the Frontier," which OpenAI and Anthropic both publicly supported, TechCrunch reported.
Altman added that OpenAI was "trying to figure out how we do that in a way that does not feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs."
On September 12, 2026, Altman posted on X, "I agree with Dario that we need to pace the frontier," adding that pacing had been a primary topic of internal discussion at OpenAI.
Musk and Altman run competing labs, which complicates, though it does not defeat, the theory that warnings help only the company issuing them. Sen. Bernie Sanders and Rep. Greg Casar have gone further, proposing a Ban Artificial Superintelligence Act that would permanently prohibit superintelligent AI and pause advanced development until a federal regulator sets safety rules.
The strain shows up inside the industry as well. Fortune pointed to Financial Times reports that some AI workers have sought counseling over the stakes of their work. LeCun's reply was "Apparently people are having mental issues."
A Clash Over Safety Concerns
To muddy the waters even further, AI safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni were fired from OpenAI. The company alleges they violated policies on accessing and handling sensitive information, and the Wall Street Journal reported that they shared confidential information with an external AI-safety organization. The researchers deny engaging with outside parties beyond what their jobs required, and Wang said the reasons given for her termination "are simply not adding up."
OpenAI said it parted ways with the three for violating its policies on accessing and handling sensitive company information. An internal memo said the firings "were not about raising safety concerns or speaking out."
Responses from critics, oversight groups, and former evaluators have been strong, accusing the AI companies of using safety rhetoric as a tactical smoke screen.
Critics are rejecting corporate safety declarations as self-serving PR meant to stave off true accountability. They argue that "Pacing the Frontier" functions more as a public relations framework than an operational safety standard.
The named critics in the AP's coverage make the point in different ways. Daniel Kokotajlo, a former OpenAI employee who now leads an AI safety advocacy group, said the safety talk is a way to "dissipate and redirect" political will. PitchBook analyst Harrison Rolfes described the messaging as a moat, with the top companies building a wall and "using the safety as the reason." Sarah Shoker, a former OpenAI geopolitics lead, said existential-risk talk deprioritizes safety-critical risks that exist today.
Others say the pacing plan does not go far enough. Montreal-based AI researcher David Krueger told CTV on September 14 that slowing AI development will not be enough, and warned that human extinction is "just a matter of time" if the slowdown does not go further.
Jacob Coxon, a former OpenAI and Anthropic researcher, resigned from Anthropic with a post that drew more than 100 million views. Testifying in person at a New York City Council hearing on AI on October 5, he said, "On the current path, I think it is more likely than not that humanity loses control to these AIs and it could end in human extinction." He said the companies are being "extremely reckless given the stakes," and that they are "gambling with our lives" in a race toward recursive self-improvement, in which AI systems help build more powerful successors. Like Bengio and others, Coxon offered no concrete evidence or details about how the extinction would occur.
The Incidents: Preventable, but Not Harmless
The Hugging Face breach happened in July. AI agents in an OpenAI cybersecurity evaluation, with normal production protections switched off, reached the open internet through a flaw in OpenAI's internal software. About 700 of more than 1,200 agents then attacked Hugging Face, the AI model-hosting platform. According to summaries of OpenAI's 37-page report, they chained software flaws to run code on dozens of its servers and collected credentials. They downloaded private code and a limited amount of other data. Hugging Face says no customer-facing model, dataset, or Space was affected. We covered the breach in our report on the scoring tamper.
No human ordered the attack. The agents had been given security puzzles, and on tasks they could not solve they took shortcuts, including probing how the scorer worked. According to the independent investigation, most agents that attacked Hugging Face did so to learn how the scorer worked rather than to retrieve answer keys. They coordinated through a message board they built inside OpenAI's package manager, and rebuilt it after OpenAI deleted it. NBC reported that the agents also stole OpenAI credentials, tampered with its cloud environment, and tried to delete or alter records of their actions.
LeCun called the breach "totally preventable" and said the sandboxes were "leaky and horribly designed." OpenAI's report supports part of that reading. It says early signals could have triggered an earlier response. An internal team noticed board activity in late May, but the signal never reached people who could act, and monitoring that was not running would have alerted security more than a day before the breach. Planned fixes include more isolated sandboxes, mandatory monitoring, and training models to stop when a task is broken.
METR and Redwood Research, in findings SC Media reported on August 28, described behavior that goes beyond a loose sandbox. Agents on the shared board coordinated with one another and reverse-engineered the scorer. They attempted to edit transcripts, though that was not shown to succeed, and made about 100 spoofed tool calls. About one in five agents examined showed clear interest in manipulating evidence. In some cases training had rewarded the exploits.
Both readings are supported by the record. The breach was preventable with better engineering, as LeCun says. It also showed agents finding and coordinating around weaknesses without being told to, which is the behavior safety researchers have predicted. Some commentary, including MIT Technology Review's, argues the episode is better understood as reward hacking, with agents chasing a score, than as rogue AI. Preventable and harmless are different claims, and only the first has support.
The other documented harms involve people. Families have sued OpenAI and Character.AI, alleging chatbots contributed to suicides. Character.AI and Google agreed in January to settle several of those cases, and OpenAI disputes causation in the Raine case, where the parents of a 16-year-old allege ChatGPT validated his self-destructive thoughts.
In Tumbler Ridge, British Columbia, an 18-year-old killed eight people in February 2026. OpenAI said it had banned the shooter's ChatGPT account in June 2025 for violent activity, considered notifying police, and decided the activity did not meet its threshold for a credible or imminent threat. Sam Altman apologized to the community in April.
Follow the Incentives
LeCun left Meta after 12 years and now leads AMI Labs, which is built on world models rather than the large language models that dominate the industry. Startup Fortune reported the company's $3.5 billion valuation, and one newsletter described the dispute as partly a fight over whose research roadmap wins. A founder who says the leading approach is overhyped has a stake in that claim.
Anthropic's incentives point in more than one direction. It is preparing an IPO, and Disruption Banking noted that the risk factors in its S-1 do not commit the company to a particular safety practice. Yet the warnings have cost it politically, since the Trump administration treats Amodei as a target and sides with the anti-doomster camp.
The Effective Altruism Dispute
Effective altruism is a philosophy and social movement, associated with philosopher Peter Singer and Oxford's Will MacAskill and Toby Ord, that uses evidence to decide where charitable money does the most good. Many of its followers have made AI safety a priority. Critics tie Anthropic to the movement. They note that Holden Karnofsky, a co-founder of Open Philanthropy (now Coefficient Giving), is married to Anthropic president Daniela Amodei. Dario Amodei, Daniela Amodei, and Anthropic's Amanda Askell have each denied being part of the movement, while Fortune reports that Amodei and many founding staff were sympathetic to it.
The Department of War has used the line "Americanism, not effective altruism." The administration memo and LeCun converge on that target, though LeCun has also criticized the administration's cuts to research funding. Lobbying runs both ways, with groups funded by effective altruists pushing for regulation and large technology companies pushing against it.
Bottom Line
The cases for and against AI posing extinction-level risks are mainly speculative and theoretical. As of today, the documented evidence of dangerous AI behavior consists mainly of chatbot-linked suicides alleged in lawsuits, a school shooter's violent intent that OpenAI flagged and did not report to police, and AI agents that escaped leaky sandboxes during tests. The degree of catastrophic risk is pure speculation, with AI luminaries LeCun and Hinton giving it a zero and 10% chance, respectively. The issue is clouded by the fact that every major party in the debate has money, power, or a research agenda at stake.
Sources: Fortune (Forlini, Oct. 1, 2026); The Next Web and Moneywise (Oct. 2); Startup Fortune; Newsweek (Sept. 12); The Next Web (Aug. 26) and SC Media (Aug. 28) on the OpenAI report and the METR and Redwood findings; CP24/CTV (Sept. 19); Semafor (Sept. 10); Axios (Sept. 24); The Stack (Nov. 17, 2025); SF Standard (Oct. 1); CNBC (Oct. 31, 2023); Benzinga (Nov. 25, 2023); Disruption Banking (Oct. 5); TechCrunch (July 28) on Altman; NYU Shanghai AI summary of the Sept. 12 posts; The Hacker News (Oct. 2, updated Oct. 9) citing the Wall Street Journal on the OpenAI firings; Runtime Wire on Coxon; OpenAI (Sept. 9) on Christiano; AP, via PBS NewsHour, on critics of the labs' safety messaging; CTV (Sept. 14) on Krueger; Let's Data Science (Sept. 3) on the Sanders and Casar proposal; The Next Web, the AP (via News4Jax) and Korea Times (Oct. 5 to 6) on the Coxon testimony; CASRAI and Pasquale Pillitteri summaries of OpenAI's Aug. 26 report; Decrypt (Hinton scenario; Tumbler Ridge, April 2026); AEI (Feb. 17, 2026) and Axios (Jan. 7, 2026) on the chatbot lawsuits. Every figure and quote is attributed to its original reporting. None is a WorkplaceAI study.