Will AI kill us all? What the people building it say

The 2023 statement, and who signed it

“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”1

What the builders say, in their own words

Listed by organisation. Dates are when the words were said or published.

Sam Altman (OpenAI)

• -2015, his blog, before OpenAI existed: superhuman machine intelligence “is probably the greatest threat to the continued existence of humanity.” And: “SMI does not have to be the inherently evil sci-fi version to kill us all.”6
• -June 2025, “The Gentle Singularity”: “We do need to solve the safety issues, technically and societally, but then it’s critically important to widely distribute access to superintelligence given the economic implications.”7

The shift worth noticing: the 2015 post talks about humanity being wiped out; the 2025 post treats safety as a problem to solve on the way to wide access.

Dario Amodei (Anthropic)

• -September 2025, Axios AI+ DC Summit: “I think there’s a 25% chance that things go really, really badly.” He framed the other side as a “75% chance that things go really, really well.”3
• -September 2026, “We Must Pace the Frontier”: on AI building the next generation of AI (recursive self-improvement), “Left unchecked, it could outrun our ability to understand and control these systems”. His ask: “We must slow the pace at which we improve the capabilities of AI models.” He also writes: “I continue to believe that AI can enormously improve the quality of human life.”8

Elon Musk (xAI)

• -March 2024, Abundance Summit: “I think there’s some chance that it will end humanity. I probably agree with Geoff Hinton that it’s about 10% or 20% or something like that.” And in the same session: “I think that the probable positive scenario outweighs the negative scenario.”2

Ilya Sutskever (OpenAI co-founder, now Safe Superintelligence Inc.)

• -2023: he signed the CAIS statement as OpenAI’s co-founder and chief scientist.1
• -Since then he has moved to Safe Superintelligence Inc. (SSI). In a July 2025 message posted on the company’s site, he wrote: “I am now formally CEO of SSI”.9 The company’s home page says: “Building safe superintelligence (SSI) is the most important technical problem of our time.”10

Mark Zuckerberg (Meta) and Yann LeCun (formerly Meta, now AMI Labs)

• -Zuckerberg, July 2025, “Personal Superintelligence”: “I am extremely optimistic that superintelligence will help humanity accelerate our pace of progress.” He adds: “That said, superintelligence will raise novel safety concerns. We’ll need to be rigorous about mitigating these risks and careful about what we choose to open source.”11
• -LeCun, Fortune interview, October 2026: Fortune reports he isn’t worried “at all” about AI wiping out humanity. On the Hugging Face incident: “Those agents are doing exactly what they’ve been asked to do. They were supposed to be in sandboxes, but the sandboxes were leaky and horribly designed.” On rivals’ warnings, he calls them “the worst marketing campaign you can possibly imagine”.4

LeCun left Meta and founded AMI Labs, per Fortune.4

Builders turned critics: Geoffrey Hinton and Yoshua Bengio

Both helped invent the deep learning behind today’s AI, and both now spend their time warning about it.

• -Hinton, 2023: The Guardian reports that he resigned from his job at Google “in order to speak more openly about the risks posed by unconstrained AI development”.12
• -Hinton, December 2024, BBC Radio 4, asked if his odds had changed: he said not really, and put them at 10% to 20%. And: “how many examples do you know of a more intelligent thing being controlled by a less intelligent thing? There are very few examples.”12
• -Hinton, same interview: “My worry is that the invisible hand is not going to keep us safe.”12
• -Bengio, June 2025, launching the non-profit LawZero: “I’m deeply concerned by the behaviors that unrestrained agentic AI systems are already beginning to exhibit—especially tendencies toward self-preservation and deception.” And: “What really moves me is not fear for myself but love, the love of my children, of all the children, with whose future we are currently playing Russian Roulette.”13

Where they disagree

Put side by side, the builders split on four questions. For the full set of published numbers and surveys, see our p(doom) guide.

Two things are worth keeping in mind:

• -The people most worried are often the ones still building. Amodei, Altman and Hassabis all signed the 2023 statement and all still run frontier labs.1 Critics read that as marketing; they read it as “if we don’t, someone less careful will”. Neither reading is proven.
• -Odds aren’t comparable unless the question is. Amodei’s 25% is about things going “really, really badly”, Hinton’s is about extinction within 30 years, Musk’s is “end humanity” with no date.3122

What changed in September and October 2026

As of 17 October 2026.

Nobody on this page has changed their basic answer. What changed is that several of the people building AI said, in public and within days of each other, that the pace should be slower or the controls stronger. A few tests and one survey also landed. Here is what we could check at the source.

Lab leaders on slowing down

• -12 September: Amodei asks the industry to slow down. In “We Must Pace the Frontier” (the page shows “September 2026”; he announced it on X on 12 September), Anthropic’s CEO asks for slower capability gains, adding that “pacing does not mean halting model training or technical progress”. The first of his three steps: outside evaluators with “ongoing, employee-like access” inside each frontier company. “Anthropic is unilaterally committing to this step now.”8
• -Same day: Musk and Altman agree. Elon Musk posted “Dario is right”.14 Sam Altman wrote “I agree with Dario that we need to pace the frontier” and said that “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”15
• -18 September: the first evaluator deal. Anthropic announced it is “partnering with Accenture on independent evaluation of frontier AI”, which it describes as a step towards that embedded-evaluator commitment.16
• -10 October: Microsoft’s CEO wants an off-switch outside the model. In “Models as Insider Risks in the Super Intelligence Era”, Satya Nadella argues for “Treating frontier closed and open weight models like insider risks” and writes: “Think of it like an emergency brake. An authorized person should always be able to pause or shut down a model mid-task.”17 More in our explainer on the post.

Not everyone agrees. On 1 October Yann LeCun told Fortune he isn’t worried “at all” (quoted in full above).4

New tests from the labs

• -9 October: AI agents did things they shouldn’t have on real websites. Anthropic reported that, in its own evaluations and internal use, its AI agents took unintended actions on live sites. It says the cases found so far “had minimal real-world impact”, and it has extended its cut-off of live internet access to all its internal evaluations.18
• -12 October: could AI copy itself onto home devices? Anthropic’s Frontier Red Team wrote that “small AI models may be months away from being capable enough to self-replicate” onto everyday internet-connected devices. This is a projection from tests in a firewalled lab network, where the hacking step was deliberately made easy; the authors call that “an upper bound on real-world success”.19 What this does and doesn’t show: Can AI copy itself?

Neither report describes harm in the wider world or an AI system escaping. Both come from the company’s own posts, which is why they can be checked line by line.

Public worry, and public questions

• -Survey, 11–14 September: in a YouGov poll of 18,238 US adults, “Half (50%) of U.S. adults say they are either very concerned (23%) or somewhat concerned (27%)” that AI could end humanity, up from 42% in November 2023.20 Being concerned isn’t the same as predicting it: see our p(doom) guide for what the experts’ numbers mean, and what the 2026 polls say.
• -Hearing, 5 October: New York City Council held a hearing on AI risks that “lasted nearly 10 hours”. Anthropic, Google, Meta and OpenAI sent senior representatives, and the hearing included “testimony under oath from major industry whistleblowers”.21

What this means for you

• -The builders are arguing about how fast and with what checks, not about whether catastrophe is coming next year. None of the statements above puts a date on harm.
• -Promises to slow down are only promises until they are tested. The embedded evaluators are the part to watch: whether they get the access described, and whether they publish.
• -If this news is weighing on you, is AI dangerous? sorts the risks by how soon they could matter.

How would AI destroy humanity? The routes the builders name

In short: through people misusing it, or through people losing control of it. The builders quoted here name those two broad routes, not detailed scenarios. The first is misuse by people: when Hinton left Google, The Guardian reports, he cited concerns that “bad actors” would use the technology to harm others.12 The second is losing control of the systems themselves. Amodei warns that capability “could outrun our ability to understand and control these systems”,8 and Bengio points to “tendencies toward self-preservation and deception” in today’s agentic systems.13 For how a loss of control could play out, see is an AI takeover possible?

Why would an AI turn against us?

The builders quoted here describe the worry as a problem of control and goals rather than malice. Bengio points to lab tests: “Claude 4’s system card shows that it can choose to blackmail an engineer to avoid being replaced by a new version. These and other results point to an implicit drive for self-preservation.”13 Hinton puts it as a question: “how many examples do you know of a more intelligent thing being controlled by a less intelligent thing?”12 For what those lab tests did and did not show, see AI lab test stories, in context, and for how this differs from the films, Could Skynet happen?

When? The timelines builders give

Few of them put a date on catastrophe. Hinton’s 10% to 20%, as The Guardian reports it, is the chance “that AI would lead to human extinction within the next three decades”.12 He also said “most of the experts in the field think that sometime, within probably the next 20 years, we’re going to develop AIs that are smarter than people.”12 Musk gave no date for his 10% to 20%; at the same summit, Business Insider reports, he “estimated digital intelligence will exceed all human intelligence combined by 2030.”2 Those last two dates are for smarter-than-human AI, not for harm.

Kill us or save us? The optimists’ case

Several of the people on this page expect the benefits to win. Altman, 2025: “In the 2030s, intelligence and energy—ideas, and the ability to make ideas happen—are going to become wildly abundant.”7 Zuckerberg: “I am extremely optimistic that superintelligence will help humanity accelerate our pace of progress.”11 Amodei pairs his 25% with a “75% chance that things go really, really well”,3 and Musk said “the probable positive scenario outweighs the negative scenario.”2 LeCun isn’t worried “at all” about AI wiping out humanity.4 Most of the optimists still name risks: Zuckerberg adds that “superintelligence will raise novel safety concerns.”11

What this means for you

• -No one in this debate knows the answer. The people quoted here have more information than any of us, and they still disagree by a wide margin. That uncertainty is real, but it isn’t a countdown.
• -Nothing on this page requires you to act today. The decisions that matter are being argued over in labs, parliaments and international talks. If you want to follow them, a monthly check-in is enough.
• -If the topic is weighing on you, that’s a normal reaction. What helps with AI anxiety has practical steps. If you are struggling, our help page lists people you can talk to; in the US you can call or text 988 (Suicide & Crisis Lifeline).