Who's Sounding the Alarm
Brief portraits of the most prominent voices in the public debate about AI risk — the doomers, the optimists, and the people inside the tent who left because they were worried about what was happening.
The public conversation about AI safety is shaped by a small group of people. Some of them invented the technology. Some of them have spent their careers studying it. Some of them have left major companies because they thought the people running those companies were not taking the risks seriously enough. They do not all agree. Several of the people on this page disagree sharply with each other in ways worth understanding.
This page covers the most prominent voices. It is not exhaustive. The aim is to give readers enough context to follow the debate when they encounter these names in news coverage, and enough video to hear them in their own words rather than filtered through reporting.
Geoffrey Hinton — the Godfather of AI
Geoffrey Hinton (born 1947, British-Canadian) is the figure most often referred to as the "godfather of AI". The label is earned: with David Rumelhart and Ronald Williams, he co-authored the 1986 paper that introduced backpropagation, the mathematical technique by which neural networks learn from their mistakes, and the algorithm that essentially everything modern in deep learning still relies on. With his graduate students Alex Krizhevsky and Ilya Sutskever, he produced AlexNet in 2012, a convolutional neural network (a kind of AI specialised for understanding images) whose victory in the ImageNet competition (an annual contest where AI systems compete to recognise objects in photographs) is widely cited as the start of the modern deep-learning revolution. He shared the 2018 Turing Award (computing's Nobel) with Yann LeCun and Yoshua Bengio. In 2024 he was awarded the Nobel Prize in Physics jointly with John Hopfield.
In May 2023, Hinton publicly resigned from Google to speak freely about what he saw as the dangers of the technology he had helped create. The departure was a turning point in the public conversation. A figure of his standing saying he was worried — that AI could be an existential threat, that he regretted some of his life's work, that humanity might not be able to keep control of systems smarter than itself — could not be dismissed as a fringe opinion. His subsequent interviews have been widely watched.
The interview below is one of his most-watched long-form conversations on the topic. He sets out, in plain terms, what he is worried about, why, and what he thinks should be done.
Geoffrey Hinton — the Godfather of AI on what he is most worried about.
Hinton's specific concerns are that AI systems will reach human-level capability sooner than most people expect (his current estimate is sometime in the next twenty years), that the people building these systems are not taking the alignment problem (the technical question of how to make AI systems reliably do what humans actually want, rather than what we accidentally told them to do) seriously enough, and that economic and political incentives are pushing development faster than safety work can keep up. He has put a probability of around 10-20% on AI causing human extinction within thirty years. That number is contested — many serious AI researchers think it is too high — but the fact that someone of Hinton's credibility is willing to say it has shifted what is sayable in public.
Yoshua Bengio — the safety pivot
Yoshua Bengio (born 1964, Canadian) shared the 2018 Turing Award with Hinton and LeCun. He runs Mila, the Quebec AI Institute in Montreal, which is one of the most-cited deep-learning research centres in the world. His own academic publications are among the most-cited in the field.
Bengio's pivot toward safety has been more recent and more striking than Hinton's. Through most of his career he worked on AI capabilities; in 2023, around the same time as Hinton's Google departure, he publicly stated that the rapid progress of large language models (LLMs — the kind of AI behind ChatGPT and Claude, trained to predict text on vast amounts of writing) had changed his timeline and that the safety community had been right to take superintelligent AI (AI substantially smarter than humans across the board, sometimes called superintelligence) seriously sooner than the mainstream had.
He chaired the 2024 International Scientific Report on the Safety of Advanced AI commissioned by the UK Government. The report, with around 100 contributors from 30 countries, is the closest thing to a scientific consensus document on what is and is not known about advanced AI risks. The 2025 update was published before the Paris AI Action Summit and is the most authoritative single document on the topic.
Bengio's framing differs from Hinton's in tone — less colloquial, more careful — and from Yudkowsky's in conclusion — Bengio thinks safe AI is achievable with sufficient effort, where Yudkowsky thinks the problem is not solvable on the current trajectory. The disagreement is substantive and worth understanding.
His major recent project is LawZero, a non-profit founded in 2025 to develop "scientist AI" — AI systems oriented toward safe, verifiable scientific reasoning rather than open-ended agency. The thesis is that the most useful path through the next decade may be AI as research tool rather than AI as autonomous agent.
Yoshua Bengio's TED Talk — what changed his mind on AI safety, and what he thinks the world should do next.
Stuart Russell — the textbook author
Stuart Russell (born 1962, British) is Professor of Computer Science at UC Berkeley. He is co-author, with Peter Norvig, of Artificial Intelligence: A Modern Approach — the textbook that almost every undergraduate computer-science student in the world has used to learn AI for the past three decades. He has been working on AI safety since long before it was fashionable, and his 2019 book Human Compatible: AI and the Problem of Control is one of the most accessible serious treatments of the topic.
Russell's 2017 TED talk below remains one of the clearest fifteen-minute introductions to the AI control problem. The framing — that the standard model of AI (give it an objective, let it pursue the objective) breaks down for sufficiently capable systems, and the solution requires a different model where AI is uncertain about human preferences and has incentive to learn them — has become foundational in the safety field.
Stuart Russell — "3 principles for creating safer AI" — TED2017. Fifteen minutes; the cleanest introduction to the AI control problem.
Russell's tone is deliberately unsensational. He does not put numerical probabilities on extinction. He argues that the question of whether AI will become an existential risk is less interesting than the question of how to build AI that is provably aligned with human interests, and that the latter is a tractable engineering problem if it is taken seriously. His more recent work has focused on autonomous-weapons regulation — he has been a leading academic voice arguing for an international ban on lethal autonomous weapons.
Eliezer Yudkowsky — the most extreme voice
Eliezer Yudkowsky (born 1979, American) is the most-quoted source for the strongest version of the AI doom argument. He has no formal academic credentials and is not part of the mainstream AI research community. He founded the Machine Intelligence Research Institute (MIRI) in Berkeley in 2000 and has spent twenty-five years arguing that building artificial general intelligence (AGI — AI that can do, in general, anything a human can do intellectually) will most likely kill everyone.
His position is that any sufficiently capable AI optimising for any goal that is not exactly aligned with human values will, by default, take actions whose side-effects are catastrophic for humanity. The technical problem of getting alignment exactly right, in his view, is enormously harder than the technical problem of building capable AI, and we are doing the second much faster than we are making any progress on the first.
The 2023 TED talk below is the cleanest fourteen-minute version of his argument. It generated significant controversy at the time — Yudkowsky's framing was much more extreme than what TED audiences are used to — but the talk has been viewed several million times and is cited in essentially every popular treatment of AI doom.
Eliezer Yudkowsky — "Will superintelligent AI end the world?" — TED2023. Fourteen minutes; the most-circulated short version of the doom argument.
Yudkowsky's call to action is what he calls "Pause AI" — a worldwide moratorium on training AI systems beyond a certain capability threshold, enforced internationally, with the willingness to use military force to enforce it. The position is more extreme than almost anyone else on this page would endorse. It is included here because it is the most rigorously-argued version of the strongest version of the case, and because the public conversation about AI risk would not be the shape it is without him.
Yann LeCun — the prominent skeptic
Yann LeCun (born 1960, French) is Chief AI Scientist at Meta and the third co-recipient of the 2018 Turing Award alongside Hinton and Bengio. His foundational work was on convolutional neural networks — the architecture that powers essentially all modern computer vision. He is not on this page as a doomer. He is on this page because the debate is incomplete without his voice.
LeCun's position is that the existential-risk framing is fundamentally misconceived. He argues that current large language models, despite their fluency, lack the capacity for the kind of long-horizon planning, world-modelling and goal-pursuit that would be required for them to pose a meaningful risk to humanity. His view is that the path to genuine general intelligence requires architectural advances beyond LLMs that we have not yet figured out, and that even when we do, the engineering challenge of giving any such system goals dangerous to humanity is much larger than the doom-leaning side acknowledges.
He has been critical of what he sees as fearmongering by his colleagues and has argued that AI safety regulation could lock in the dominance of a small number of large firms — what economists call "regulatory capture", when the rules end up serving the interests of the incumbents that helped write them. He is one of the most prominent advocates for open-source AI as a counterweight to that risk; Meta's release of the Llama family of open-weight models (AI models whose internal numbers are made public so anyone with the hardware can download and run them on their own computer) reflects his thinking.
The disagreement between LeCun and Hinton — three Turing Award co-winners, two of whom think the technology they invented is an existential risk and one of whom thinks that fear is overblown — is one of the more striking facts of the contemporary AI debate. Reasonable, knowledgeable people disagree about this. The disagreement is not resolved.
Yann LeCun — the case against the existential-risk framing, from the most prominent skeptic in the field.
Demis Hassabis — the laureate
Demis Hassabis (born 1976, British) is the co-founder and CEO of DeepMind and a 2024 Nobel Laureate in Chemistry for his work on AlphaFold, an AI system that predicts the three-dimensional shape of proteins from their underlying chemistry. Working out protein shapes used to take biology PhDs months of laboratory work each; AlphaFold does it in seconds and has done it for almost every known protein. The free database that resulted is the closest thing modern AI has to a public good. Hassabis is the rare case of an AI lab leader whose lab has produced unambiguously beneficial science alongside the more contested capabilities work.
His public position on AI risk is moderate-pessimistic. He has signed the 2023 Centre for AI Safety statement that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war". He is in favour of international coordination on the most powerful models. He has consistently argued that we are likely to see human-level AI within a decade or so and that the science of alignment is genuinely unsolved.
Where he differs from Hinton or Bengio is in his belief that the risks are manageable with sufficient effort, that the benefits are likely to be enormous (he points to AlphaFold as a worked example), and that the answer is to build the technology carefully rather than not at all. His framing is that we are entering "the most consequential technological transition in human history" and that getting it right is the central problem of his generation.
Demis Hassabis — Nobel Laureate and DeepMind co-founder on what's coming and how to build it carefully.
Dario Amodei — the inside-the-tent voice
Dario Amodei (born 1983, American) is the co-founder and CEO of Anthropic, formerly VP of Research at OpenAI. He left OpenAI in 2021 with several other senior researchers to found Anthropic specifically because he was concerned that the largest AI labs were not prioritising safety highly enough. Anthropic's stated mission is "AI safety research at the frontier". Claude — discussed throughout the rest of this site — is Anthropic's product.
Amodei's 2024 essay Machines of Loving Grace is the most coherent published statement of an inside-the-tent position. The argument: powerful AI is coming, probably this decade; it is likely to compress decades of biomedical, scientific and economic progress into a few years; the benefits could be transformative; the risks (concentration of power, misuse, loss of control) are real and have to be managed deliberately. He has called for international coordination on frontier models (the most powerful AI systems being built at the cutting edge of the field) comparable to nuclear-weapons-era arms control.
Whether this position is genuinely safety-first or commercially convenient (Anthropic, like OpenAI, is a multi-billion-dollar AI company) is something readers will form their own view on. His public statements have been consistent over time, and the technical safety work Anthropic publishes is taken seriously by the academic safety community.
Dario Amodei — Anthropic CEO on building AI safely at the frontier, and why he left OpenAI to do it.
Ilya Sutskever, Jan Leike and the OpenAI safety departures
Through 2024-2025, a series of senior researchers left OpenAI citing safety concerns. The most prominent were Ilya Sutskever (a co-founder of OpenAI, Hinton's former student, and the lead author on much of the technical work that produced GPT-2, GPT-3 and GPT-4) and Jan Leike (head of OpenAI's superalignment team — the unit dedicated to keeping smarter-than-human AI on humanity's side — departing in May 2024 with a public statement that "safety culture and processes have taken a backseat to shiny products"). Leike subsequently joined Anthropic. Sutskever founded Safe Superintelligence Inc, a new lab whose stated and only product is safe superintelligent AI — no consumer chatbots, no incremental commercial releases, just the long bet.
The pattern of departures is meaningful regardless of one's prior on the doom debate. The people closest to the technology, with the most professional reason to be measured in their public statements, have been leaving the company they helped build because they think the path it is on is unsafe. That is not proof of anything by itself, but it is information.
Ilya Sutskever — co-founder of OpenAI, founder of Safe Superintelligence Inc — on what he sees coming, and why he left.
How to read the debate
If you are coming to this debate fresh, three things are worth knowing.
The disagreement is real, not theatrical. Hinton and LeCun are both Turing Award co-winners, equally credentialed, looking at the same technology, reaching opposite conclusions about the existential-risk question. Neither is being dishonest; both are being thoughtful. Anyone telling you the answer is obvious in either direction is selling something.
Probability statements are estimates, not measurements. When you read that Hinton thinks there is a 10-20% chance of AI causing extinction, that is a personal probability based on his judgement, not a calculation. Different experts give different numbers; the spread is wide because the underlying uncertainty is wide.
The doom debate is not the only safety debate. Most of this site is about how AI is being deployed today — and most of the harms documented in the When AI Goes Wrong page have already happened, to real people, without any superintelligent AI being involved. The present-tense risks of opaque algorithms, biased systems and unaccountable automated decisions are happening now. The future-tense risks of superintelligent AI may or may not be real. Both deserve attention; do not let the second crowd out the first.
The honest takeaway is that we are at a moment in this technology where the people who know it best are publicly divided about how worried to be. That is a reasonable thing for the public to know, and to factor into how they think about it.