Ask ten AI researchers how likely it is that their own work eventually wipes out humanity, and you won't get ten similar answers. You'll get numbers ranging from "basically impossible" to "worse than a coin flip," sometimes from people who work two floors apart in the same San Francisco office building.
That gap isn't a rounding error. It's the entire debate.
Over the last three years, the question of superintelligence extinction risk has moved out of late-night forum threads and into congressional hearings, Nobel laureate open letters, and boardroom strategy decks at the very labs building this technology. Some of the people racing hardest to build advanced AI are also the ones putting the highest numbers on the risk that it kills us all. That contradiction alone is worth sitting with for a second.
This post pulls together 15 real, sourced numbers behind that debate. No guessing, no rounding up for drama, no cherry-picked outliers dressed up as consensus. Just what researchers, institutions, and surveys have actually put their names to, so you can see exactly where the disagreement comes from and decide for yourself how worried to be.
Table of Contents
- The 14.4% Average Nobody Explains Properly
- The 5% Median That Changes the Whole Story
- Anthropic's Own CEO Puts It at 10-25%
- Toby Ord's 1-in-6 Odds
- Metaculus Forecasters Land on 9%
- Eliezer Yudkowsky's 90%+ Estimate
- Yann LeCun's Near-Zero Counterpoint
- Roman Yampolskiy's 99%+ Outlier
- The American Public's 35.6% Average
- The 70% Number From a Room Full of Convinced People
- Only 21% Know What "Instrumental Convergence" Means
- 800+ Signatures on a Single Page
- Only 5% of Americans Want the Current Free-for-All
- 350+ Names Signed a 22-Word Warning
- 61% of Americans Already Think AI Threatens Humanity
1. The 14.4% Average Nobody Explains Properly
Start with the number sitting in the headline, because it deserves more context than most articles give it. In the 2023 Expert Survey on Progress in AI, researchers from the AI Impacts project reached out to thousands of published AI researchers and asked a very specific question: what's the probability that advanced AI leads to human extinction, or something similarly severe and permanent, within the next 100 years. A total of 2,778 researchers answered. The mean response came back at 14.4%.
This wasn't a Twitter poll or a self-selected group of doomers. It was one of the largest and most methodologically serious surveys of working AI researchers ever conducted, run by a group that has been tracking these numbers since 2016, according to the AI Impacts 2023 Expert Survey on Progress in AI. When people quote "AI researchers say there's a 14% chance of doom," this is almost always the survey being cited, even though the number rarely gets the context it needs to actually mean something.
📌 Related read: "Is AI Really Humanity's Last Invention? 12 Warnings From the Scientists Who Built It"
2. The 5% Median That Changes the Whole Story
Here's the part most coverage of that same survey conveniently leaves out. While the mean response was 14.4%, the median, meaning the answer sitting right in the middle once every researcher's guess is lined up from lowest to highest, was only 5%.
That's not a small difference. It's the gap between "researchers on average think there's a serious risk" and "half of researchers think the risk is fairly low, while a smaller group thinks it's extremely high." Statisticians call this a skewed distribution, and it happens whenever a handful of extreme answers pull the average away from where most people actually sit, a pattern confirmed in IEEE Spectrum's detailed breakdown of the AI Impacts survey results. In plain terms, most AI researchers are not walking around believing there's a 1-in-7 chance their own work ends humanity. A smaller, vocal minority is driving that number upward, and that minority happens to include some of the field's most respected names, which is exactly why their opinions carry outsized weight in the public conversation even though they don't represent the typical view.
3. Anthropic's Own CEO Puts It at 10-25%
What makes this debate genuinely strange is that some of the people building the most advanced AI systems on the planet are also the ones assigning it the highest personal risk estimates. Dario Amodei, CEO of Anthropic, a company whose entire business model depends on developing frontier AI models faster and more capably than competitors, has publicly placed his own estimate somewhere between 10% and 25% for things going "really, really badly," a figure reported directly from Amodei's own remarks, covered by Axios.
Sit with that for a moment. This isn't a critic or an outside academic making that estimate. It's a sitting CEO, actively raising billions of dollars to scale up the exact technology he's describing as carrying a double-digit chance of catastrophic failure. Anthropic has built much of its public identity around being the "safety-first" lab, releasing detailed alignment research and pushing for external regulation even while competing aggressively for market share, and Amodei's willingness to state a specific number publicly is part of that positioning.
📌 Related read: "Brother vs. Brother: 15 Explosive Moments in the Anthropic-OpenAI AI Civil War"
4. Toby Ord's 1-in-6 Odds
Not every serious estimate comes from inside an AI lab. Philosopher Toby Ord, an Oxford researcher known for his work on existential risk, spent years building a systematic risk model across every major threat to humanity's long-term survival, from nuclear war to engineered pandemics to climate collapse. He published his conclusions in his book The Precipice: Existential Risk and the Future of Humanity, estimating the overall probability of human extinction by the year 2100 at roughly 16.7%, or about 1 in 6.
What stands out in Ord's model is how much of that total risk he attributes specifically to unaligned artificial intelligence. Of his overall 16.7% figure, roughly 10 percentage points come from AI alone, according to the breakdown referenced in AEI's analysis of the economics of AI existential risk. That makes AI the single largest contributor in his entire risk model, ahead of nuclear weapons, climate change, and pandemics combined, which is a striking claim from someone whose entire academic career is built on comparing risks like these against each other rigorously rather than reacting to headlines.
5. Metaculus Forecasters Land on 9%
Surveys of researchers are one thing. Prediction markets are another, and they work differently. Metaculus is a well-established public forecasting platform where thousands of participants, many with strong track records, make and continuously update predictions on real-world questions, with their performance scored over time. It's the same style of platform intelligence agencies and research institutions increasingly study because aggregated forecaster judgment tends to outperform individual expert opinion.
On the specific question of human extinction or near-extinction by 2100, the Metaculus community estimate sits at around 9%, with AI-related causes accounting for roughly 8 of those 9 percentage points, according to figures cited in AEI's coverage of prediction-market extinction risk estimates. In other words, forecasters who spend their time evaluating and re-evaluating this exact question see AI as almost the entire story when it comes to existential risk this century. Nuclear war, pandemics, and climate change combined barely register next to it in their models.
📌 Related read: "AI 2027: The Predictions Everyone's Talking About"
6. Eliezer Yudkowsky's 90%+ Estimate
At the far high end of the spectrum sits Eliezer Yudkowsky, a researcher who has been writing about AI alignment since long before it was a mainstream concern. Yudkowsky, co-author of the 2025 book If Anyone Builds It, Everyone Dies, has publicly stated his own probability estimate for AI-caused extinction at above 90%, a figure documented on Wikipedia's page tracking public P(doom) estimates.
Yudkowsky's position isn't a passing comment. He has spent years arguing that once a sufficiently capable AI system exists, humanity will not get a second chance to correct course, a thesis he laid out publicly as far back as his 2023 essay calling to halt advanced AI training, referenced in the same Wikipedia summary of his public statements. He's become one of the most quoted, and most disputed, voices in this entire debate precisely because his number sits so far from the researcher median covered in fact two.
7. Yann LeCun's Near-Zero Counterpoint
On the exact opposite end sits Yann LeCun, Meta's former chief AI scientist and one of the three researchers who won the 2018 Turing Award for foundational work in deep learning. LeCun has publicly put his own estimate at below 0.01%, according to figures compiled in CompaniesHistory's research roundup on AI existential risk statistics, arguing that today's systems, no matter how impressive they look in a demo, are nowhere near the kind of autonomous, goal-directed capability that a genuine extinction scenario would require.
LeCun and Geoffrey Hinton, once close collaborators who shared that same 2018 Turing Award, have since become public opposites on this exact question, a rift covered in detail in The New Yorker's profile of the AI doomsayer movement. Two researchers with nearly identical technical backgrounds and a shared career-defining award, landing on estimates that differ by several orders of magnitude, tells you almost everything about why this debate refuses to settle.
8. Roman Yampolskiy's 99%+ Outlier
If Yudkowsky sits at the high end of the researcher spectrum, Roman Yampolskiy sits past it entirely. Yampolskiy, an AI safety researcher and director of the Cyber Security Laboratory at the University of Louisville, has spent much of his career studying what he calls "AI containment," the question of whether a sufficiently advanced system could ever be reliably controlled once it exists. His own conclusion is stark: he publicly estimates the probability of AI-caused catastrophe at above 99%, a figure listed alongside other named researcher estimates in CompaniesHistory's compiled research on AI takeover statistics.
What makes Yampolskiy's number worth including isn't that it's the most extreme figure in this list, it's why he arrives at it. His argument isn't rooted in a specific doomsday scenario the way pop culture usually frames this fear. It's rooted in control theory, the branch of engineering and mathematics concerned with whether a system can reliably constrain another system that may become more capable than the constraint itself.
Yampolskiy's position is essentially that the control problem has never been mathematically solved for a system smarter than its controller, and until it is, the default outcome trends toward catastrophe rather than away from it. Agree or disagree, it's a technical argument, not a gut feeling, which is part of why it keeps getting cited in academic AI safety circles even by researchers who land on very different numbers themselves.
📌 Related read: "The AI Arms Race: How Dangerous Is It Really?"
9. The American Public's 35.6% Average
Every number so far has come from researchers, forecasters, or philosophers who study this professionally. But ordinary people have opinions too, and it turns out the general public is considerably more worried than most AI researchers are. A public opinion survey asked respondents directly: what's the probability that AI technology eventually leads to the extinction of all human life, on a scale of 0 to 100. The mean response came back at 35.6%, more than double the AI Impacts researcher mean from fact one, with a median of 20% and 31% of all respondents placing their personal estimate above 50%, according to Survey160's published research on public p(doom) beliefs.
That gap between expert and public opinion is worth pausing on. It's rare to find a technical topic where the general public is significantly more pessimistic than the specialists building the technology, rather than the other way around. Usually it's the reverse, insiders see risks the public doesn't fully grasp yet. Here, something closer to a third of ordinary respondents believe there's a coin-flip or worse chance that AI ends humanity, a level of concern that outpaces even the more pessimistic named researchers covered earlier in this list.
10. The 70% Number From a Room Full of Convinced People
Not every number in this debate comes from a random or representative sample, and this one is worth including precisely because of that caveat. In March 2026, Harvard hosted a public event built around the book If Anyone Builds It, Everyone Dies, featuring a talk, a two-sided conversation, and an audience Q&A on AI existential risk. Researchers surveyed the 89 matched attendees both before and after the event.
The post-event median estimate for the probability of AI-caused extinction or severe disempowerment came back at 70%, with 96% of attendees agreeing that mitigating AI existential risk should be treated as a global priority, according to the published study documenting the Harvard survey results.
That 70% figure will almost certainly get quoted out of context somewhere, so it deserves an honest caveat right here. These were self-selected attendees at a talk specifically built around a book arguing superhuman AI would kill everyone, and their pre-event median was already 50% before the talk even started.
This isn't a representative snapshot of public opinion, it's a measurement of how a specific persuasive event moved the views of people who showed up already leaning concerned. Still, the fact that a room of educated, engaged adults can be pushed to a 70% median in a single evening says something about how persuasive the doom case can be once it's laid out in detail, even if it doesn't tell us what a random cross-section of the public actually believes.
11. Only 21% Know What "Instrumental Convergence" Means
Here's a number that complicates almost everything above it. Despite 78% of the public agreeing that AI risk deserves serious attention, only 21% could correctly identify "instrumental convergence," a foundational concept in AI safety research referring to the prediction that sufficiently advanced systems tend to pursue self-preservation, resource acquisition, and goal-preservation as sub-goals regardless of what their original objective actually was, according to survey data compiled in CompaniesHistory's research on public AI risk literacy.
That gap between concern and comprehension matters more than it looks. It means most of the public conversation happening around superintelligence extinction risk, including a lot of what gets shared and argued about online, is happening without a shared technical vocabulary for what the actual concern even is.
People are worried, correctly or not, but relatively few could explain in a sentence why a smarter-than-human AI system pursuing a harmless-sounding goal might behave in ways nobody intended. That's not a criticism of the public, these are genuinely technical ideas that took AI safety researchers years to formalize. It's just a reminder that "78% think this is important" and "21% understand the core mechanism" can both be true about the same group of people at the same time.
📌 Related read: "Big Tech CEOs' AI Job Warnings"
12. 800+ Signatures on a Single Page
Surveys measure opinion. Open letters measure willingness to put your name on the record, which is a different and arguably higher bar. In October 2025, the Future of Life Institute released its "Statement on Superintelligence," a text barely thirty words long that called for a prohibition on developing superintelligent AI until there is broad scientific consensus that it can be done safely and controllably, along with strong public buy-in. Within days, it had drawn over 800 signatories, according to CNBC's reporting on the superintelligence ban statement.
What makes this list genuinely unusual isn't the number itself, it's who's on it. Geoffrey Hinton and Yoshua Bengio, the two "godfathers of AI" who between them helped build the deep learning techniques every modern AI system relies on, both signed. So did Apple co-founder Steve Wozniak, Virgin Group's Richard Branson, UC Berkeley computer scientist Stuart Russell, five Nobel laureates, former Joint Chiefs of Staff chairman Mike Mullen, and a politically scattered mix of public figures ranging from Prince Harry and Meghan Markle to Steve Bannon and Glenn Beck. It's rare to find a single petition that pulls signatures from that wide a political and cultural spread, which is part of why it made headlines well outside the usual tech press.
📌 Related read: "The AI Arms Race: How Dangerous Is It Really?"
13. Only 5% of Americans Want the Current Free-for-All
Buried inside the reporting on that same superintelligence statement is a smaller number that's arguably more telling than the signature count itself. The Future of Life Institute cited internal polling showing that just 5% of U.S. adults support the current status quo, meaning fast, largely unregulated development of superintelligent AI, according to the same CNBC coverage of the FLI petition.
That 5% figure cuts against a common assumption in this debate, that public disagreement over AI risk falls somewhere close to a coin flip. It doesn't, at least not on the specific question of pace. Regardless of whether someone's personal p(doom) estimate looks more like the 5% median from AI Impacts researchers or the 70% median from the Harvard event attendees, an overwhelming majority of ordinary Americans appear to agree on one narrower point: the current speed of development, with minimal external oversight, isn't something most people are comfortable with.
That's a much easier thing to build political consensus around than an actual extinction probability, which may be exactly why organizations pushing for regulation keep leading with it.
14. 350+ Names Signed a 22-Word Warning
Rewind further back, to May 2023, before most of today's headline AI models even existed. The Center for AI Safety published what might be the single most consequential sentence in this entire debate: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Just twenty-two words, released with no elaboration, no policy proposal, and no specific timeline attached.
At launch, it was signed by more than 350 leading researchers and executives, according to AI Wiki's documented history of the Center for AI Safety statement. The signatory list included Geoffrey Hinton and Yoshua Bengio again, but also something rarer, the sitting CEOs of three of the most prominent and directly competing AI labs on Earth at the same time: Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Dario Amodei of Anthropic.
These are companies locked in an aggressive, well-documented race for market share, funding, and talent, and yet their chief executives found common ground on a single sentence about their own industry's worst-case outcome. CAIS deliberately kept the statement short, according to its own press release, specifically so that a broad coalition who disagreed on almost every detail of AI risk could still agree on the headline claim.
📌 Related read: "Big Tech CEOs' AI Job Warnings"
15. 61% of Americans Already Think AI Threatens Humanity
The final number on this list predates almost everything else covered here, and that's exactly what makes it worth including last. In its press release accompanying the May 2023 statement, the Center for AI Safety cited a poll finding that 61% of Americans already believed AI threatens humanity's future, well before ChatGPT-era systems had matured into anything close to today's capability level, and years before the Future of Life Institute's superintelligence petition or the Harvard survey covered earlier in this post, according to the official CAIS press release.
That timing matters. Public unease about AI's long-term risk didn't emerge as a reaction to any single dramatic headline or viral incident. It was already sitting at a clear majority before most of the numbers in this entire article existed. Every survey, petition, and CEO statement that's followed since has been layered on top of a public that was already more than 60% convinced something worth worrying about was happening, which helps explain why each new data point tends to land on receptive ears rather than a skeptical blank slate.
📌 Related read: "Google vs. OpenAI: Inside the AI War"
Frequently Asked Questions (FAQ)
What is "superintelligence extinction risk" exactly?
It refers to the possibility that a future AI system, one that substantially exceeds human intelligence across most or all cognitive domains, could act in ways that cause irreversible harm to humanity, up to and including extinction, either through misaligned goals, loss of human control, or unintended consequences of pursuing an otherwise reasonable-sounding objective.
Is there an actual scientific consensus on the risk percentage?
No, and that's the central point of this entire article. Estimates from named, credentialed researchers span from below 0.01% to above 99%. The most-cited large-scale researcher survey shows a mean of 14.4% and a median of 5%, but no single number represents anything close to a field-wide consensus.
Why do researcher estimates vary so widely?
Largely because "p(doom)" isn't a standardized metric. Different researchers define the outcome differently, some include only literal human extinction, others include permanent loss of human autonomy or control even without physical extinction. Disagreements also stem from which historical risks or technological analogies a given researcher finds most persuasive when reasoning about an unprecedented scenario.
Do the CEOs of AI companies actually believe their own products are dangerous?
Several have said so publicly and put specific numbers behind it, including Anthropic's Dario Amodei at 10-25% and signatories of the 2023 CAIS statement, which included the CEOs of OpenAI, Anthropic, and Google DeepMind. Whether that stated concern changes their companies' actual development pace is a separate and more contested question.
Should I be personally worried about this?
That's a judgment call this article can't make for you, but it's worth knowing the full range of credible opinion rather than just the loudest voice on either side. The researcher median sits far lower than most viral headlines suggest, while the public and some individual experts sit considerably higher. Forming your own view means weighing both ends honestly.
Final Thoughts
Fifteen numbers, and not one of them agrees with the others. That's not a flaw in this article, it's the actual state of the field. Superintelligence extinction risk isn't a settled scientific question with a single accepted probability the way something like vaccine efficacy or climate sensitivity increasingly is. It's an open, contested estimate made by very smart people using very different assumptions, arguing about a technology that doesn't fully exist yet.
What does seem to hold steady across nearly every number in this list, from the AI Impacts researcher median to the Future of Life Institute's 5% status-quo support figure, is a general discomfort with the current pace of development, even among people who disagree sharply on the actual odds of catastrophe. Whether that discomfort translates into meaningful global coordination, the kind these open letters are explicitly asking for, remains the real open question behind every one of these numbers.
Curious how this connects to the broader AI safety and lab rivalry landscape? Check out our full "AI Wars" series covering the Anthropic-OpenAI civil war, the global superintelligence arms race, and the researchers sounding the alarm from inside the labs themselves.