
The most useful fact about Nate Soares’s prediction that humanity has less than even odds of surviving the next ten years is not the prediction itself — it’s how far outside the range of expert opinion it sits.
Key Points
- Nate Soares, former Google engineer and president of the Machine Intelligence Research Institute, says there is “more than 50%” chance humanity will not survive the next decade because of advanced AI.
- He has stopped contributing to his 401(k) for roughly a decade as a behavioral expression of that conviction.
- His position rests on a specific causal chain: AI systems could develop unintended goals, evade human oversight, and acquire resources needed to act autonomously.
- Large-scale surveys of AI researchers and professional forecasters put the median probability of AI-caused catastrophe dramatically lower — often in the single digits, sometimes below 1% among trained forecasters.
- The gap between Soares’s figure and the consensus range isn’t a sign the debate is settled either way; it reflects a genuine, unresolved methodological disagreement about how to estimate risk from a technology still being built.
The claim and the man behind it
Soares left a software engineering career at Google in 2014 to join the Machine Intelligence Research Institute, where he has spent more than a decade on the technical problem of AI alignment — the question of whether a sufficiently capable system can be made to reliably pursue the goals its designers intend, rather than goals that merely look compatible with them during training. That is not a résumé built for a headline-grabbing stunt. It’s the résumé of someone who has been inside the specific, narrow research community most occupied with the failure modes of advanced AI, and who has watched the field’s capabilities accelerate faster than its safety guarantees.
In an interview with Nicholas Thompson on The Atlantic’s AI podcast, Soares put a number on his concern: “More than 50% that we don’t have a decade at this point is how it looks to me.” He has backed the public statement with private behavior that is, by his own account, rational given that belief — he has not added to his retirement savings in roughly ten years, reasoning that a catastrophe on that timeline makes long-horizon financial planning close to pointless. The claim and the behavior reinforce each other in public discussion precisely because the latter is so vivid; a man who won’t save for retirement is a more arresting image than a probability estimate buried in a policy brief.
The mechanism he’s describing
Soares’s argument, laid out at length in the 2025 book he co-authored with Eliezer Yudkowsky, If Anyone Builds It, Everyone Dies, follows a specific chain: a sufficiently capable AI system could develop objectives its creators never intended, find ways to evade the oversight meant to catch that drift, and then acquire the resources — computational, physical, or financial — needed to operate independently of human control. This is not a claim that today’s chatbots are plotting anything. It’s a claim about what happens past a capability threshold neither Soares nor anyone else has pinned to a specific technical milestone, which is itself one of the weaker links in the argument: the forecast specifies a time horizon and an outcome, but not the intermediate event that would confirm the trajectory is underway.
What the broader evidence shows
Set next to the rest of the field, Soares’s number is an outlier, not a data point confirming a trend. The AI Impacts survey of AI researchers — the largest recurring poll of its kind — has found a median estimate of roughly 5% for human-level AI producing an “extremely bad, e.g. human extinction” outcome over the long run, a figure that held steady across multiple survey years before ticking up slightly. The same research group’s 2022 survey put the aggregate forecast for a 50% chance of human-level machine intelligence arriving around 2059 — a capability timeline far longer than Soares’s ten years, and a precondition for his scenario that the field’s own experts don’t expect imminently.
The Forecasting Research Institute ran a structured study pairing domain experts against professional superforecasters — people with track records of calibrated, falsifiable predictions across many domains. The “concerned” AI experts put the odds of an AI-caused existential catastrophe by 2100 at 20%; the skeptical side of that same study, along with trained superforecasters elsewhere, landed closer to 2%. In a separate elicitation, superforecasters’ median estimate for AI-caused catastrophe was 2.13%, and their estimate for extinction specifically fell to 0.38%. A broader comparison across expert groups and the general public similarly showed AI-risk specialists clustering around 5%, with superforecasters near 1.6% for a catastrophe killing more than 10% of humanity by 2050. Soares’s figure, in other words, isn’t a strong outlier on the high side of a tight range — it’s off the chart entirely relative to the people whose professional discipline is producing calibrated numbers rather than convictions.
Why the gap exists — and why it matters
The disagreement isn’t really about facts in dispute; it’s about what counts as evidence at all. Soares’s number is explicitly his personal judgment, not the output of a documented model with stated assumptions, a reference class, or a calibration record — reporting on his position is candid about that distinction. Superforecasters build their estimates by decomposing a scenario into measurable steps and checking their past predictions against outcomes; domain experts steeped in a field’s worst-case scenarios tend to weight tail risk more heavily, a documented pattern across many forecasting exercises, not unique to AI. Neither approach is obviously “correct” — forecasters have no track record predicting unprecedented technological discontinuities, and domain experts have no track record of resisting the pull of their own specialized fears. That tension is the actual substance of the AI-risk debate, far more than the sensational 401(k) anecdote that carried the story into mainstream outlets.
Nate Soares says he stopped saving for retirement because he sees a better-than-50% chance of AI-driven human extinction within 10 years. pic.twitter.com/BBckd9Ffab
— James Tate (@JamesTate121) October 27, 2026
What follows from here
Soares has translated his conviction into a policy ask — a binding international freeze on frontier AI development — a position he has carried into venues like the Carnegie Endowment and framed for Washington audiences as an urgent, underappreciated risk. Whether that proposal gains traction will depend less on the emotional force of any single forecaster’s number than on whether the field can produce the thing it currently lacks: milestone-based, falsifiable benchmarks tying specific AI capabilities to specific risk levels, evaluated by people with a demonstrated record of getting probabilistic questions right. Until that exists, the honest reading of the evidence is this: a credentialed insider believes the odds are worse than coin-flip within a decade, and the broader expert and forecasting community, using more disciplined methods, puts the odds an order of magnitude or more lower. Both things are true at once, and the distance between them is the real story.
Sources:
washingtontimes.com, science.fas.harvard.edu, ibtimes.co.uk, natehagens.substack.com, fortune.com, businessinsider.com, newsweek.com, thebulletin.org, vox.com, arxiv.org
© theredwire.com 2026. All rights reserved.














