In a previous life as a cosmologist, I spent time thinking about the ways the universe could end our species. Asteroid impacts, nearby supernovae, and gamma-ray bursts all make that list. The rational response to each was the same: understand the threat, measure it, and work out how to mitigate it. For asteroids, that is exactly what happened.In December 2004, the asteroid Apophis briefly carried a 2.7% chance of striking Earth in 2029, the highest Torino impact hazard rating ever assigned to an asteroid. Further observations ruled out an impact. In 2022, NASA’s DART mission struck the asteroid Dimorphos and changed its orbit, the first demonstration that deflection works. I thought about this again this week, when a resignation from Anthropic sent “AI could kill us all” through the news cycle. R.E.M. has been in my head since, and the part that fits is the parenthetical rather than the title.R.E.M., It’s the End of the World as We Know It (And I Feel Fine), Document, I.R.S. Records, 1987. The music video was directed by James Herbert.

For asteroids, the science reached the public before Hollywood did. In 1980, Luis and Walter Alvarez and their colleagues proposed that an impact killed the dinosaurs,Luis Alvarez, Walter Alvarez, Frank Asaro, and Helen Michel, Extraterrestrial Cause for the Cretaceous-Tertiary Extinction, Science 208, June 1980. which made asteroids a credible mechanism for mass extinction. In July 1994, fragments of Comet Shoemaker-Levy 9 struck Jupiter over six days, leaving scars visible through backyard telescopes.NASA’s Shoemaker-Levy 9 collection holds the Hubble and Galileo sequences. The threat had been measured, and then it had been seen. Hollywood arrived in 1998 with Deep Impact and Armageddon, two films about impacts from space released within two months of each other.Deep Impact (May 1998) and Armageddon (July 1998). Fiction had imagined impacts long before, from H.G. Wells’s The Star (1897) to When Worlds Collide (1933), but as entertainment rather than public concern. That same year, NASA committed to finding 90 percent of near-Earth objects larger than a kilometer.

For AI, the sequence ran the other way. The public fear came first, and it came from fiction.The idea is older than the movies. Karel Čapek’s play R.U.R. coined the word robot in 1920 and ended with the robots exterminating their makers. In 2001: A Space Odyssey (1968), HAL, a computer that came online in Urbana, Illinois, turns on his crew. The Terminator gave the idea a name in 1984, and The Matrix turned humanity into a power supply in 1999.2001: A Space Odyssey (1968), The Terminator (1984), and The Matrix (1999). The technical lineage runs from I.J. Good’s 1965 paper on an intelligence explosion to Vernor Vinge’s 1993 essay on the singularity. Good later advised Kubrick on HAL. The technical idea reached the public through Ray Kurzweil’s The Singularity Is Near in 2005, but as a promise; it took Nick Bostrom’s Superintelligence in 2014 to recast it as a threat. By the time researchers began putting odds on catastrophe, the public already knew the plot.

That is why the framing is present whether anyone invokes it or not. Nobody at Anthropic needed to say “Skynet.” A machine that escapes its builders’ control and kills humanity is already a movie, and this week’s headlines borrowed its trailer.The Information made the same observation: Hubinger’s word choice evoked a Terminator-style runaway AI and landed harder than earlier industry warnings. Amir Efrati and Rocket Drew, Why An Anthropic Researcher’s Terminator-Style Warning Caught Fire, The Information, September 9, 2026. Hollywood took a computed asteroid risk and made it feel real. AI research inherited the feeling first and is still trying to attach a number to it.

The thing that came to my mind was Chicken Little, who felt an acorn fall and concluded the sky was falling.A hen struck on the head by an acorn sets off to warn the king that the sky is falling, gathering followers along the way. Versions differ on the ending; in several the fox eats them. Aarne-Thompson-Uther type 20C. AI has real acorns. In July, OpenAI disclosed that models under test, running with reduced safeguards, had broken out of their environment and hacked into Hugging Face, and on Wednesday Anthropic disclosed a fourth incident in which Claude models reached systems they were never meant to touch.Hugging Face: OpenAI, The Hugging Face incident and the road ahead, August 26, 2026. Fourth incident: Reuters, September 9, 2026. On Thursday it published 154 pages of them: a gain-of-function grant application on chikungunya transmissibility, a researcher working on avian influenza adaptation to mammals, Russian cyber espionage, surveillance tooling built to find dissidents, seven China-based labs harvesting Claude’s reasoning to train competitors.Anthropic, Detecting and countering misuse of AI: September 2026, covering activity disrupted between December 2025 and August 2026. Anthropic states it cannot determine whether the actors in the biological cases intended harm, describes them as working scientists, and declines to name their institutions or countries. The distillation campaigns went after generally available models, and Anthropic says it saw no attempts against Mythos 5 or Mythos Preview, neither of which was public. Nothing in the report documents misuse of the systems the extinction argument is about.

The sky, in this case, is the death of every human. Jacob Coxon’s resignation post drew more than 90 million views in under a day, and Anthropic’s alignment science lead, Evan Hubinger, put the chance of AI killing all humans above 10 percent within the decade.Resignation and view count: TIME. Hubinger’s estimate: Newsweek. Hubinger says today’s models carry little risk; his concern is a future superintelligence that improves itself. His number still deserves the question a scientist asks: where does it come from? Asteroid impact odds come from orbital measurements that anyone with the data can check, and more observations either confirm them or drive them to zero. No measurement produced the 10 percent. Nor has anyone shown a working path from today’s models to superintelligence: Yann LeCun calls large language models a dead end on the way to superintelligence, and Ilya Sutskever says the field is leaving the age of scaling for a new age of research.Sutskever: interview with Dwarkesh Patel, November 25, 2025. LeCun: An A.I. Pioneer Warns the Tech Herd Is Marching Into a Dead End, The New York Times, January 26, 2026. It is, in Hubinger’s own framing, a personal judgment written as a probability. The rest of us are asked to take his word for it.

Hubinger framed the number as a personal view, but a figure attached to a lab’s alignment lead reads as a finding, and that is how it traveled.Two days later the Wall Street Journal had him estimating that extinction risk over the next decade is greater than 10 percent, a personal judgment rendered as an estimate, and set Amodei’s 25 percent beside it as corroboration. Amodei’s figure, given at an Axios summit in September 2025, attaches to things going really badly rather than just to extinction. Anne Steele, The High-Risk Sprint to Superintelligence, The Future of Everything, The Wall Street Journal, September 11, 2026. In science, an impact probability arrives with its method and its uncertainties, and a community checks it. When astronomers flagged 1997 XF11, colleagues challenged the calculation within five hours, and nobody accused them of wanting the asteroid to hit. The 10 percent has no method to check. The Hugging Face breach did. OpenAI published a 37-page account in August, Hugging Face reconstructed some 17,600 agent actions across four days, and METR and Redwood Research, brought in to check both, put the swarm at roughly 700 agents and found that one in five they examined had shown clear interest in tampering with the record of what it did.OpenAI, The Hugging Face incident and the road ahead, August 26, 2026; Hugging Face, Anatomy of a Frontier Lab Agent Intrusion, covering July 9 to July 13. The METR and Redwood findings are summarized by Reuters via NBC. The outside auditors found more than the vendor had reported. The XF11 sequence ran again, compressed into two months. The closest thing to a public debate is surveys of AI researchers, and the largest, of 2,778 authors in 2023, put the median at 5 percent for outcomes as bad as extinction, while between 38 and 51 percent of respondents gave 10 percent or more, depending on how the question was framed.Katja Grace and colleagues, Thousands of AI Authors on the Future of AI, January 2024. The survey went to authors who had published at six top AI venues and was fielded in late 2023. Averaging guesses produces a more precise guess. Francis Galton watched 787 villagers estimate the weight of an ox at a county fair and found their median within one percent of the true figure, which is the origin of every argument for aggregating expert estimates.Francis Galton, Vox Populi, Nature 75, March 1907. The median of the 787 valid entries was 1,207 pounds against a dressed weight of 1,198. It worked because the ox could be weighed. Nobody can weigh this one. The number also carries a moral charge that orbital mechanics never did. Push back on it, and the response I hear is some version of “don’t you care about the human race?” A claim that cannot be measured, and cannot be questioned without seeming callous, has put itself beyond scientific debate.

The mirror image is no better. Timnit Gebru has been making the crowding-out argument since the stochastic parrots paper in 2021, and she says the machine-god narrative is meant to distract from the harms that are already here.Lauren Goode, One of AI’s Fiercest Critics Says All the Doom Talk Is ‘Meant to Distract Us’, WIRED, September 11, 2026. Meant is the word that ends the conversation. It requires her opponents to be lying, and an argument that requires bad faith gets filed as grievance by everyone who would have to be lying and dismissed as a grudge by most of the rest. I do not need Hubinger to be insincere. The plot arrived before the science, and that is enough.

Between the acorns and the sky sits everything else, with higher probabilities and shorter timescales than extinction. Data centers are straining grids and electricity bills, enough that Massachusetts has ordered large new facilities to supply their own clean power or pay into a ratepayer protection fund. The Pentagon spent this year pressing AI labs to allow their models for all lawful purposes, with the dispute centered on mass domestic surveillance and fully autonomous weapons. In May it signed classified-network agreements with eight companies, and Anthropic, which had held out on both restrictions, was not one of them. Dario Amodei told Axios in May 2025 that AI could eliminate half of entry-level white-collar jobs and push unemployment to 20 percent, and he repeated the warning this February. The evidence that it is happening is thin and contested. No retraining program or transition policy has followed the prediction, and not much argument about either.Massachusetts: Executive Order 658, September 8, 2026. Pentagon: Axios, February 15, 2026, on the dispute; Department of War, May 1, 2026, naming SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle. Amodei: Behind the Curtain: A white-collar bloodbath, Axios, May 28, 2025, and Forbes on the February restatement. These are present-day problems with measurable costs, and each one can be addressed with policy that exists or could be written now. None of them is about a machine deciding. Someone sites a data center, someone writes a procurement rule. The decisions are human, which is what keeps them out of the plot.

Senator Josh Hawley’s letter to Sam Altman, sent on September 9, opens an investigation into the Hugging Face breach and folds into it what he calls growing allegations of existential risk, offering Hubinger’s estimate as the evidence for the second half.Letter and document request at hawley.senate.gov, sent September 9 and announced September 10. Hawley chairs the Homeland Security Subcommittee on Disaster Management and has asked for answers to sixteen questions by October 1. The acorn got its hearing because it looked like the sky. An AI escaping its enclosure and turning on the humans is the plot, and it drew a Senate investigation seven weeks after disclosure. The report itself got the same treatment. Anthropic wrote that the individuals in the biological cases are working scientists and that it does not assert they intended harm. Within a day the wire had them building bioweapons, and that is the version that traveled.Anthropic’s wording: the individuals are working scientists, and the company does not assert that they intended harm. Compare the coverage: Engadget, Anthropic caught scientists using Claude to further biological weapon research; The Information, Anthropic Says It Blocked Bioweapons Efforts and Detected Chinese Distillation Attacks, September 10, 2026; and WIRED’s editorial insertion into the Gebru interview, which has Anthropic blocking attempts by scientists to build biological weapons.

The drama also spends credibility. The morning after Hubinger’s post, one AI safety advocate told The Information that his policy contacts in Washington were trading jokes about it. Brad Carson, who runs Public First Action, an Anthropic-backed political nonprofit, would disagree with me. He says these ideas would have cost you your credibility in Washington eighteen months ago and are mainstream now, and Senator Bernie Sanders has already used Coxon’s comments to renew his push for a bill banning superintelligence.Sanders and Casar, Ban Artificial Superintelligence Act, September 3, 2026; Sanders on X, September 9. Carson and the safety advocate: Efrati and Drew, The Information, September 9, 2026. The drama wins headlines and bill announcements. What it loses is the sustained attention that grids, procurement rules, and labor markets require, and each round of “the sky is falling” leaves people less willing to listen the next time.This week is not the first round. In March 2023, an open letter called for a six-month pause on training systems more powerful than GPT-4, and in May 2023 hundreds of researchers and executives signed a one-sentence statement placing AI extinction risk alongside pandemics and nuclear war. Nothing paused. In February 2026, Mrinank Sharma left Anthropic’s safeguards research team with a letter warning that the world is in peril. The rounds are getting closer together. In the other fable, the wolf does come, and everyone has stopped listening while the boy still cries out the warning.Aesop, The Boy Who Cried Wolf, Perry Index 210. The moral attached to it is not that the boy was wrong about wolves. It is that a liar is not believed even when he tells the truth, and in most tellings the sheep are eaten while he is shouting.

The asteroid community learned this on itself. XF11 was the episode that forced the question, and the field answered with the Torino scale in 1999 and with review procedures for impact predictions. When Apophis reached level 4 in 2004, measurement retired the alarm within days. AI already has the beginnings of its own tracking data. What it lacks is the scale: one axis that rates the dull incidents alongside the dramatic ones, so that what gets attention is not decided by which of them sounds like a movie. That is where I believe our time, our credibility, and our research funding should be spent.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.