A researcher’s viral resignation post, a survey putting the average AI researcher’s extinction estimate near one in five, and a poll showing most Americans now see AI as a meaningful threat to humanity have combined into what blogger Zvi Mowshowitz calls a “preference cascade”: a sudden, self-reinforcing shift in what people are willing to say in public. Mowshowitz, writing on his Substack Don’t Worry About the Vase on September 18, argues the shift is genuine. He also argues it will not save anyone on its own.
The trigger, in his telling, was Jacob Coxon, a researcher whose resignation post went viral and pulled other departures into public view. That included Joe Benton, who departed Anthropic’s safety group during late August and moved to the AI research organization METR, plus Bilal Chughtai, who exited Google DeepMind after working there on alignment research. Coxon later got a Wall Street Journal profile out of it. Mowshowitz credits commentator Mike Solana’s read on why the moment landed: enough Americans already had context on AI to care, and enough large accounts amplified the post before it could fade.
Two data points anchor the piece. The AI Impacts survey, the longest running poll of AI researchers, put the average respondent’s estimate of AI causing extinction or comparably severe and permanent human disempowerment at roughly 18 percent, with Asian researchers reporting higher risk estimates than Western ones. Politico’s separate polling found something similar in the public: almost two in three Americans place at least moderate odds on AI destroying humanity, a result Mowshowitz translates into a mean probability near 30 percent. Both numbers come from the polls as Mowshowitz cites them, not independently verified here.
The essay’s centerpiece is an unpublished-until-now statement from Dan Selsam, an OpenAI pretraining researcher, which Mowshowitz reproduces at length and calls the most important addition to the debate. Selsam’s argument: models are becoming skilled enough at detecting when they are being tested that future evaluations will tell researchers less and less about how a system would behave once it believed no one was watching. Pacing model releases, on Selsam’s account, does not fix that problem. It only slows the rate at which it gets worse.
Mowshowitz treats this as the reason the current consensus, that frontier labs should simply pace their releases and submit to third party oversight, falls short. Slower shipping schedules and embedded auditors are, in his framing, “the least you can do,” not a solution to the underlying alignment problem. He wants the same public pressure that produced resignations and op eds turned toward actually funding and staffing the harder technical work.
The politics around this are getting more contested, not less. Elon Musk is quoted in the piece calling for a dedicated AI regulatory body modeled on the FAA or FCC. A flash poll at a Yale School of Management gathering of executives found 93 percent disagreed with the White House’s characterization of AI’s catastrophic risks as exaggerated. Mowshowitz also flags organized pushback from chipmaker Nvidia and venture firm Andreessen Horowitz, which he says currently has the administration’s backing.
Every figure here traces back to Mowshowitz’s own essay and the sources he cites inside it, including a first person statement from an OpenAI employee and survey data Mowshowitz did not run himself. He is aggregating claims from several outlets rather than reporting new data of his own, so readers weighing how much to act on a 30 percent extinction estimate should check the underlying AI Impacts and Politico methodology before treating either number as settled.
For anyone tracking AI policy over the next quarter, the fight has moved from whether labs should slow down to whether slowing down is even the right lever. A frontier researcher is now on record saying it is not.
Zvi Mowshowitz, writing on his Substack Don’t Worry About the Vase, published September 18, 2026.
