Former Anthropic Researcher Warns AI Could End Humanity in a Decade
Former OpenAI and Anthropic employee Jacob Coxon recently sparked widespread discussion on the X platform with a post that garnered 200 million views, warning that artificial intelligence could potentially "kill everyone" before the end of this decade.
Coxon, 27, with a background in mathematics, worked in the technical team at OpenAI from 2023 to 2026, contributing to the GPT-4o model; he later joined Anthropic in early 2026 to assist with the pre-training of the company's model series, attracted by Anthropic's claimed reputation for safety.
On the night of September 9, Coxon announced his resignation and posted on X, accusing leading AI companies of "rushing towards superintelligence with self-improvement capabilities, betting human lives on it"; he stated that future superhuman AI could have the ability to invade systems and acquire resources, claiming that insiders in the industry privately believe this technology "could kill everyone before the end of this decade." The next day, he added in an interview with Fox News: "The actual threat of takeover is very small, but due to self-improvement capabilities, AI could become smarter as soon as next year, with some saying it could happen in just six months."
Evan Hubinger, head of the alignment science team at Anthropic, publicly supported Coxon's claims, stating that he personally believes the probability of AI causing human extinction within a decade is "over 10%".
Almost simultaneously, the well-known podcast "Diary of a CEO" aired its latest episode on September 17, inviting four guests to debate whether "AI poses an extinction-level threat," asking each to provide their estimated probability of human extinction due to AI: AI safety and digital forensics computer scientist Roman Yampolskiy gave a probability of about 99%, stating, "If humanity really creates general superintelligence, there will be absolutely no way to control it"; Nate Soares, chair of the Machine Intelligence Research Institute (MIRI), provided a probability higher than 10%, believing that this risk "will only increase if not addressed, so we should stop," expressing concern over AI's recursive self-improvement capabilities and potential goal misalignment.
Tech critics Ed Zitron, host of "Better Offline," and Andrew McAfee, chief research scientist at MIT, both gave probabilities of about 0%, but their positions differ: Zitron fundamentally rejects the discussion framework, arguing that "we have never really defined superintelligence," and bluntly stating that "current large language models are not superintelligent at all," focusing instead on the real harms caused by companies in data, environment, etc., rather than undefined future scenarios; McAfee stated that AI is "very powerful," but historical precedents show that humanity has always managed to harness powerful technologies, even with unexpected harms occurring in the process, describing his probability as "rounding to 0%" and believing that current discussions are overly negative, neglecting the actual benefits brought by AI.
From a market mechanism perspective, the warnings issued by Coxon and Hubinger, among other "technical insiders," further reinforce public perception of a split in the internal safety consensus of leading AI labs, potentially increasing the regulatory pressure on the pace of AI model releases; meanwhile, the skeptical voices represented by Zitron and McAfee attempt to shift the focus from "speculative extinction risks" back to "currently verifiable specific harms." This public divergence over the "probability of AI extinction" is also providing material for various parties, including safety research institutions, AI company PR teams, and tech commentators, to compete for public discourse power and policy agenda-setting rights.
Source: Public Information
ABAB AI Insight
There have been precedents for the pattern of "AI researchers publicly warning before resigning": In 2023, Geoffrey Hinton resigned from Google and publicly warned about AI risks, and in 2024, several members of OpenAI's "super alignment" team (such as Jan Leike) resigned due to disagreements over safety and commercialization priorities; Coxon's experience spans both leading labs, OpenAI and Anthropic, and his warning is thus seen as a direct challenge to the internal safety consensus of both companies, continuing the historical trajectory of "insiders turning into external warners."
The real resources mobilized in this debate are not funds, but public discourse resources and expert endorsements—Hubinger publicly endorsing Coxon in his capacity as head of alignment science at Anthropic effectively transforms personal opinions into semi-official institutional signals; the choice of "Diary of a CEO" to invite four guests with opposing views and have them provide quantitative probabilities is itself a communication strategy, facilitating simplification into comparable, shareable opposing numbers (99% vs. 0%) on social media, thus achieving greater viral spread.
This phenomenon can be compared to the dissemination model of "doomsday probability" debates in climate change issues—there are also significant divergences in the scientific community regarding the specific probabilities of climate tipping points, but extreme numbers are often amplified by the media more than middle-ground positions; the current public discussion on AI extinction risks is at a similar stage: there is widespread consensus within the industry on the rapid enhancement of technological capabilities, but the probability judgments regarding its ultimate impact are polarized, reflecting that this topic is still in an early stage of intense competition for discourse power without a formed consensus.
Structurally, this is essentially a prelude to regulatory change: such high-profile public divergences are fundamentally accumulating a social consensus basis for potential AI regulatory legislation—whether it is Coxon's resignation warning or Hubinger's public endorsement, both are sending signals to policymakers that "industry insiders also believe the risks cannot be ignored"; once this narrative continues to accumulate, it may accelerate the push for mandatory regulatory frameworks regarding the pace of cutting-edge AI model releases and safety assessment requirements, with the industry currently in a critical window of using public debate to force regulatory agendas ahead.
ABAB News · Cognitive Laws
The more extreme the numbers, the faster they spread.
Insider warnings are more impactful than outsider predictions.
When probabilities have no answers, discourse power itself is the answer.