Article — "Anthropic's Alignment Chief Puts Extinction Risk Above 10%—and Admits No Plan Exists": Two things happened this week that cannot both be fully true in the way we normally use the word "believe."
A researcher named Jacob Coxon resigned from Anthropic, saying the company was "gambling with our lives," and stopped building. And the head of Anthropic's Alignment Science team, Evan Hubinger, publicly agreed with him—"we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade"—conceded that "we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to," and kept building.
Same forecast. Opposite behavior. That gap is the whole story.
Same forecast. Opposite behavior. That gap is the whole story—a sincere belief that reorganizes nothing is, as public evidence, indistinguishable from a marketing signal.
This week the machines recede and the people step forward, because for the first time the leaders have attached numbers to their fear, and numbers are supposed to be appraisable.
Consider the formal structure of Hubinger's post as an act. An alignment lead at a company whose entire brand is safety publicly affirms a double-digit extinction probability, concedes no plan exists, and returns to work Monday. The reflex is to call this hype—a valuation play, a bid for regulatory capture that locks in incumbents. Critics quoted by Axios say exactly that.
I want to refuse that reading, because the more interesting possibility is that it's sincere. And sincerity, here, is the more damaging diagnosis.
Paul Bloom has the sharpest frame for this. As a teenager, his rabbi reportedly believed the Lubavitcher Rebbe was the Messiah and the world would soon end—and was, in every other respect, perfectly normal. His wife read Stephen King.
Years later Bloom encountered, per his account, a computer science professor who put AI extinction at 99% within a decade, and delivered the forecast with "great comic energy," cracking the room up as he described a supervirus killing everyone. Bloom, older and less polite now, reportedly asked whether he saved for his children's education. The professor ducked.
Sit with the strangeness Bloom names. The man was laughing about a 99% chance his children would never reach adulthood. The rabbi's cheer at least made sense—he expected the end to be glorious. What is funny about everyone dying soon?
Bloom's conclusion is not that these people are lying. It's that social forces block the full adoption of certain beliefs. We are conservatives about consensus; when nearly everyone around us rejects an argument, a part of the mind whispers "this can't be right," and we keep funding the college accounts.
So the doomer holds two things at once: the argument he finds compelling, and the quiet social veto that stops it from reorganizing his life. He asserts the belief. He does not inhabit it.
Bloom is careful about the exceptions, and they matter for parsing Anthropic. He grants that some misalignments between belief and action are innocent. You reach for your broken phone in the Apple Store—you didn't stop believing it's broken, you just forgot. Phobias, hot emotions, momentary lapses. These don't impeach the belief.
But he insists this case is not like the broken phone. When people genuinely believe they and those they love will die by a date—a terminal diagnosis, say—they reliably change how they live. They stop saving for a retirement they won't see. The absence of that adjustment is evidence.
Now line up the three Anthropic responses against that test, because they don't score the same.
Coxon resigned. Whatever else you think of his read on the technology, his action discriminates: he claims the race is a "hubristic gamble," per The Decoder, and he removed himself from it. Belief and behavior point the same direction. This is the terminal-diagnosis response.
Samuel Marks is the partial case. He wrote that developers believe their technology "could cause human extinction," that seniority correlates with concern—and, per The Decoder, that he signed the open letter calling for a slowdown while continuing to work. He registered dissent through a channel that costs little and kept his position. Belief hedged by consensus, exactly Bloom's mechanism.
Hubinger is the one the broken-phone excuse can't rescue. He didn't forget. He didn't act under a phobia's grip. He calmly stated a greater-than-10% chance of human extinction, conceded there is no plan, and continued the work that the number indicts. This is a considered, repeated, public position—the opposite of a momentary lapse.
And that is precisely where the credibility problem detonates.
Notice what the number was supposed to do. A quantified forecast from an insider is meant to be more appraisable than a vague warning—that's the point of putting a figure on it. Axios editors, the messengers themselves, conceded the opposite: "these vague warnings are impossible to validate or appraise."
The reason they can't be appraised is not that 10% is fuzzy. It's that the action attached to it doesn't discriminate between two very different states of the world.
State one: Hubinger sincerely holds the belief, but social consensus vetoes its full adoption, so he behaves like a man who doesn't hold it—Bloom's mechanism running inside a frontier lab.
State two: the claim is reputation-serving performance—safety branding, regulatory positioning, the doom-signal that flatters a company selling the danger.
From the outside, these produce identical behavior. In both, the person states the catastrophic number and keeps building. The action is the same, so the action tells you nothing. The signal carries no information because it doesn't move under the two hypotheses that matter.
Stated fear is the mirror image of moral licensing as an audience effect. A publicly asserted catastrophic forecast that reorganizes nothing in the speaker's life functions, to the audience, indistinguishably from a marketing signal—regardless of what's happening inside the speaker's head.
That's the corrosion. Not that Hubinger is lying—I don't think he is, any more than Bloom thinks the professor was. It's that a sincere belief which changes no behavior is, as public evidence, worthless. It can't be told apart from the cynical version, so a rational observer is forced to treat it as if it might be the cynical version.
Coxon anticipated this and tried to inoculate against it: "This is not a marketing stunt." But the way you demonstrate that a doom claim isn't a stunt is by acting on it—which is why his resignation reads as the only fully legible act of the three. The person who quit made his belief appraisable. The person who quantified it did not.
Which leaves Congress—the audience Axios explicitly names—holding a number they cannot use. A greater-than-10% extinction probability from the alignment lead of the safety lab sounds like the most decision-relevant fact imaginable. It is inert, because the man who stated it did the same thing the next morning he'd have done at 0%.
Anthropic and OpenAI, Axios notes, are "full speed ahead, even as they practically beg for regulation or a global pause." Begging for a pause you have the power to take unilaterally is not the terminal-diagnosis response. It is the professor ducking the question about the college fund.
The tragedy Bloom flags at the end applies here too: the fact that doomers don't act on their beliefs doesn't make them wrong. Some true things are very hard to believe. Recursive self-improvement may be exactly as dangerous as Hubinger says.
But belief is not the assertion of a proposition. Belief is the proposition rearranging your life. Until the number does that to the people stating it, the rest of us are right to read it as we'd read any claim whose author's conduct contradicts it—as something other than a report on the world.