The most trusted take last week about AI's power came from Jacob Coxon, 27, four months into his job at Anthropic (and two months before his equity vested), who posted that he'd resigned because the labs are "racing straight to self-improving superintelligence and gambling with our lives," and that "the people building AI earnestly believe that it could kill us all by the end of the decade." Shortly thereafter, Evan Hubinger, an alignment science lead who still works at Anthropic, replied that Coxon was right, "we really do earnestly believe AI could kill all humans!" and put his own odds above 10 percent within the next decade.

So when Anderson Cooper sat down with Dario Amodei on Saturday, he asked the question the way any of us scared human beings would: "Do you earnestly believe that AI could kill all humans? And what percent chance do you think it is?"

On camera, Amodei's opened with a smile (yes, really) saying, "it's funny, I agree with Jacob much more than I disagree with him." Not exactly comforting words. Then he picked his way to a pivot, telling Cooper he used to talk in "unconditional probabilities," and the chance "is not all that high," but it's "more illuminating" to ask "how do you decompose those probabilities." A ten percent number "sounds like it's a roll of the dice"; it’s better to say it "forks into different paths," and "if we take the right paths, then the chance of something going wrong is very low," while the wrong paths make it "even higher than that." The right way forward, apparently, was the three-step plan Amodei had published that morning on his personal blog.

That's not an accident of course because it’s Amodei's habit to put his biggest messages in long essays on his own site, and The Information's Cory Weinberg argued last night that it's been a smart strategy for a CEO who isn't a natural on camera: the essay comes across as careful, it sits above the X food fight, and this one almost immediately prompted Altman and Musk to agree (there was clearly a preview there). It's also why the interview went the way it did because he'd written the framework and sat down with Cooper to explain it. When a frightened person asked a frightened person's question, he answered with the document. 

Sam Altman got the same line of questioning from Fortune's Alyson Shontell on Friday, and he said, "I think it is unacceptable to be taking like a 10 percent chance of killing everybody by the end of the decade." Pressed, he wouldn't take the number either, "I don't know how you can put a number like that," "whether it's 10 or eight or six," but only after saying what he'd do about it. OpenAI has "been pausing training runs until we can make a safety case that we're more comfortable with." The IPO will need to wait: "right now would be an ill-advised moment to go public.” Asked whether a pause would really happen if it cost money, he said "that's on us" if people doubt it, and that he'd stand in front of investors and say "I am sorry" for "a decision that is extremely against your financial interest right now." Melt the GPUs if humanity's survival required it? "Easy yes," he volunteered.

The policy is nearly identical for both companies, both want outside evaluators inside the labs and both say slowing down doesn't mean stopping.

What differs from a communications perspective is who each man thought he was talking to, and it certainly wasn't me, a lay person scared for the future of humanity. Amodei answered the question an AI philosopher would ask: how should we reason about the risk? Altman answered the question a shareholder would ask: what will this cost you? "As we have said, we're not rushing into an IPO," he said. Slowing down is "not a hard decision," he said, because "even with Astra, we can grow revenue hugely, even if we never shipped another model." So the sacrifice doesn’t cost him much at all.

This question is now leaving the labs though. Marc Benioff shares a stage with Amodei this week at the 2026 Dreamforce conference. Every bank, hospital system, retailer and airline that has bragged about being "AI-first" for two years now runs on one of these models, and the next earnings call, town hall or CNBC interview could very likely include a reporter pressing your CEO on whether they'd keep using it if it turned out the labs were gambling with our lives. Voters aren't only afraid of superintelligence by the way. They're afraid of their jobs, their kids and their own attention, and those fears belong to the companies deploying AI far more than to the ones building it.

Which is why I'd argue this is the biggest responsibility our profession has ever been handed. This is the most consequential technology any of us will ever work on, the stakes are whether our kids get a future at all, and the way the world talks about it, argues about it and makes decisions about it will be shaped by how the people building it and deploying it answer the question. That's our job. Whether a CEO says something a frightened person can hold onto, or something that sounds like a plan, is on us.

If your company is running heavy on any of these models, a few things you should have ready:

Work with your leadership team on a pov here asap. Your CEO doesn't need to calculate a probability of doom, just a message that acknowledges whether they're worried and how they’re thinking about risk mitigation. Something a person could say at a kitchen table like, “I don’t know the odds but I do know the rule WE run by which is that if we can't stop a system, we don't use that system. And we test that we can stop it every quarter."

Answer the everyday fear with substance, not philosophy. The existential question belongs to the labs but the jobs question belongs to you. This should be the question your CEO is prepared to answer with specifics: what kinds of skills you're retraining for, what you've decided not to automate and why. Don't answer "are you replacing people" with a framework either because that’s just the same mistake one level down.

Don't borrow the labs' vocabulary. "Pacing the frontier," "alignment," "safety case," "recursive self-improvement." The moment your CEO says any of those on camera, they've joined the industry's conversation with itself. You must be able to translate into, "we can turn it off," "we check its work," "we don't let it decide alone."

Never name a sacrifice that's costs you nothing. If the CEO says they're slowing something down for safety, it has to be something that costs you something.

I'm less sure about the bigger question. When the most believed voice on your company's risk is the person who just quit, what does your CEO have to say to get the microphone back? That one remains to be seen.