Axios co-founder Jim VandeHei tells For Humanity why Congress is suddenly paying attention to AI risk, and the review board he says still doesn't exist.
Jim VandeHei has spent close to three decades covering Washington, first as a reporter, then as co-founder of Politico and later Axios. On this week's For Humanity, recorded two days after former Anthropic researcher Jacob Coxon's resignation post went viral, he tells host John Sherman that the reaction in Congress was unlike anything he has seen since AI became a mainstream concern, and that he still does not think it produces real legislation this year.
"More activity from lawmakers in 24 hours than I've seen since the advent of AI," is how VandeHei describes the days after Coxon's post, according to his own account on the show. Axios's own reporting backs that up: within days, Rep. Ted Lieu was pushing legislation requiring AI "kill switches," Sen. Bernie Sanders and Rep. Greg Casar were preparing a bill to pause advanced AI development, and Sen. Ruben Gallego proposed a bipartisan AI select committee.
VandeHei is blunt about what he thinks happens next: probably not much. He estimates that only 2 to 5 percent of the 535 members of Congress understand the technology "at any level of expertise that would make you feel satisfied," and says most of Washington's attention goes to problems that predate AI entirely. He also points to the economics: by his account, roughly 40 percent of the economy is now propped up by AI-related investment, which he says makes elected officials wary of doing anything that could look like slowing it down in an election year.
"There'll be a lot of hoopla, there'll be a lot of talk," VandeHei says. "There won't really be any legislation." In his view, the next real turning point is either another incident "on steroids" or an election where voters punish a candidate for being on the wrong side of the AI debate.
VandeHei's own proposal is specific. He argues the industry and government should have set up an independent body years ago, staffed with real expertise in biotech, data, and the workforce, with the ability to review frontier models before release and the authority to hit "some kind of kill switch" if a lab is about to release something "catastrophic or wildly unpredictable." According to VandeHei, most lab leaders he has spoken with privately agree with some version of this idea, even if none of them has built it.
He points to Anthropic CEO Dario Amodei's own public estimate, given in an interview roughly a year earlier, that there is a 10 to 25 percent chance AI development "ends very, very badly." VandeHei says that when he pressed Amodei on why he keeps building the technology anyway, the answer was that Anthropic believes it can do it more responsibly than its competitors would. VandeHei is careful to frame this as Amodei's own reasoning, not an endorsement of it.
He floats a parallel accountability idea for data centers, arguing that operators should be required, going forward, to fully offset the electricity and water they use and disclose deal terms to the communities where they build. Sherman, for his part, describes a reporting trip to Meta's Hyperion data center in rural Louisiana, where he interviewed residents about the project's effect on their community.
Sherman has said publicly that his own estimate of AI causing his death, informally known as a P(doom) figure, has run as high as 85 percent, and dropped to around 70 percent this week after the Coxon story broke. Asked directly whether he thinks there is a better than 50 percent chance AI kills him personally, VandeHei says no. He attributes part of that to his own temperament, describing himself as a lifelong entrepreneur and optimist, and part of it to a belief that people underestimate their own capacity to course-correct under pressure, pointing to the Civil War and Cold War-era nuclear arms control as precedents the country did not seem capable of managing in advance.
He is careful not to overstate the comparison. Getting the same kind of restraint with China on AI, he says, is not guaranteed, and he places real weight on an upcoming conversation between President Trump and Chinese leadership as a possible, though uncertain, opening.
The moment VandeHei says stuck with him most did not come from a lab or a hearing room. At an advertising industry panel the day after Coxon's story broke, Sherman asked a room of about 40 working professionals, all of whom use AI daily, whether they would support throwing the technology "back in the river" and pretending it never existed. By Sherman's account, nearly every hand in the room went up.
VandeHei says he was not surprised. According to VandeHei, most people are not thinking about AI the way researchers and journalists do. They are going to work, raising kids, and hearing mostly that the technology might take their job or, in some framings, end humanity, without seeing a version of it that has clearly made their own life better yet.
VandeHei's newest book, "Simplify," co-written with Mike Allen and Roy Schwartz, argues that AI often makes organizations more complicated rather than less, by making it easier for anyone to generate more drafts, more decks, and more noise. On the show, he connects that same argument to the public conversation about AI risk: the technology is moving fast enough that, in his words, the country does not have years left to build the kind of independent oversight he is describing. Whether that oversight gets built before the next incident, he says, is still an open question.
To support work on AI risk, you can donate to Guard Rail Now.
Watch the full conversation: For Humanity #93 on The AI Risk Network
Read more AI safety coverage: The AI Risk Network on Substack
