Is Anthropic's Interview Process Built to Reward the Best Liar in the Room?

Axios reporter Madison Mills reported on August 24 that candidates interviewing at Anthropic sit for a dedicated culture interview, and that some of them get asked a pointed hypothetical about mission versus money. The version one applicant described to Axios: how would you feel if the company abandoned its AI ambitions for safety reasons and the decision sent the stock to zero?

A few other details from that reporting. The culture interview is run by an employee nominated for the job rather than a standing panel. Candidates are also asked to describe a moral quandary they've faced and explain how they handled it. Compensation is fixed and there's no negotiation. And the questions are suggested rather than fully scripted, which Axios offers as the likely reason some candidates recall a stock-crashing-to-zero framing while others remember a vaguer conversation about safety trade-offs that might cost revenue. Two former employees told Axios the money question is new. Co-founder Daniela Amodei has said publicly, in a Bain Capital video, that some form of culture interview has been in place since the company started.

The framing sits directly on a tension Axios has been tracking. CEO Dario Amodei has questioned whether newer hires are joining for the right reasons, and the company's careers page uses the word "mission" six times in its description of the hiring process.

The story got picked up as a curiosity, one more strange artifact from the most-watched company in AI. It deserves better than that. Anthropic is treating values alignment as a real selection criterion, weighted alongside technical skill, and this is the first time a frontier lab has done it openly enough for anyone outside to get a look.

Good instinct. Execution is where it gets difficult.

Culture used to be the soft part of the process

Culture fit spent most of the last decade as the soft round. A lunch, a gut read, and a nicer name for "would I want to get a beer with this person."

Frontier labs can't afford that, for reasons specific to what they build. Axios points to two occasions where Anthropic's stated values cost it something commercially. The company declined a U.S. military arrangement that would have permitted unrestricted government use of its technology, and it limited access to Mythos, its most cyber-capable model, over safety concerns. Whether either decision hurt revenue on net is hard to establish. Axios makes the point that the company is still private, and that Claude's consumer app hit number one on U.S. app store charts after the Pentagon dispute.

Either way, the arithmetic isn't really the interesting part. At a company where individual researchers make judgment calls with real downside attached, publish or don't, flag it or let it go, the distance between someone who believes the mission and someone renting a seat until the IPO turns into a real operational risk. Anthropic appears to have worked that out early. Every other serious lab will reach the same conclusion inside a year.

Josh Millet, CEO and founder of Criteria, agrees, up to a point. "The labs have the right instinct," he says. "Values alignment does predict who stays and who speaks up when speaking up costs them something, and in this industry those moments carry real weight."

"But an instinct isn't an instrument," he adds. "The minute a values question gets asked differently by every interviewer and scored on gut feel, you've stopped measuring the candidate and started measuring the interviewer. That's the piece we need to solve."

The problem is buried in the reporting

Read the Axios piece with a selection-science background and one detail stands out. Candidates remember the question differently because nobody scripted it.

That matters more than it sounds like it does. A question that shifts in wording and position depending on who asks it produces scores you cannot compare across candidates. There is no ranking a group on a measure that changes shape with the administrator.

The evidence on this is not close. Sackett, Zhang, Berry and Lievens (2022), the most recent large-scale meta-analysis of personnel selection, puts the predictive validity of structured interviews at .42 against .19 for unstructured, roughly double. The same paper found something the field is still absorbing: once the range-restriction corrections are fixed, structured interviews outrank cognitive ability as the single best predictor of job performance. The older and more widely cited Schmidt and Hunter (1998) figures were closer together, .51 against .38, but the direction has held for decades. Nearly all of the advantage traces back to standardization: identical questions, identical order, a fixed rating scale, more than one rater where you can manage it.

There's a second issue, and Axios raises that one too. Citing Bloomberg, the piece notes that candidates are paying thousands of dollars for private coaching and for mock interviews with engineers from top labs. Aline Lerner, who founded the prep company Interviewing.io, laid the math out for Bloomberg in plain terms: a few thousand spent against a possible two hundred thousand in salary.

Now drop a known, unscripted, high-stakes values question into a market with a professional coaching industry attached to it. It gets posted on Blind, which this one was, and from that point it stops measuring conviction and starts measuring preparation. Anyone sharp enough to clear a frontier lab's technical bar can produce the answer the room wants to hear. Judging by the Blind post Axios cites, the candidate who admitted they would not be happy watching their equity go to zero may well have been the most honest person in the process. By that candidate's account, the interviewer wasn't impressed. Additional posts suggest the question was already circulating in 2025.

What a rigorous version would look like

Assessing values is fine. Assessing them casually and then rejecting people on the result is the problem. Four things separate one from the other.

Start by defining what you're actually measuring. "Mission-driven" is not a measurable trait. The competencies underneath it are: tolerance for ambiguity, ethical decision-making under pressure, willingness to raise a concern when the incentives point the other way. Anthropic's second question, the one about a real moral quandary the candidate has faced, sits much closer to the mark. Past-behavior questions outperform hypotheticals, because a hypothetical rewards imagination and a behavioral question demands evidence.

Then score against a rubric written in advance. Behaviorally anchored rating scales spell out what a 1 answer contains and what a 5 answer contains before anyone sits down in the room. Without them, "the interviewer didn't seem to like that answer," which is how the Blind poster described the outcome, becomes the entire evaluation. That's a defensibility problem as much as a predictive one.

Ask everyone the same thing in the same order. Consistency is what makes the comparison legitimate, and it's what you'll want on file the first time a rejected candidate asks why.

Last, treat alignment as a two-way match. The useful question is whether what a candidate needs from an employer overlaps with what the organization actually provides, not whether they impressed the room. A workplace alignment assessment gets at that by comparing a candidate's ranked workplace priorities against the organization's own profile and flagging the gaps for a structured follow-up. It also avoids the bias trap in conventional fit interviews, where fit tends to collapse into similarity.

This won't stay inside the labs

Frontier lab hiring practice travels. Structured technical interviews, take-home work samples, leveling frameworks: all of it moved from a handful of tech companies into the general market inside a few years.

Values screening is on that path now. Plenty of companies with nothing to do with AI have concluded that misalignment drives their regrettable attrition, and they will borrow the question long before they borrow the rigor behind it.

Anthropic is asking about the right thing. The harder part, for them and for everyone about to copy them, is asking it the same way twice.

Original reporting by Madison Mills for Axios, August 24, 2026. Additional reporting cited by Axios from Bloomberg. Analysis is Criteria's own.