Mercor Killed the Take-Home Test. It Didn’t Solve Interviewing.

Osvald Nitski didn’t mince words on the 20VC podcast, according to The Times of India. Mercor’s head of product said the $10 billion AI hiring startup wants candidates who “still can have good judgment and know what they’re doing and not just, like, regurgitate what comes out of Claude.” That is why Mercor has junked take-home assignments and moved to whiteboard interviews and systems-design discussions instead. Nitski is right that AI broke the take-home as a hiring signal. He is wrong to assume that changing the format fixes the underlying problem. Move the exercise into a different room and you can still be running the same unstructured guesswork you started with, just live instead of asynchronous.

The Take-Home Stopped Measuring What It Was Built to Measure

Take-home assignments worked on a simple premise: give a candidate unsupervised time and see what they produce alone. That premise depended on the work reflecting the candidate’s own thinking. Nitski’s account makes clear that assumption no longer holds. Mercor still runs one round to check whether a candidate knows how to use AI tools, but it has moved away from assignments that AI can complete with little effort. That is a rational response to a broken signal, not a retreat from rigor. A test that no longer distinguishes candidates has stopped being a test.

Whiteboards Carry Their Own Baggage

The report notes that whiteboard interviews, where a candidate explains their reasoning while writing or drawing on a board, were widely used at Google, Meta and Microsoft during the 2010s. Mercor is not introducing a new format. It is reaching back into an old one. Watching someone solve a problem in front of an unfamiliar interviewer while narrating every step tests composure under performance pressure at least as much as it tests engineering judgment. In our view, that combination has never been the clean signal companies wanted it to be. Reaching for a format because it is harder for AI to fake is not the same as reaching for a format proven to predict who does the job well.

Judgment Is a Skill You Have to Train Interviewers to See

Nitski frames the goal precisely: “We care a lot about being able to set up good experiments and understanding statistics, having good judgment, and then systems design as well,” he said, describing what the whiteboard sessions are meant to surface. That is a defensible target. It is also only a usable signal if two different interviewers, watching two different candidates work through two different whiteboard problems, come away with comparable evidence of how each one thinks. Judgment is not self-evident. It has to be probed for with consistent follow-up questions and scored against a shared standard, or the “good judgment” round becomes exactly as subjective as the take-home tests it was meant to replace. An unstructured whiteboard session is exposed to interviewer bias and inconsistency in the same way a take-home was exposed to AI. Swapping one vulnerability for another is not progress. It is a lateral move.

There Is No Format Consensus, and That Is the Real Signal

Mercor’s retreat from take-homes is not the industry’s verdict. Cognition, an AI coding startup, argues the opposite case. Its head of people and operations, Emily Cohen, told Business Insider, “I guess this is like asking a kid to take a math test without a calculator,” about banning AI use in interviews. She argues that for most of the work a hire would actually do on the job, candidates should be free to use AI tools. Other AI coding startups, including Lovable, Cursor and Kilo, have said practical work trials and take-home projects remain a better way to judge both technical and soft skills. Two camps of AI-native companies looked at the same shift and landed on opposite formats. Our read is that this divergence means format is not the variable doing the real work. If it were, the market would have converged on one answer by now. What separates a defensible process from a fragile one is not whether candidates get a calculator or a marker. It is whether the exercise, whatever shape it takes, produces structured, comparable evidence of how a candidate actually thinks. That evidence has to hold up when compared across multiple candidates and multiple interviewers, not live in one interviewer’s gut feel.

What an AI-Resistant Interview Still Has to Prove

If you are reviewing your interview process for AI exposure, resist the temptation to stop at which format is harder to fake. That question has an answer for exactly as long as it takes someone to build a tool that closes the gap, and our read is that gap does not usually stay open for long. The harder, more durable question is whether your interviewers know what to probe for, whether they ask it the same way across candidates, and whether what they capture can be defended against each other’s notes. Mercor solved a real problem. It has not yet shown it solved the one that actually determines whether a hiring process holds up: not what candidates can fake, but what your interviewers can consistently see.

Original reporting: The Times of India.