How to Assess Work Samples When Candidates Use AI

As a standalone signal, the whiteboard interview is dead. A controlled NC State and Microsoft study found candidates in watched technical interviews performed about half as well as those solving the same problem privately, and generative AI now solves both whiteboard puzzles and take-homes in minutes. The durable replacement is a work-sample assessment: a job-relevant task that ends in a live defense, where the candidate explains and changes real decisions out loud.

That last move is the whole game. AI overlays can write code during a screen-share and finish a “3-hour” take-home in minutes, and can also assist with real-time explanations. The strategic response from the best-run engineering teams in 2026 is not surveillance software. It is a format shift toward assessments that test judgment and communication, without treating that assessment as proof against AI assistance. This guide gives you the decision framework: which format to use now, how to make it both fair and cheat-resistant without spyware, and how Anthropic, Stripe, Vercel, and Linear actually choose.

Are Whiteboard Interviews Dead? (Yes, and AI Is Only Half the Reason)

Yes, as a standalone signal. The whiteboard interview was broken before AI ever touched it, and AI removed whatever signal was left.

The first problem is that it never measured the right thing. In a controlled experiment, Behroozi and colleagues at NC State and Microsoft (2020) had candidates solve the same problem in two conditions: alone, and watched by an interviewer in a traditional whiteboard setup. Candidates in the watched condition performed roughly half as well. The format largely measures performance anxiety and working-memory load under observation, not engineering competence. It also penalizes exactly the people you want to hire fairly: introverts, neurodivergent candidates, and anyone whose communication style does not match a high-pressure verbal performance.

The second problem arrived in 2025. Overlay tools like Cluely, Interview Coder, and Leetcode Wizard now feed answers invisibly during a screen-share. A standard LeetCode-style problem is solved silently in the background while the candidate types. If your screen still relies on competitive-programming puzzles, you are no longer measuring the candidate. You are measuring their tooling.

This does not mean live coding is worthless. It means the watched-puzzle format is. The version that survives is collaborative live coding: pairing in a real IDE on a realistic problem, where the interviewer is a partner thinking alongside the candidate, not a proctor waiting for the right answer. That tests how someone reasons, asks questions, and works in unfamiliar code, which is both closer to the job and far harder to fake with an overlay.

Why AI Broke the Take-Home Too (and Why Surveillance Is the Wrong Fix)

The unsupervised take-home is now the most AI-exposed format of all. The fix is not detection software. It is design.

Take-homes always had the best real-world-validity story, and they remain valuable. But an ungraded, undefended async task is the easiest thing in your loop for AI to complete. Assessment vendor Fabric reports that a take-home designed to take three hours can be finished by AI tools in roughly eight minutes, and that cheating adoption in its candidate pool more than doubled across 2025, from about 15% to 35%. Treat those exact numbers as directional rather than gospel; they are vendor self-reported and uncited. The direction, though, is not in dispute, and any engineering leader who has reviewed a take-home that “felt too clean” already knows it.

The tempting response is to buy your way out with proctoring: eye-tracking, keystroke logging, screen lockdown, browser spyware. Resist it, for three reasons.

• It is adversarial and brand-damaging. Engineering is a small, talkative community. Candidates share surveillance horror stories, and your best applicants self-select out before they apply.
• It creates its own bias and accessibility problems. Lockdown and eye-tracking tools penalize neurodivergent candidates, disabled candidates, and anyone with a nonstandard setup. The EEOC and DOJ have made clear that employers remain liable when an automated assessment tool causes adverse impact, regardless of who built it.
• It does not even work. Surveillance fights the symptom. A second monitor or a phone defeats most of it. You spend trust and budget to lose anyway.

The durable answer is design-based resistance: build formats where the signal lives in live reasoning, so there is nothing for an overlay to rescue. In some roles you can go further and explicitly allow AI during the task, then score how well the candidate directs and critiques it, because that mirrors the actual job.

Which Assessment Format Should You Use Now? A Decision Framework

Match the format to the daily reality of the role, and make sure at least one round forces real-time judgment. There is no single best format; there is a best format for this role.

The through-line across all four is the same: ask candidates to explain decisions about work they can demonstrate. Pick the format that looks most like a normal Tuesday in the role, then make sure the candidate has to explain their thinking to a human at least once.

A practical default for most startup engineering roles is the second row: a short, paid, realistic take-home that becomes the agenda for a live conversation. You get the ecological validity of real work plus the cheat-resistance of a live defense. If you want the tactical mechanics of designing that take-home itself, scope, time budget, and grading, see our deep-dive on how to structure code assignments candidates don’t hate.

What a Live Discussion Adds to the Assessment

The single most durable anti-cheating mechanism is to end every async artifact with a live defense: “Walk me through this. Now change requirement X. Why did you choose this over the alternative?”

Here is why it works. An overlay LLM can produce the code. It can also assist with explanations, changed requirements, and debugging. A discussion adds evidence of understanding but does not prove unaided authorship. The artifact stops being the final signal and becomes the agenda for a 20- to 30-minute conversation about judgment. Hesitation alone does not establish that a candidate copied the work.

Discussing work prepared in advance may reduce pressure to produce code from scratch under observation. It still requires clear criteria and appropriate accommodations. The cited NC State study does not establish predictive validity for this particular format.

Concretely, the move looks like this in any loop:

1. Candidate completes a small, realistic, paid work sample async.
2. A 25-minute live session opens with “walk me through your approach.”
3. You change one requirement live and watch them adapt.
4. You ask them to debug or extend one piece on the spot.
5. Reviewers score the reasoning, on a rubric, before anyone debriefs.

No spyware. No accusations. Assess the explanation against agreed criteria; the format cannot guarantee that no AI assistance is used.

How Anthropic, Stripe, Vercel, and Linear Actually Hire

The best-run engineering teams have already made this shift. None of them rely on watched whiteboard puzzles, and none of them rely on surveillance. They rely on realistic work plus live judgment.

Anthropic runs a recruiter screen, a technical phone screen, then either a take-home or a roughly 60-minute live assessment (role-dependent, in CodeSignal, and explicitly not LeetCode-style), followed by four to six onsite rounds including system design and a heavily weighted values round. Most notably, the company that makes Claude publishes an explicit candidate AI policy. As of a July 2025 reversal, candidates may use AI to polish application materials, but it is prohibited in live interviews and take-homes: “Complete these without Claude unless we indicate otherwise. We’d like to assess your unique skills.” That is an example of explicit rules, not proof that assistance cannot occur.

Stripe runs a deliberately practical loop: debug an unfamiliar codebase, build a small integration from scratch, work multi-part problems while narrating your thinking. Some rounds run as pairing. It is closer to real engineering than competitive programming on purpose.

Vercel uses a collaborative, build-style coding session plus system design, weighted toward frontend product judgment and communication.

Linear uses a short (around three-hour), paid, work-trial-style project followed by a code-review discussion, and requires a near-unanimous “strong yes” from the panel to extend an offer. Structure, a high bar, and work-relevance, in one loop.

A useful contrast is the GitLab-style pattern: an async code review of a real merge request as the basis for a live discussion. It tests reading and critiquing real code rather than generating it, which fits a remote-async culture. The point of listing five different approaches is not that one is correct. It is that each company matched the format to how it actually works, and every one of them ends in a moment of live, defensible judgment.

Is the New Format Actually Fairer? What the Evidence Says

Job-relevant work samples are among the most valid and lowest-bias selection methods, but only when they are structured. Fairness comes from structure, job-relevance, and consistency, not from the format label.

Be careful with the numbers, because the canon was recently corrected. Sackett, Zhang, Berry, and Lievens (2022) re-analyzed decades of personnel-selection research and lowered several long-cited validity estimates:

• Structured interviews are now the single best predictor at an operational validity of about .42 (revised down from .51).
• Work-sample tests sit at about .33 (revised sharply down from the long-quoted .54).
• General cognitive ability sits at about .31 (down from .51).

These are broad estimates across methods and occupations. They do not establish a validity of .42 for every live discussion of an assignment. Use common questions and criteria, then evaluate your own format.

On bias specifically, use peer-reviewed effect sizes rather than the recycled marketing stats. Aamodt’s meta-analysis found unstructured interviews far more susceptible to bias (d = .59) than structured ones (d = .23), and racial score gaps shrink as structure increases. Add one more lever: pay candidates for substantial work-sample stages. Campion and colleagues (2025) found that practice and paid work-sample testing reduce subgroup score differences, and paying for real work also raises completion and helps caregivers and lower-income candidates who cannot donate unpaid hours.

Why the “42% / 81% Bias Reduction” Stats You’ve Seen Are Unreliable

You will find dozens of vendor blogs claiming structured interviews “reduce gender bias 42%, racial bias 35%, and improve accuracy 81%.” Those three numbers have no traceable primary study; they are copied from one source to the next. Use the peer-reviewed figures above instead. The credibility of your fairness argument depends on citing research that actually exists, especially in a regulatory environment where the EEOC and DOJ expect you to defend your process.

Combine the Assignment and Review in Kit

AI broke the whiteboard and the unsupervised take-home in the same year. The fix is not surveillance. It is designing the right format: job-relevant work samples, paid and structured, that always end in a live defense. The problem with doing this by hand is that the pieces, the realistic task, the payment, the scheduled defense, the independent scoring, live in five different tools and tend to drift apart. Kit makes them one composable pipeline.

• Composable process templates let you encode the thesis directly: an application form flows into a code-assignment stage, then into a live-interview round, then into team review and an offer. The take-home is built to be the agenda for the defense, not the final signal.
• The code-assignment stage creates a private GitHub repository from a template. You can set a deadline and ask the candidate to use a branch and pull request. Kit does not create the pull request for them. Job relevance depends on the assignment your team prepares.
• Per-stage payouts let you pay candidates for substantial work-sample stages, which is both the fairness move backed by Campion (2025) and a clear respect signal.
• Team review with per-stage reviewers gives you structured, independent scorecards before the debrief, a record that helps compare assessments and inspect the process. It does not itself establish legal compliance.
• Live-interview scheduling productizes the defense round, so the “walk me through this” conversation is a built-in stage rather than an afterthought.

If you want the validity case in depth, read structured interview scorecards and predictive validity, and for the broader shift away from puzzle screens, see why LeetCode is obsolete in a post-AI interview.

The whiteboard is gone and the unsupervised take-home went with it. What replaces them is not a new gadget. It is a format choice: paid, structured, job-relevant work that a candidate defends out loud. No format guarantees fairness or immunity to AI assistance. Start a free trial and build an assignment-and-review pipeline, or browse the role templates to start from a pre-built one.