Are Async Technical Interviews Fair? Validity, Integrity, and Candidate Experience
The three concerns
When teams hesitate to adopt async technical interviews, they almost always raise three concerns:
- Predictive validity — does async performance correlate with on-the-job performance?
- Integrity — are async results worth anything when ChatGPT exists?
- Candidate experience — does requiring a 2-hour take-home filter for who has free time, not who has skill?
These are reasonable concerns. Below is what the research actually says and what the operational fixes look like.
Predictive validity
The research base for work-sample tests — which is what async coding exercises are, technically — is solid. Schmidt & Hunter's 1998 meta-analysis on hiring methods, updated by Sackett et al. (2022), places work-sample tests near the top for predicting job performance: validity coefficients in the 0.3–0.5 range, comparable to structured interviews and significantly above unstructured interviews.
That said, the validity number assumes:
- The exercise reflects actual job tasks (not algorithm puzzles)
- Scoring is rubric-based, not vibes-based
- Multiple raters score the same submission (or at least one rater uses a calibrated rubric)
An async stage that violates any of these has lower validity than a coin flip. The format is not magic; the implementation is.
Integrity in the AI era
The honest answer: an async stage with no integrity controls and a generic algorithm prompt produces close to zero signal in 2026. Candidates who use AI score the same as candidates who don't, except faster.
But "async is broken" does not follow. The fixes:
- Use AI-resistant question formats. Read-this-codebase and tradeoff-design questions are far harder to outsource than implement-X questions.
- Capture process signals. Paste events, tab focus, keystroke rhythm — none are individually proof, but together they flag the submissions to scrutinize.
- Pair with a live follow-up. A 20-minute follow-up where the candidate explains their own code is the single highest-ROI integrity control.
- Set expectations. Tell candidates upfront whether AI use is allowed, partially allowed (autocomplete OK, full generation not), or banned. Most candidates will tell the truth if you tell them the rules.
An async stage with these four controls produces materially better signal than a typical unstructured live interview, which has its own integrity problems (interviewers who recall the wrong answer, candidates who memorize question banks, etc.).
Candidate experience and accessibility
The "async filters for free time" critique is real and has two parts.
The time-budget part. A 4-hour take-home does filter for candidates who have 4 spare hours, which correlates with not currently being employed, not having caregiving responsibilities, and not being in another active hiring process. Fix: cap async at 90 minutes for screening, and pay candidates for anything longer.
The accommodations part. Candidates with disabilities sometimes need more time, screen reader compatibility, or different formats. Fix: state the accommodation process in the invite, accept requests without requiring medical documentation, and design exercises that work without strict time limits where possible (the time pressure is rarely the signal you care about).
A well-designed async stage is more accessible than a typical live coding interview, because the candidate controls when they take it. A candidate in a different time zone, with a fluctuating schedule, or with social anxiety can perform closer to their actual ability in an async format than in a 9 AM live whiteboard session.
The data on candidate drop-off
Internal ClarityHire data across 4,000 take-home invites in Q1 2026:
- Async exercises ≤30 minutes: 87% completion rate
- 30–60 minutes: 71%
- 60–90 minutes: 58%
- 90–120 minutes: 41%
- 120+ minutes: 24%
The drop-off is steep. If your funnel cannot afford to lose 60% of candidates, your async stage cannot exceed 60 minutes. Knowing this number lets you set scope correctly.
Putting it together
Async technical interviews are fair when they are:
- Scoped to the actual job (work samples, not puzzles)
- Bounded in time (≤90 min for screening, paid above that)
- Scored with a published rubric by multiple raters
- Combined with a live follow-up and process-level integrity signals
- Transparent about AI rules and accommodations
When they are not, they are at best a coin flip and at worst a polite way to reject candidates the team had already decided against. The format is fine. The implementation is what teams get wrong. For the operational playbook, see our best-practices guide and our question examples.