Most technical screening falls into one of three buckets. Each one claims to measure "can this person engineer." Each one actually measures something narrower — and the gap between the claim and the reality is where bad hires (and missed great ones) come from.
Here's the honest version.
Whiteboard / live algorithm
What it claims to measure: problem-solving.
What it actually measures: recall of a specific class of data-structure puzzles, performed from memory under social pressure, with no tools.
Whiteboarding rewards people who recently ground LeetCode and punishes people who've spent the last three years shipping production systems and forgot how to invert a binary tree on demand. It correlates with interview prep, not with the job. And in an AI world it's doubly strange: you're testing the one condition — no assistant, no reference — that will never occur again after the offer.
It's fast, which is its only real virtue.
Take-home project
What it claims to measure: real-world engineering.
What it actually measures: how much unpaid time a candidate is willing to spend, and how well they (or a friend, or an AI) can polish a greenfield project with no deadline pressure.
Take-homes are more realistic than whiteboards — real editor, real code, real tradeoffs. But they have three problems. They're slow (multi-day turnaround, then a reviewer has to read it all). They're unfair to anyone with a job or caregiving duties who can't sink a weekend into your funnel. And they're trivially AI-completable now, with no visibility into whether the candidate understood a line of it.
A take-home you can't see the process behind is just a code sample of unknown provenance.
AI-native interview
What it claims to measure: judgment under realistic conditions.
What it actually measures: decomposition, prompting, verification, and recovery — observed live, with the assistant the candidate will actually use on the job turned on.
This is the whiteboard alternative that doesn't pretend the assistant away. The candidate gets a realistic task in a real editor with a real AI assistant. A silent watcher captures the whole process — prompts, edits, test runs — and scores it against your rubric. It's async like a take-home (no scheduling, no live proctor) but delivers a scorecard in minutes, not after a reviewer slogs through it.
Side by side
| Whiteboard | Take-home | AI-native | |
|---|---|---|---|
| Measures the real job | ✗ | ~ | ✓ |
| Resistant to AI gaming | ✗ | ✗ | ✓ (assumes AI) |
| Fast turnaround | ✓ | ✗ | ✓ |
| Fair to busy candidates | ~ | ✗ | ✓ |
| Process visibility | ✓ live | ✗ | ✓ |
| Candidates don't resent it | ✗ | ~ | ✓ |
The point isn't "AI-native always wins"
Every format trades something. If you need a 20-minute phone screen, a quick exercise is fine. But for the round that actually decides the hire, the question is: are you measuring the job, or an artifact of the format? Whiteboards measure prep. Take-homes measure free time and provenance. An AI-native interview measures judgment — which is the thing you were trying to hire for in the first place.
Related: why banning AI won't fix your interview · async coding interviews that save senior time.