The Technical Interview Measures Interview Ability, Not Engineering Ability

The technical interview does not measure engineering ability. It measures interview ability. And that's a different skill.

Most people approach this problem the wrong way. They ask why skilled engineers keep failing, as if the interview is a legitimate test and the engineer is simply underperforming. The data suggests the opposite: the test itself is broken in a way that even its biggest defenders have quietly confirmed.

Google ran an internal analysis of tens of thousands of interviews. The result was not mildly inconclusive. They found zero relationship between individual interview scores and subsequent job performance. Not "a weak correlation." Not "statistically noisy." Zero. If the world's largest engineering employer has evidence that its own hiring filter predicts nothing, the puzzle isn't why good engineers fail. It's why the filter still exists.

The answer has to do with what happens when you can't measure the thing you actually care about.

Engineering work is mostly invisible. Debugging a production issue, refactoring legacy code, deciding whether to pay down technical debt or ship the feature — none of these map cleanly onto a 45-minute scored exercise. What can be measured is whether someone can invert a binary tree on a shared document. So companies measure that instead. Not because it's a good proxy. Because it's the only proxy that scales.

The structural mismatch runs deeper than lazy hiring practices. There's a cognitive mismatch baked into the format.

Psychologists who study skill acquisition describe a progression from novice to expert. Novices think step by step. They can articulate every move. Experts — the ones who've seen 100,000+ situations — think intuitively. They skip the steps. An expert engineer looks at a system and immediately sees where it will break. They can't always explain how they saw it, because they didn't reason their way there. They recognized a pattern.

The interviewer, who's usually not an expert, hears the expert skip steps and thinks: they don't have a process. They're not communicating their reasoning. They're failing.

A 2020 study by Behroozi and colleagues found that in observed coding conditions, all the women in their test group failed — while all the women in private conditions passed. Performance anxiety under observation cut results by half. The test wasn't measuring ability. It was measuring who can perform ability under someone else's gaze.

And there's data on the volatility. Research from interviewing.io shows 75% of candidates perform inconsistently across different interviews. Even strong candidates face a 22% chance of failing any given round. That means a genuinely excellent engineer could walk through the same company's door twice, pass once and fail once, and neither result tells you whether they'd be good at the job.

The system selects for a specific type of person: someone who performs well under artificial pressure, articulates thinking at a pace that matches interviewer expectations, and happens to know the algorithm patterns that show up most often. This person may be an excellent engineer. They may also be mediocre but excellent at the interview. The system can't tell the difference.

Meanwhile, the real costs are substantial. A 2012 Center for American Progress review found that replacing a bad senior hire costs 213% of annual salary. A Harvard Business School study of 50,000 workers found that avoiding a toxic employee yields twice the return of hiring a star. The hiring filter that claims to protect companies from bad hires is simultaneously rejecting the people it can't evaluate and admitting people whose primary qualification is interview fluency.

This isn't an argument for abolishing interviews. It's an argument for understanding what you're actually measuring. If you're testing algorithmic puzzle-solving under pressure, you'll hire people who are good at algorithmic puzzle-solving under pressure. That's not a bug. It's exactly what the test does.

The companies that've moved toward work-sample assessments — where candidates build something resembling actual work — find better performers and lower turnover. Harvard Business Review found employees hired through skills-based assessments deliver 25% higher performance and 40% lower turnover. But inertia is powerful. Switching to better methods requires giving up the comfort of a system where everyone knows the game.

Here's the thing most people miss: the interview preparation industry is the most honest evidence of what the system measures. Companies sell candidates 15 core algorithm patterns, timed practice sessions, and behavioral story frameworks. If engineering ability were the goal, the prep would look like building something. Instead, it looks like studying for an exam you'll never take.

The next time someone tells you that technical interviews are just imperfect — that they have signal mixed with noise — ask them to point to the signal. Google looked at tens of thousands of data points and couldn't find any. That's not imperfection. That's a different measurement entirely.

The test I'd use for any hiring process is simple: run it on yourself. Would the person currently doing the job pass the interview they'd have to go through to get hired? If the answer is no, you're not measuring job performance. You're measuring something else. Figure out what.