skip to content
‹ All posts

By Riya Madaan

Proctoring is an arms race you're funding

Interview proctoring pits detection against generation, and detection always trails. Every false positive burns a candidate you wanted. Stop funding it.

#opinion

Interview proctoring is a detection game, and detection games always favor the evader: the defender has to win every time, the evader only has to win once.

Somewhere right now, a senior engineer is showing her webcam a 360-degree pan of her bedroom to prove she isn't hiding a second monitor, so that a company can trust the results of a multiple-choice quiz.

This is interview proctoring in 2026, and the strangest part is that everyone involved knows it isn't working. The candidates know, because the workarounds are a genre of content. The vendors know, because the roadmap is an endless list of counter-countermeasures. The buyers know, because the whole point of buying proctoring is that they didn't trust the assessment - and afterward, they still don't. The only thing the money reliably purchases is candidate resentment.

The structure of the interview proctoring arms race

Interview proctoring is a detection game, and detection games have a shape. The defender ships a detector; the attacker studies it and routes around it; repeat. Eye-tracking gets defeated by screen placement. Browser lockdown gets defeated by a second device. AI-writing detection trails AI writing by definition, because generation models train faster than classifiers can chase them, and every published detector is a training target.

The asymmetry is brutal. The evader needs to work once; the detector needs to work every time. And in hiring, the defender pays for both sides: you fund the vendor directly, and you fund the evasion indirectly through burned candidates, lengthened processes, and the reputational cost of treating professionals like exam-hall teenagers.

Arms races can still be worth running when the thing defended is priceless. So it's worth asking what's actually being defended here. Usually: a question bank. Reusable questions with known answers, whose entire value depends on secrecy. The proctoring industry exists to protect the secrecy of assets that are already on Reddit. You are hiring guards for an empty vault.

False positives are the real product

Every detector has a false-positive rate, and in proctoring the false positives are people.

Think about what a proctoring flag actually is: an accusation of fraud, generated by a statistical model, attached to a specific human being who wanted to work for you. The candidate who glanced away because they think while looking at the ceiling. The candidate whose internet hiccuped. The candidate whose eye movements or typing rhythm don't match the model's training distribution - and it's worth sitting with the question of which candidates that systematically describes.

When the flag is wrong, what's the appeal process? Mostly there isn't one. The candidate is quietly rejected with a scarlet letter in an ATS field they'll never see. When it's right, what did you win? You caught someone gaming a format so gameable it required surveillance to administer. The format was the vulnerability. The proctoring was a confession.

And your best candidates - the ones with competing offers, the ones you built the pipeline for - see the webcam requirement and simply decline. Candidate experience is a funnel metric; surveillance is a tax on it that the strongest candidates don't have to pay, so they don't.

Surveillance vs. evidence

Here's the distinction that gets lost. Proctoring watches the person. It records faces, rooms, keystrokes, and gaze, hunting for signs of deception. Evidence records the work. What was tried, in what order, with what result.

The first is adversarial by construction - it starts from the assumption that the candidate is probably lying and searches for proof. The second is neutral: it's just the truth of the session, equally available to exonerate and to inform. A candidate can be furious about being surveilled; nobody is furious that the interview happened and the record shows it.

The deeper point: if your assessment needs surveillance to be trustworthy, the assessment is the problem. A quiz needs proctoring because the answers exist outside the room. A real broken system doesn't, because there's nothing to smuggle in. The fault is live, the answer is a working system rather than a string you can paste, and the pass condition is "the system actually recovered" - checked by scripts against ground truth, not by a human squinting at a webcam feed. (Randomized variants are on our roadmap to close the remaining gap.) Even AI assistance stops being contraband and becomes part of what's measured. When cheating is structurally pointless, the entire surveillance apparatus becomes weight with no wings.

This is why Faultybox logs commands and never cameras - a record of what happened in the environment, nothing of the human's face or room. We sell no proctoring and no AI detection, and this is a position, not a gap. The recording exists to show the work, and a candidate can run the same labs on their own account and keep that record, which tells you whose side the format is on.

Exit the race

You can't win an arms race, but you can decline to enter it. The move is to stop defending secrets and start generating evidence: assessments where the task is the job, the answers don't exist in advance, and the record is the work itself.

Every dollar spent watching candidates' eyes is a dollar not spent building an interview worth taking. The vendors will keep selling the race as long as you keep funding it. Stop funding it.


Faultybox runs your candidates through a real broken cluster and shows you exactly how they fixed it - command log included. Start free →