Skin Tone and Face Detection

The performance difference in face recognition by skin tone is among the best-documented findings in applied computer vision. Proctoring inherits it. Testing across representative users should follow the kind of implementation discipline outlined in the implementation overview, with failed recognition treated as a product defect rather than a candidate problem.

How it shows up

Failure to detect a face at check-in, producing repeated attempts before an exam starts, under time pressure.

Loss of face during the session, which generates flags for absence when the student never moved.

Identity match failure, which puts the student into a manual verification process at the worst possible moment.

And more flags overall, because a system that intermittently loses the face registers that as anomaly.

Why it happens

Training data historically over-represented lighter skin, and evaluation showed higher error rates on darker skin across many systems.

Detection depends on contrast, and default camera exposure settings are calibrated for lighter faces.

Poor lighting compounds it, and a student sitting an evening exam in a room lit by an overhead bulb is in exactly that condition.

What institutions can do

Ask suppliers for error rates disaggregated by skin tone, using a published scale.

A supplier that has not measured this has not tested for it.

Compare candidates on that figure, not only on headline accuracy.

What to do in operation

A quick human fallback at check-in, so a student who fails detection reaches a person in under a minute rather than retrying for ten.

Lighting guidance given in advance, plainly and without implying fault.

Reviewer awareness that repeated face-loss flags on one student may be a detection failure rather than behaviour.

And counting it: face-loss flags by student, because a student with twenty is experiencing a technical problem.

The harm beyond the flag

A student who has to prove their identity three times before an exam starts is disadvantaged in the exam itself.

And being repeatedly not-recognised by a system, in a high-stakes setting, is a specific experience that students describe as humiliating.

That harm occurs even when no case follows, and it is invisible in any measure based on outcomes.

The procurement question

This is one of the few areas where a supplier's answer is checkable against independent evaluations.

Published benchmark results exist for many underlying face recognition systems.

Ask which engine is used and look it up. Suppliers rarely volunteer this and usually will say if asked directly.

What to check

Do you know your system's error rate by skin tone?

How long does a failed check-in take to resolve, and does it reach a human?

Are face-loss flags counted per student?

And has anybody asked affected students about their experience of check-in?

The point

A student who has to prove their identity three times before an exam starts is disadvantaged in the exam itself, and that harm occurs even when no case follows.

It is invisible in any measure based on outcomes.

Worth stating

Provide a quick human fallback at check-in so a student who fails detection reaches a person in under a minute rather than retrying for ten.

Count face-loss flags per student, because a student with twenty is experiencing a technical problem.

Also worth knowing

Detection depends on contrast, and default camera exposure is calibrated for lighter faces.

A student sitting an evening exam under an overhead bulb is in exactly the condition where the difference bites hardest.

And finally

Ask suppliers for error rates disaggregated by skin tone using a published scale, and compare candidates on that figure rather than on headline accuracy.

A supplier who has not measured it has not tested for it.

Summary

Provide a human fallback that resolves a failed check-in in under a minute, and count face-loss flags per student so technical problems are visible as technical.

In summary

Face detection performs measurably worse on darker skin, which produces failed check-ins, lost-face flags and identity challenges at the worst possible moment.

The harm occurs even when no case follows. For wider institutional context, consult the National Institute of Standards and Technology.