Testing a vision model for bias across skin tones before you ship it

Face and vision models trained on unbalanced datasets perform unevenly across skin tones. What to test for, deliberately, before trusting a model in production.

Written by

Pranjul Rathour

GenAI Engineer at SCULT INDIA, based in Kanpur, India. Builds production RAG systems, fine-tunes open LLMs and ships agentic AI products; has mentored 200+ college developers. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks.

pranjulrathour41@gmail.com LinkedIn GitHub X Instagram Bluesky Dev.to

In a packed college auditorium

Face detection and recognition models trained on datasets skewed toward lighter skin tones have a well-documented history of performing worse on darker skin tones. This isn't a hypothetical risk — it's a measurable, testable failure mode.

What to actually test

Detection and matching accuracy broken out by skin tone group in your test set, not just an aggregate accuracy number.
Performance under the lighting conditions where the gap tends to widen further — low light disproportionately affects darker skin tones in many models.
Whether your enrolment/training data itself is balanced, since an unbalanced test set can hide a real gap.

If a gap is found

Document it honestly rather than shipping silently, and prefer a model or vendor with published fairness evaluations if the application is anything consequential — access control, identity verification — where a false accept or reject has real cost.

See evaluating a vision model beyond accuracy.

— Pranjul Rathour, GenAI Engineer from Kanpur, India. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at any campus: pranjulrathour41@gmail.com.