An auto-correction whose target is measured on skin has a skin tone baked into it

Fact

Face detection makes it easy to meter on the subject instead of the scene, and the obvious next step — "drive exposure until the face sits at luma 0.55" — is a constant that only one complexion satisfies. Measured over cheek and nose in Rec.601 luma, ordinary skin under the same light spans roughly:

very light   0.78      medium   0.63      deep   0.33

A single target therefore is a target skin tone: it lifts dark faces and pulls down pale ones, and does so most confidently when detection is working best. The same holds for white balance keyed on skin. Skin is red-leaning by construction, so mean(Cr) − mean(Cb) runs about 0.21…0.31 across complexions while a genuinely neutral grey card sits near 0; a "neutral skin" target reads a warm complexion as a warm lamp and desaturates the person.

Why it matters

Nothing looks broken. The correction is smooth, the code has no branch on anything sensitive, and on the developer's own face it is visibly right — which is how this ships. A consistently mis-exposed preview then reads as a bad camera rather than as a bad constant.

How to apply

• Split the measurements by what they are a property of. Exposure, dynamic range, clipping and colour cast belong to the light: measure them over the whole frame. Only quantities that describe variation — texture energy, chroma spread, local contrast — are safe to measure on skin, because they say the same thing about any skin.
• Where a skin-referred correction is genuinely wanted, act on the excess outside a deadband wider than the spread of complexions, so the common case is no correction at all.
• Prefer clipping to level as a tone-independent signal.
• Test it directly. Shift the skin statistics across a wide tonal range and assert the exposure, contrast and colour outputs are identical; that is a property the implementation either has or does not.