A new scale to curb skin tone bias in AI

The concern over FST was related to limited scale for darker skin. This leads to some facial recognition tech working well for golden brown skin but failing for certain red tones.

Tech firms have been using a skin tone scale developed in 1970s to grade colour of objects in an image. The Fitzpatrick Skin Type (FST) scale classifies skin tone based on six-point scale.

This measure has been used by several companies to know whether facial recognition systems, smartwatch heart-rate sensors, or AI cameras perform equally well across skin tones. But the FST scale lacked representation from darker skin tones.

The FST scale has four categories for “white” skin, and only one each for “black” and “brown” skins. Several experts recommended abandoning FST as it poorly represents colour in the latter two skin types.

Their concern over FST was related to limited scale for darker skin. This leads to some facial recognition tech working well for golden brown skin but failing for certain red tones.

In one study, Facebook researchers said FST does not encompass the diversity within brown and black skin tones. In a separate move, Microsoft acknowledged FST’s imperfections.

An increasing number of products now offer a far richer palette than the one developed in the 1970s. In 2020, Crayola launched 24 skin tone crayons, and a year later, Barbie dolls maker Mattel introduced nine tones.

Last June, Google said it is developing a new colour scale that will be more representative of the various skin tones. The search giant’s move came after a backlash within its own AI division.