Some people have a neurological condition called face blindness (also known as ‘prosopagnosia’) which means that they are unable to recognise people, even those they know well – this can include their own face in the mirror! They only know who someone is once they start to speak but until then they can’t be sure who it is. They can certainly detect faces though, but they might struggle to classify them in terms of gender or ethnicity. In general though, most people actually have an exceptionally good ability to detect and recognise faces, so good in fact that we even detect faces when they’re not actually there – this is called pareidolia – perhaps you see a surprised face in this picture of USB sockets below.
What if facial recognition technology isn’t as good at recognising faces as it has sometimes been claimed to be? If the technology is being used in the criminal justice system, and gets the identification wrong, this can cause serious problems for people (see Robert Williams’ story in “Facing up to the problems of recognising faces“).
In 2018 Joy Buolamwini and Timnit Gebru shared the results of research they’d done, testing three different commercial facial recognition systems. They found that these systems were much more likely to wrongly classify darker-skinned female faces compared to lighter- or darker-skinned male faces. In other words, the systems were not reliable.
The Gender Shades Audit
Facial recognition systems are trained to detect, classify and even recognise faces using a bank of photographs of people. Joy and Timnit examined two banks of images used to train facial recognition systems and found that around 80 per cent of the photos used were of people with lighter coloured skin.
If the photographs aren’t fairly balanced in terms of having a range of people of different gender and ethnicity then the resulting technologies will inherit that bias too. Effectively the systems here were being trained to recognise light-skinned people.
The Pilot Parliaments Benchmark
They decided to create their own set of images and wanted to ensure that these covered a wide range of skin tones and had an equal mix of men and women (‘gender parity’). They did this by selecting photographs of members of various parliaments around the world which are known to have a reasonably equal mix of men and women, and selected parliaments from countries with predominantly darker skinned people (Rwanda, Senegal and South Africa) and from countries with predominantly lighter-skinned people (Iceland, Finland and Sweden).
They labelled all the photos according to gender (they did have to make some assumptions based on name and appearance if pronouns weren’t available) and used the Fitzpatrick scale (see Different shades, below) to classify skin tones. The result was a set of photographs labelled as dark male, dark female, light male, light female with a roughly equal mix across all four categories – this time, 53 per cent of the people were light-skinned (male and female).
The Fitzpatrick skin tone scale (top) is used by dermatologists (skin specialists) as a way of classifying how someone’s skin responds to ultraviolet light. There are six points on the scale with 1 being the lightest skin and 6 being the darkest. People whose skin tone has a lower Fitzpatrick score are more likely to burn in the sun and not tan, and are also at greater risk of melanoma (skin cancer). People with higher scores have darker skin which is less likely to burn and they have a lower risk of skin cancer.
Below it is a variation of the Fitzpatrick scale, with five points, which is used to create the skin tone emojis that you’ll find on most messaging apps in addition to the ‘default’ yellow.
Testing three face recognition systems
Joy and Timnit tested the three commercial face recognition systems against their new database of photographs – a fair test of a wide range of faces that a recognition system might come across – and this is where they found that the systems were less able to correctly identify particular groups of people. The systems were very good at spotting lighter-skinned men, and darker skinned men, but were less able to correctly identify darker-skinned women, and women overall.
These tools, trained on sets of data that had a bias built into them, inherited those biases and this affected how well they worked. Joy and Timnit published the results of their research and it was picked up and discussed in the news as people began to realise the extent of the problem, and what this might mean for the ways in which facial recognition tech is used.
There is some good news though. The three companies made changes to improve their facial recognition technology systems and several US cities have already banned the use of this tech in criminal investigations, and more cities are calling for it too. People around the world are becoming more aware of the limitations of this type of technology and the harms to which it may be (perhaps unintentionally) put and are calling for better regulation of these systems.
• Study finds gender and skin-type bias in commercial artificial-intelligence systems (11 February 2018) MIT News
• Facial recognition software is biased towards white men, researcher finds (11 February 2018) The Verge
• Go read this special Nature issue on racism in science (21 October 2022) The Verge
More technical articles
• Joy Buolamwini and Timnit Gebru (2018) Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, Proceedings of Machine Learning Research 81:1-15.
• The unseen Black faces of AI algorithms (19 October 2022) Nature News & Views
See more in ‘Celebrating Diversity in Computing‘
We have free posters to download and some information about the different people who’ve helped make modern computing what it is today.
Or click here: Celebrating diversity in computing
This blog is funded through EPSRC grant EP/W033615/1.