Facial Recognition
Overview

Enrollment photos
23

7

2

5
4

Joy Buolamwini, MIT

Dr. Timnit Gebru, Google

Woman
Dark Skin

Woman
Light Skin

Man
Dark Skin

Man
Light Skin

2018 MS
Face API
Error Rate

20.8%

1.7%

6.0%

0.0%

2019 MS
Face API
Error Rate

1.5%

0.3%

0.3%

0.0%

Microsoft

IBM

Face++

Error rate:

20.8%

34.7%

34.5%

Error rate:

1.52%

16.97%

4.1%

Change (ppts)

-19.28

-17.73

-30.4

Buolamwini & Gebru, 2018

Buolamwini & Raji, 2019

"By highlighting the issue of classification performance disparities and amplifying public
awareness, the study was able to motivate companies to prioritize the issue and yield significant
improvements within 7 months." Raji & Buolamwini, 2019

Facial Recognition Accracy of Humans
AUC (.5 is "random" and 1 is perfect)

Facial Recognition Accuracy of Humans + Algos
AUC (.5 is random and 1 is perfect)

Facial examiners

0.93

Facial reviweers

0.87

Superrecognizers

0.83

Fingerprint examiners

Students

0.76

0.68

One examiner + best algorithm

1

2 facial examiners

0.96

Best algorithm (of 4)

0.95

One examiner + 2nd best algorithm

0.95

1 facial examiner

0.93

From: Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms, Phillips et al., 2018

1. Fairness. We will work to develop and deploy facial recognition technology in a manner that
strives to treat all people fairly.
2. Transparency. We will document and clearly communicate the capabilities and limitations of
facial recognition technology.
3. Accountability. We will encourage and help our customers to deploy facial recognition
technology in a manner that ensures an appropriate level of human control for uses that may
affect people in consequential ways.
4. Non-discrimination. We will prohibit in our terms of service the use of facial recognition
technology to engage in unlawful discrimination.
5. Notice and consent. We will encourage private sector customers to provide notice and secure
consent for the deployment of facial recognition technology.
6. Lawful surveillance. We will advocate for safeguards for people's democratic freedoms in law
enforcement surveillance scenarios and will not deploy facial recognition technology in
scenarios that we believe will put these freedoms at risk.