en.wikipedia.org/wiki/Timnit_Gebru
1 correction found
The pair investigated facial recognition software, finding that in one particular implementation Black women were 35% less likely to be recognized than White men.
Gebru and Buolamwini’s Gender Shades paper did not measure whether Black women were "recognized" less often than White men. It evaluated commercial gender-classification systems and found much higher misclassification rates for darker-skinned women.
Full reasoning
The cited research here is Gender Shades by Joy Buolamwini and Timnit Gebru. That paper was not a study of whether people were recognized or identified as specific individuals. It explicitly studied commercial gender classification systems.
The paper’s title is "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification." Its abstract says the authors "evaluate 3 commercial gender classification systems" and found that "darker-skinned females are the most misclassified group (with error rates of up to 34.7%)." The comparison group was lighter-skinned males, whose maximum error rate was 0.8%.
MIT’s own summary of the work says the systems’ task was determining gender, not identifying or recognizing a person: the systems had error rates of 0.8% for light-skinned men and up to 34.7% for dark-skinned women in gender classification.
So the article’s wording is incorrect in two ways:
- It says the study investigated facial recognition software, when the paper evaluated facial-analysis / gender-classification systems.
- It says Black women were less likely to be recognized, but the measured outcome was whether the systems correctly classified gender.
Those are materially different tasks in computer vision, and the published sources describe this work as gender-classification bias, not person-recognition performance.
2 sources
- Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification
We evaluate 3 commercial gender classification systems using our dataset and show that darker-skinned females are the most misclassified group (with error rates of up to 34.7%). The maximum error rate for lighter-skinned males is 0.8%.
- Study finds gender and skin-type bias in commercial artificial-intelligence systems | MIT News
In the researchers' experiments, the three programs' error rates in determining the gender of light-skinned men were never worse than 0.8 percent... For darker-skinned women, however, the error rates ballooned ... to more than 34 percent.