本頁4 分鐘閱讀
概述
It found their gender classification was far less accurate for darker-skinned women than for lighter-skinned men. The study matters because it made intersectional auditing, which breaks results down by combined attributes rather than one at a time, a standard way of exposing hidden AI failures.
深入探討
Joy Buolamwini, then at the MIT Media Lab, noticed that face detection software did not register her face until she put on a white mask. With Timnit Gebru, she designed a structured audit and published it at the 2018 Conference on Fairness, Accountability and Transparency. Existing benchmarks were dominated by lighter-skinned men, so they built their own, the Pilot Parliaments Benchmark. It contains 1,270 images of members of parliament from three African and three European countries, labeled by binary gender and by skin type on the dermatologists' Fitzpatrick scale. They tested the gender classification services of IBM, Microsoft and Face++. All three did better on men than on women and better on lighter skin than on darker skin. The worst results appeared where the two overlapped. Error rates for darker-skinned women reached about 35 percent for the worst system, compared with under 1 percent for lighter-skinned men. A single overall accuracy figure hid this gap. A 2019 follow-up by Inioluwa Deborah Raji and Buolamwini, "Actionable Auditing," found that the audited companies had narrowed their gaps after being publicly named. Vendors that had not been audited, including Amazon, showed similar disparities. Amazon disputed the methodology. A common misconception is that Gender Shades measured the face identification used by police. It measured gender classification. Evidence on identification came from NIST's 2019 demographic effects report, which tested around 200 algorithms. Many had higher false positive rates for some groups, including African and East Asian faces, often by large factors, while the most accurate algorithms showed much smaller differences. Real harm followed. Robert Williams, Nijeer Parks and Porcha Woodruff, all Black, were each wrongly arrested after facial recognition leads. In June 2020, IBM said it would leave the general-purpose facial recognition business, Amazon paused police use of Rekognition, and Microsoft said it would not sell to US police until a federal law was in place.
戰略影響
風險與安全
災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
更明確的決策
民眾和專業素養決定強而有力的安全政策在政治上是否可行。
突破炒作
清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
The Future of Gender Shades and Facial Recognition Bias Audits
Face recognition has become more accurate on average, and NIST's continuing tests show demographic differences shrinking for leading algorithms, though not disappearing, and staying large for many others. Policy remains fragmented. Some US cities restrict government use, the EU AI Act sharply limits real-time remote biometric identification in public spaces by police, and several vendors have withdrawn some products. Wrongful arrest cases continue to push police departments to treat matches as leads only. The Gender Shades method, publishing disaggregated results and naming vendors, is now widely used for auditing other AI systems, including language and image generators.
現實世界的實施
A product team reports face verification accuracy separately for darker-skinned women, darker-skinned men, lighter-skinned women and lighter-skinned men, instead of publishing one headline number that can hide a failing subgroup.
Robert Williams, a Black man in Detroit, was arrested in 2020 after facial recognition matched his driver's license photo to shoplifting footage. The charges were dropped, and a later settlement changed Detroit police rules on using such matches.
Before approving a vendor, a city procurement office checks NIST's demographic test results for the specific algorithm version on offer, because error differences vary widely between algorithms.
A researcher builds a test set balanced by skin type and gender, following the Pilot Parliaments Benchmark approach, to check a new face model for gaps before release.
風險與防護欄
將存在風險視為科幻小說，同時能力複合。
混淆了表面產品安全與高度自治下的對準。
只給非英語和非專業觀眾留下低品質的資源。
實施路線圖
單獨的產品危害、誤用和失控/失調風險。
詢問哪些證據會改變您對時間表和嚴重性的看法。
比起行銷主張，更喜歡主要來源和具體評估。
確定一條行動路徑：職業、政策、資金或技能——而不僅僅是意識。
不斷探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Gender Shades and Facial Recognition Bias Audits quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常見問題
What is Gender Shades and Facial Recognition Bias Audits?
Gender Shades is a 2018 study by Joy Buolamwini and Timnit Gebru that audited commercial face analysis systems from IBM, Microsoft and Face++. It found their gender classification was far less accurate for darker-skinned women than for lighter-skinned men. The study matters because it made intersectional auditing, which breaks results down by combined attributes rather than one at a time, a standard way of exposing hidden AI failures.
Which companies' face analysis services did the original Gender Shades study audit?
The 2018 study tested IBM, Microsoft and Face++. Amazon was examined in the 2019 follow-up.
What was the Pilot Parliaments Benchmark made of?
The benchmark used 1,270 images of parliamentarians from six countries, chosen to balance gender and skin type.
Which scale did the researchers use to label skin type?
They used the Fitzpatrick scale from dermatology. Some later work uses broader scales such as the Monk scale.
What task did Gender Shades actually measure?
The study measured gender classification. A common misconception is that it measured the identification systems used by police.
Which subgroup had the highest error rates in Gender Shades?
Errors peaked where the two attributes overlapped, at darker-skinned women, reaching about 35 percent for the worst system.
繼續學習
相關指南
為此主題精選的更多指南