ДаліНаступний посібник
Representation and Sampling Bias in Datasets
технічний
Технічний КЕРІВНИЦТВО
Intersectional bias appears when a system performs poorly for people at the intersection of multiple social categories, even when aggregate results for each category separately look acceptable.
Evaluating race and gender one at a time can hide higher errors for a combined group. Intersectional analysis helps reveal who is missed by broad averages and one-dimensional audits.
Intersectionality describes how social categories and systems of disadvantage can interact. In machine learning audits, an intersectional disparity is a performance or outcome difference for a combined subgroup that may disappear when results are aggregated by one attribute at a time. It is not enough to test “race” and “gender” separately if the largest error affects a particular race-by-gender group. The concept also cautions against treating categories as independent or interchangeable variables. Buolamwini and Gebru’s Gender Shades study examined three commercial gender-classification systems using a benchmark balanced by gender and skin type. The tested systems had their highest misclassification rates for darker-skinned women, up to 34.7%, while the maximum for lighter-skinned men was 0.8%. The study also found that two existing benchmarks were overwhelmingly lighter-skinned. These results show why subgroup composition and intersectional evaluation matter, but they apply to the tested systems, data, and binary gender-classification task at that time. They do not establish the accuracy of current products or validate gender inference as appropriate. Single-axis audits can mask compounded effects in hiring, lending, health care, and content moderation. A model can satisfy a threshold for each broad category while failing a smaller intersectional group. Evaluation requires enough examples to estimate performance and uncertainty; small samples can make metrics unstable and may expose sensitive attributes. Teams should justify which intersections are relevant, protect privacy, consult affected communities, and avoid presenting sparse estimates as definitive. Mitigation can include collecting better data with consent, improving measurement, changing the target or workflow, and testing multiple performance measures. No single metric captures every form of harm. Document limitations and provide appeal or human review for high-impact decisions. Intersectional analysis is a continuing practice, not a one-time check before release.
Архітектурні рішення збільшують продуктивність і експлуатаційні витрати протягом багатьох років.
Технічна освіта допомагає командам вибрати правильний стек, а не лише найновіший.
Кращий інженерний вибір зменшує проблеми з надійністю у виробництві.
Population categories and deployment contexts change. Repeat intersectional checks after model updates or expansion to new groups. Treat sparse estimates as uncertainty requiring better evidence or safeguards, not proof of equal performance. Maintain a route for people to report systematic errors; use appeal data to identify groups missing from the initial audit. Document groups that cannot be evaluated and reassess when collection methods or policies change. Track which intersections are covered by evidence and which remain untested as product reach changes.
A facial-analysis system performs reasonably for women overall and for darker-skinned people overall, yet has much higher errors for darker-skinned women.
A résumé ranker passes separate age and gender checks but downgrades older women more than either single-attribute analysis reveals.
A loan model shows similar approval rates by race and disability separately while outcomes are much worse for disabled applicants in one racial group.
A content generator reproduces stereotypes combining gender and ethnicity that single-attribute prompts did not surface.
Оптимізація одного тесту може приховати ширші слабкі сторони системи.
Витрати на інфраструктуру та обслуговування часто недооцінюються.
Прогалини в безпеці та спостережуваності можуть зростати в міру ускладнення систем.
Визначте цільові показники затримки, якості та вартості перед впровадженням.
Тест за реалістичних умов навантаження та даних.
Моніторинг інструментів на наявність помилок, дрейфу та впливу користувача.
Перед масштабуванням підготуйте шляхи відкату та реагування на інциденти.
Free newsletter
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
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
Intersectional bias appears when a system performs poorly for people at the intersection of multiple social categories, even when aggregate results for each category separately look acceptable. Evaluating race and gender one at a time can hide higher errors for a combined group. Intersectional analysis helps reveal who is missed by broad averages and one-dimensional audits.
Intersectional analysis examines performance for people at intersections of categories, which separate aggregate checks can conceal.
The study reported the highest misclassification rates for darker-skinned women.
The paper evaluates particular commercial systems and binary gender classification; its findings should not be generalized beyond that scope.
Broad one-attribute averages can hide a worse error rate for a combined subgroup.
Small intersectional groups can yield unstable estimates, so counts and uncertainty are needed.
Продовжуйте вчитися
Інші посібники, вибрані для цієї теми
ДаліНаступний посібник
Representation and Sampling Bias in Datasets
технічний