How I'm fighting bias in algorithms | Joy Buolamwini
Joy Buolamwini uses repeated failures of facial-analysis systems to make algorithmic bias concrete, then traces harm from training data and development teams into policing, credit, hiring, and other institutions. The talk is historically important; its statistics and regulatory descriptions remain dated to 2016.
Plan for this page
Watch actively with preparation, signposts, and a comprehension check.
You will leave with
How bias can enter through data, design, teams, and institutional use, and why errors at scale can become exclusion or discrimination.
Time
9 min
Before you begin
General learners, builders, and decision-makers who need a human-centered entry into bias and fairness.
Do this now
Read the terms and attention points before loading the video.
How I'm fighting bias in algorithms | Joy BuolamwiniHow bias can enter through data, design, teams, and institutional use, and why errors at scale can become exclusion or discrimination.
00:00 · The coded gaze
00:45 · A white mask makes the face visible
01:15 · The failure repeats
03:15 · Training examples shape recognition
04:08 · From exclusion to discrimination
What it teaches
How bias can enter through data, design, teams, and institutional use, and why errors at scale can become exclusion or discrimination.
General learners, builders, and decision-makers who need a human-centered entry into bias and fairness.
Pay attention to
Follow the path from one person’s failed detection to a reused software component around the world.
Separate a technical error rate from the institutional consequence of acting on that error.
Notice that the proposed response includes people, process, purpose, auditing, and data—not only model accuracy.
Bias can enter at problem framing, data, labels, model, threshold, interface, institution, or feedback—not at only one stage.
Average accuracy can hide concentrated errors and unequal consequences for particular groups.
Fairness requires named groups, outcomes, baselines, participation, audits, oversight, correction, and recourse.
Active check
Why is improving dataset diversity useful but insufficient by itself?
Caveats
The talk was delivered in 2016; its policing figures, regulatory descriptions, websites, products, and institutional context should not be treated as current without checking.
The central examples concern face detection and facial analysis. Other AI systems can produce bias through different data, objectives, proxies, thresholds, workflows, and institutions.
More representative data can improve some performance gaps but may also expand surveillance; fairness cannot be reduced to collecting more faces.
Accessibility
Manually supplied English and French caption tracks and an official English transcript were available. The face-detection demonstration is visually important, but the speaker narrates the key contrast; applause and audience reaction are marked in the transcript.
How I'm fighting bias in algorithms | Joy BuolamwiniTED / TEDxBeaconStreet
The exact reviewed video, delivered in November 2016 and published on YouTube in March 2017. It combines personal evidence, institutional examples, and advocacy.
inAi