The Limitations and Danger of Facial Recognition
summary
Commercial systems show 20–35% higher error rate on darker‑skinned women, prompting vendors to publish fixes like IBM’s Diversity in Faces dataset (built partly from ethically dubious photo scraping). Algorithmic audits use curated corpora — one paper analyzes 2400 Instagram images across seven gender tags and finds trans and nonbinary people misgendered 10–30% more. Law‑enforcement misuse (celebrity‑lookalike leads), government tracking of minorities, and deployment failures in housing show what breaks and who gets harmed.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:21DISCLAIMER
- 0:46Face Recognition in the News
- 2:31Phrenology v2.0
- 4:01Why Algorithmic Accountability Matters?
- 5:56Gender Shades Project
- 7:28Gender Shades - Impact
- 9:09Auditing Facial Recognition Systems
- 11:41Common Emotions & Micro-Expressions
- 12:42Facial Action Coding System
- 13:52Why Emotion Recognition won't scale
- 16:23The Misgendering Machine
- 17:57Computer Vision and Gender
- 19:53THE AGE OF SURVEILLANCE
- 20:39Misuse of Face Recognition
- 22:26Technology and Power
- 25:05When Not to Build, Design, or Deploy
related talks
45mThe Myth of Neutrality: How AI is widening social divides
28mEthical AI
47mEthics & AI: Identifying & preventing bias in predictive models
41mMeasures and Mismeasures of Algorithmic Fairness
54mWe are the guardians of our future
31mDoes This Algorithm Have a Soul?
41mYeah, But Should We....
37mThe AI 'Black Box'