Ethical AI
summary
Analysis of commercial automatic gender recognition APIs shows error rates drop to near-random for darker-skinned women while lighter-skinned men have much lower errors, prompting a claim that some tasks should be retired rather than only relabeled. Methods include auditing paid APIs, diversifying training sets, partnering with archives and community groups, and proposing data sheets for datasets and community-participatory archives; recommended processes are practical and partially implemented but require institutional hiring and consortiums to scale.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 11:05Algorithmic Fairness
- 12:00Respect Data
- 12:30Predatory Inclusion
- 13:20Anthropology
- 13:35Archives
- 14:35Data Collection Spectrum
- 16:45Data Collection Plan
- 17:45Archive Concerns
- 19:30Community Participatory Archive
- 20:25Partnership in AI
- 23:55Exclusivity
- 24:35Conclusion
related talks
22mGender Bias in Artificial Intelligence
47mEthics & AI: Identifying & preventing bias in predictive models
27mThe Limitations and Danger of Facial Recognition
37mThe AI 'Black Box'
34m(Un)Ethical Machines? Why AI Bias Is a Problem and What to Do About It
47mEthical Machine Learning
41mMeasures and Mismeasures of Algorithmic Fairness
56mGlobal Bias in AI: When Western Data Shapes the World