Global Bias in AI: When Western Data Shapes the World
summary
Shows that models trained on Western-heavy sources actually fail specific groups: image generators (Stable Diffusion trained on LAION) output >80% darker-skinned inmates and underrepresent women in high‑paying roles; face systems still mislabel Black people and error rates reach 34.7% for darker‑skinned women; medical devices and datasets (pulse oximeters, Fitzpatrick scale) miss clinical signs in darker skin; data annotators in Kenya and elsewhere are paid 1–5 cents per label and face trauma; roughly 49% of online content is English, creating a large representation gap.
AI-written summary · about
chapters (14)
- 0:00Introduction and context
- 3:31Personal anecdotes on AI
- 7:31Data representation gaps
- 10:20The human side of AI training
- 14:59Architecture of bias
- 17:37Model collapse and feedback
- 20:26The WEIRD bias problem
- 22:00Case studies in imagery
- 26:57Medical AI and diagnostics
- 33:24Fintech and global economies
- 37:28Algorithmic colonialism
- 42:38Societal impacts of automation
- 46:09Ethical frameworks and governance
- 52:43Closing thoughts
