Reducing Bias in AI with Open Source
summary
AI models skew toward Western perspectives because training data and tokenization favor English — 7 tokens vs 16 tokens for the same sentence in English and Vietnamese — and commercial incentives ignore low-use languages. Real harms include facial‑recognition misidentification, hiring models reproducing gender bias, and paid annotators in Kenya earning $2/hour annotators whose work contributed to harmful outputs and distress. Open‑source practices proposed: publish datasets and models (Hugging Face, LLaMA), fund shared GPU pools, add mentorship and governance for 54 countries 200 languages representation and explainability.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 3:19AI bias concerns
- 7:01AI bias examples
- 10:58Open source for bias reduction
- 19:32Challenges in open AI models
- 28:27Open floor discussion on solutions and ideas
- 42:44Summary
related talks
43mReducing Bias in AI with Open Source
05mWhat Do We Mean When We say Open Source AI?
11mMachines of Love & Grace – Building Truly Inclusive AI
05mToward Ethical AI
19mMoral Outsourcing: Humanity in the age of AI
26mBuilding ethical AI: Transparency, bias mitigation, and trust
40mCulturally-Inclusive AI: Enhancing User Experience With Global Perspectives
26mPutting AI to Work on Our Biggest Challenges