Designing More Inclusive AI with More Inclusive Workflows
summary
Finds that 90% of top "nurse" results are women and that this matches US census numbers, then questions whether matching a biased status quo is the right goal. Trains word2vec on raw search phrases and surfaces gendered associations (business→CEO, business feminine→fashion/secretary). Reweights search ranking to a global ethnic distribution for "wedding" and reports customers prefer the more diverse results. Builds custom evaluation sets from Getty Images data, encodes diversity into loss functions, and iterates.
AI-written summary · about
chapters (12)
- 0:00Inclusive AI workflows intro
- 1:12Human brain unconscious biases
- 3:20Bias and fairness in ML
- 6:27Why bias matters in AI
- 8:32Gender bias in search case study
- 11:00Customer data revealing biases
- 13:48Ethnic diversity in wedding images
- 15:55Fairness vision from leadership
- 16:37Best practices for inclusive AI
- 20:50Open source dataset challenges
- 21:53Inclusive imagery in communication
- 23:15Q&A
related talks
28mBreaking the Chains of Oppressive Software
22mGender Bias in Artificial Intelligence
52mKeynote address
38mSocial Implications of Bias in Machine Learning
11mAddressing Privacy, Bias, and Fairness in AI Systems
41mAll About Biases in Artificial Intelligence
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
30mWhy the Future of AI Is Feminist