Afternoon Session #1: Science, Pseudo-science, Canon and Belief
Talkat Wrong at the Root: Racial Bias and the Tension Between Numbers and Words in Non-Internet Data · 2019-06-05..06-07·
summary
Google’s 2019 diversity numbers (tech hires: 25.7% women, 4.8% Black, 6.8% Latino, 1.1% Native American) set a context for concrete failures: MEDLINE/MeSH still indexes Caucasoid, Negroid, Australoid and MeSH on‑demand/MTI can resurrect those legacy terms; ProPublica finds the COMPAS recidivism tool skews against Black defendants; public‑health data tie toxins and infection to cognition—lead exposure costs 23 million IQ points per cohort per year; NLP word embeddings encode gendered analogies (doctor→nurse) and can be corrected by projecting out a gender subspace, but legal scholars warn against outsourcing judgment to fMRI/IAT metrics.
AI-written summary · about
chapters (7)
- 0:00Personal and professional context
- 4:26Bias in medical classification
- 14:10The persistence of outdated categories
- 23:40Technology as a change agent
- 28:22Technophilia and behavioral realism
- 53:41Race, intelligence, and environmental factors
- 1:13:45Algorithmic bias in language models
related talks
28mBreaking the Chains of Oppressive Software
38mSocial Implications of Bias in Machine Learning
41mEthics Panel
37mThe AI 'Black Box'
47mDiversity and Inclusion for Algorithmic Bias Considerations
31mInclusive Storytelling With Data
01h 00mThe AI ethics session nobody knew we needed
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....