Reflective, not Objective: Intentionality in AI
summary
Claims that AI is neutral get overturned with concrete failures: machine learning is just algorithms trained on data and often reproduces bias, e.g. word embeddings encode gender stereotypes so word embeddings map man:computer programmer :: woman:homemaker, a coreference model links "programmer" to his but breaks for "her", facial and speech systems fail on darker skin and accents, and a recidivism risk model labels Black defendants as higher risk yet they do not reoffend at nearly twice the rate of whites; fixes include re‑sampling data, auditing, and interpretable models.
AI-written summary · about
chapters (24)
- 0:00Intro
- 0:34About me
- 1:20Twitter Bots
- 2:53Outline
- 4:41Algorithms
- 5:54Applications
- 7:17Machine Learning
- 8:16Machine Learning Pipeline
- 9:33Prediction Model
- 11:33Kathy ONeal
- 12:16Why does this matter
- 13:09Bias in AI
- 14:03Data imbalance
- 14:54Uncertainty bias
- 15:35Implicit systemic bias
- 16:09Word embeddings
- 17:17NLP example
- 18:11Feature representation
- 19:31Overfitting
- 20:26Chatbots
- 21:02Cost function
- 22:29Algorithms are designed by people
- 23:20AI is reflective
- 24:58Conclusion
related talks
34m(Un)Ethical Machines? Why AI Bias Is a Problem and What to Do About It
05mToward Ethical AI
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
41mWhen AI Failures Reflect Our Own Social Biases
35mZero Bias AI
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
59mIn search of inclusive AI
38mSocial Implications of Bias in Machine Learning