Ghost in the Machine: The Unexpected Results of Not Understanding Your Users
summary
Claim: ML/NLP code can accidentally exclude 25% of users or only work 50% of the time for some groups. Evidence includes Hello Barbie misrecognizing speech, a Nikon camera labeling Asian faces as blinking, Microsoft’s Tay turning racist, biased criminal-risk scores that flagged darker-skinned people as higher risk, and Facebook ad targeting that excluded Jewish users. Tools and fixes named: high-fidelity simulations, gradual rollouts, PIN-AUC and open datasets for audits, crowdsourced labeling and manual bias checklists.
AI-written summary · about
chapters (5)
- 0:00Start
- 0:01Ghost in the Machine
- 22:20Make the problem well-defined
- 27:22Representation matters
- 30:46Teach your algos to recognize bias
related talks
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
41mWhen AI Failures Reflect Our Own Social Biases
38mSocial Implications of Bias in Machine Learning
39mBecoming Conscious of Our Unconscious Bias in Design
47mEthical Machine Learning
38mSocial Implications of Bias in Machine Learning
35mDesigning Inclusive and Accessible Experiences in the Era of AI
49mAn AI with an Agenda