Testing for Cognitive Bias in AI Systems
summary
Claim: AI inherits human bias from training data and can produce consistent but wrong outcomes. Evidence: the Amazon résumé crawler that favors men and a Microsoft Twitter bot that learned abusive language. Methods shown: design many edge-case test cases, measure confidence not binary pass/fail, run canary release and side-by-side human comparisons, retrain with diverse data. Recommendations: map hypotheses to requirements, watch for confirmation bias, expect to rework datasets and budget far more testing than conventional software.
AI-written summary · about
chapters (13)
- 0:00Cognitive bias in AI testing
- 2:21Origins of bias in AI
- 5:09Real-world AI bias examples
- 8:43Bias impact across domains
- 10:08Testing for bias challenges
- 13:38Holistic bias testing approaches
- 16:35Key takeaways and data selection
- 17:20Q&A: Political bias and trust
- 18:06Q&A: Human bias criteria
- 18:48Q&A: Cognitive biases in AI
- 19:50Q&A: AI testing AI for bias
- 20:32Q&A: Stakeholder buy-in
- 21:58Q&A: Bias risks in tools
related talks
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
35mZero Bias AI
47mFight AI Bias With… Bias
31mAvoid mistakes building AI products
21mFairness and bias in AI
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
49mAn AI with an Agenda
47mYour AI Is Still Biased (Even After You Checked)