An AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
summary
Humans’ biases routinely enter models because training data records past decisions and social patterns, producing concrete harms: ProPublica study shows risk scores no better than chance and Amazon’s hiring system repeatedly fails to remove gender bias. Practical fixes include IBM AI Fairness Toolkit and What‑If tool for auditing, adversarial debiasing and reweighting to change training distributions, and distributionally robust optimization to protect minority subgroups. Known failure modes include pricing bots that learn to collude, simulation hacks like the half‑cheetah exploiting boundaries, and predictive‑policing feedback loops that amplify arrests.
AI-written summary · about
chapters (22)
- 0:00Intro
- 4:11REAL CONSEQUENCES
- 10:55KNOW WHAT QUESTION YOU'RE ASKING UP FRONT
- 11:23USE CONDITIONAL PROBABILITY OVER CORRELATION
- 12:40MORTGAGE LENDING ANALYSIS
- 20:52BUILD A BETTER DATA SET!
- 21:26BEWARE SHADOW COLUMNS
- 22:33MAKE SURE YOUR SAMPLE SET IS REPRESENTATIVE
- 29:09KEEP IN MIND YOU NEED TO KNOW WHO CAN BE AFFECTED IN ORDER TO UN-BIAS
- 29:53PRICING ALGORITHMS
- 32:53WHAT IF AMAZON BUILT A SALARY TOOL INSTEAD?
- 33:32THE BRATWURST PROBLEM
- 37:25MORE COMPLEX ALGORITHMS THAT INCLUDE OUTSIDE INFLUENCE
- 40:29VERIFY AND CHECK SOLUTIONS DERIVED FROM SIMULATION
- 42:19MANY AI/ML TOOLS ARE TRAINED TO MINIMIZE AVERAGE LOSS
- 43:25REPRESENTATION DISPARITY
- 45:09DISTRIBUTIONALLY ROBUST OPTIMIZATION
- 46:00WHAT HAPPENS TO PEOPLE WHO USE DIALECT?
- 46:25PREDICTIVE POLICING
- 49:09LOOK TO CONTROL ENGINEERING
- 54:22ABIDE BY ETHICS GUIDELINES
- 54:34Transparency of Use Transparency of Algorithms
related talks
49mAn AI with an Agenda
34m(Un)Ethical Machines? Why AI Bias Is a Problem and What to Do About It
21mFairness and bias in AI
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
47mFight AI Bias With… Bias
05mToward Ethical AI
35mZero Bias AI
30mFighting Prejudice in Artificial Intelligence