The Case for Auditing Algorithms: Disloyalty and Bias M
summary
INRIA builds regalia audits that probe black boxes by behavioral testing and by constructing black-box surrogate cloning to mimic a service without triggering platform adaptations. A real bug where alphabetical agent IDs produced worst shifts motivates looking for measurable harms. Regulators find 40% of commercial sites with unfairness issues and France lists about 25 sensitive variables to guard. Techniques include blocking sensitive fields (insufficient), matching and reweighting to minimally repair outcomes, and legal and scientific limits on exhaustive probing.
AI-written summary · about
chapters (7)
- 0:00Introduction to algorithm auditing
- 3:15Regulation and legal pressure
- 5:47Challenges in detecting bias
- 9:20Solutions and future hope
- 10:24Developer responsibility and ethics
- 14:18Sensitive variables in segmentation
- 16:50Reporting unethical practices
related talks
01h 12mInvestigating algorithmic systems - Algorithm auditing.
01h 21mEmpirical methods fairer AI - Auditing a Dutch public sector risk profiling algorithm
53mIt’s Not Fair! Detecting Algorithmic Bias with Open Source Tools
50mEthical Machine Learning
01h 12mObscured by algorithms: how to prove "yes" if computer says "no"
47mAlgorithmic bias is everywhere (especially at Breeze) - what can we do about it?
01h 06mPanel on Interpretability in the Law
33mPath to a more equitable testing process: Addressing unconscious bias and ethics