Empirical methods fairer AI - Auditing a Dutch public sector risk profiling algorithm
summary
Audit of a Dutch education risk‑profiling pipeline uses aggregated population statistics (300k+ students) and GitHubed code to show a jump from 22% non‑European background at intake to 36% in the 42,800 person high‑risk group, 63.2% of 3,180 human‑selected controls, and 88.9% at appeal—near the media’s 97.6% claim—linking three proxy features to disparate outcomes. Methods center on a selection rate metric and privacy‑preserving protocol for lawful, actionable bias testing.
AI-written summary · about
chapters (12)
- 0:00Fair AI auditing introduction
- 2:28Dutch risk profiling algorithm
- 10:14Data collection for fairness
- 20:30Auditing challenges for Dutch DPA
- 25:55Sensitive data necessity for bias
- 30:05Fairness metrics and legal frameworks
- 44:14AI Act and DSA bias provisions
- 52:43Data access for bias testing
- 1:01:20Transparency in impact assessments
- 1:05:34Scaling analytic capacity
- 1:15:46Causation vs correlation in AI
- 1:20:19Closing and Q&A
related talks
29mResponsible AI Challenges and Recommendations
01h 10mWorking on Current and Future Governance and Control of AI and Algorithms
01h 16mBEYOND ETHICS WASHING:IMPACT ASSESSMENTS, AUDITS, AND OVERSIGHT FOR AI
23mRemoving Unfair Bias in the Machine Learning Pipeline
11mAddressing Privacy, Bias, and Fairness in AI Systems
40mEthical AI: Addressing Bias & Algorithmic Fairness
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
46mThe Measure and Mismeasure of Fairness