Fairness and Accountability Design Needs for Algorithmic Support in High-Stakes Public Sector ...
summary
An interview study with 27 practitioners across 5 OECD countries finds a practical gap: researchers offer fairness metrics and explainers but operational teams face shifting data, political constraints and institutional incentives. Examples include a geospatial policing model that shortens weekly model building to a day and relies on human augmentation, a deployed system that becomes a car-wash detector after investigative actions, and cases where performance tanked in four weeks due to feedback loops and untracked concept drift. Design needs center on usable transparency and drift-tracking tools.
AI-written summary · about
chapters (10)
- 0:00Introduction and context
- 2:13Public sector ML systems overview
- 3:58Canonical fairness problems
- 5:44Research focus and methodology
- 8:57Political challenges of transparency
- 11:24Data feedback loops and secondary uses
- 13:49User discretion and model reliance
- 15:36Institutional routines and adoption
- 17:19Conclusions and future needs
- 18:44Discussion and Q&A
related talks
40mQuestions and Practices to Make Algorithmic Decision Making More Fair
44mProcedural Justice in Algorithmic Fairness: Transparency, Control and Participation Perspectives
01h 10mJournalist-in-Residence Talk: The Limits of Fairness
01h 16mBEYOND ETHICS WASHING:IMPACT ASSESSMENTS, AUDITS, AND OVERSIGHT FOR AI
35mOpen Source Fairness
30mBias, Explainability, and Accountability: The Data Scientist's Burden
42mFairness in Machine Learning: From Theory to Practice
47mEthical Machine Learning