The Social Cost of Strategic Classification

on YouTube ↗36m

" summary

Argues that raising classifier thresholds to optimize institutional accuracy via the Stackelberg equilibrium forces higher expected costs on truly qualified people by defining an individual burden and aggregating it into a social burden. Uses an outcome model with an outcome-monotonic cost assumption to reduce to threshold classifiers, proves institutional utility is quasi‑concave while social burden and a resulting social gap across subgroups are monotone increasing. Illustrates with Colombia welfare scores and FICO score manipulation.

AI-written summary · about

" chapters (11)

1. 0:00Intro
2. 2:04Means testing for social program eligibility
3. 4:17What we normally focus on
4. 6:29Improvement or gaming?
5. 7:43A causal perspective
6. 9:05Machine learning invites gaming
7. 13:16A formal model
8. 19:33Alternative measures of social cost
9. 21:39Lemma: Reduction to threshold classifiers
10. 26:56Measure of unequal burden
11. 29:41How can differences in cost arise?

" related talks

• 45mAGT and Machine LearningChristos Papadimitriou · Algorithmic Game Theory and Practice 2015
• 30mIndividual Statistical FairnessAaron Roth · Beyond Differential Privacy 2019
• 28mClassification Through Regression:Unlock the True Potential of Your LabelsTopaz Gilad · PyData Global 2022
• 29mEvaluation DesignAshia Wilson · Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026)
• 43mIn defence of uncertaintyAbeba Birhane · Code Mesh LDN 2019
• 38mTracking and Improving Information in the Service of FairnessMichael Kim · Beyond Differential Privacy 2019
• 14mFair And Reliable Machine Learning For High-Stakes Applications:approaches Using Information TheorySanghamitra Dutta · Meet the Fellows Welcome Event
• 35mOn Machine Learning for Prediction and Prioritization in the Allocation of Scarce Societal ResourcesSanmay Das · Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026)