(Improving) Muddled Information
summary
Heterogeneous gaming ability causes manipulable signals to decouple from true types, so raising incentives reduces information: a two-by-two model and a linear‑quadratic‑elliptical model both show higher stakes lower R-squared (information loss), pooling equilibria exist, and the Stackelberg designer optimally chooses a smaller sensitivity than the Nash fixed point. Remedies include lowering stakes (ban credit use in hiring), hardening features (use incoming links not site text), or flatten allocation rule via noise/regularization.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:00Allocation Problem
- 4:27Signaling Game
- 6:44Formal Language
- 9:39Observations
- 10:58Cost Function
- 11:19Natural Action
- 12:31Higher Gaming Ability
- 14:46Higher Natural Action
- 15:31Preliminary Observations
- 18:27Main Results
- 23:37Interventions
- 31:30Commitment
- 33:06Model
- 39:50Fixed Point
- 42:24Improving Information
related talks
14mInformation, Learning, And Incentive Design For Societal Networks
45mAGT and Machine Learning
42mRethinking AI Alignment and Evaluation under Heterogeneous Preferences
12mLearning In The Presence Of Strategic Agents: Dynamics, Equilibria, And Convergence
41mOne AI, two AI, red AI, blue AI: Impacts of Aggregation on Model Diversity and Consumer Utility
37mSelfish Robustness and Equilibria in Multi-Player Bandits
23mFalse Consensus, Information Theory, and Prediction Markets
38mTracking and Improving Information in the Service of Fairness