Michael Kearns: Algorithmic Fairness, Privacy & Ethics | Lex Fridman Podcast #50

Michael Kearns' Background and Book Focus

• Kearns is a University of Pennsylvania professor and co-author of An Ethical Algorithm.
• The book addresses algorithmic fairness, bias, privacy, and ethics, bridging machine learning with game theory and computational social science.
• Kearns transitioned from an English major to computer science after realizing an English degree taught reading but not writing.
• He cites David Foster Wallace's Infinite Jest and his uncle, a moral philosopher, as significant intellectual influences on his work.

The Scope and Limitations of Algorithmic Ethics

• Algorithmic solutions can fix "low-hanging fruit" issues like blatant privacy violations or extreme unfairness but cannot resolve upstream data biases (e.g., racist policing data).
• Fairness vs. Privacy: Privacy has a settled technical definition (differential privacy), whereas fairness lacks a universal definition and involves unavoidable trade-offs.
• Impossibility Theorems: Research shows three desirable fairness metrics cannot be simultaneously satisfied in all scenarios.
• Subjective Fairness: Current definitions are "received wisdom" from scholars; the field lacks extensive human-subject data on how ordinary people perceive fairness.
• Group vs. Individual Fairness: Optimizing for marginal group fairness (e.g., by race or gender independently) can lead to "fairness gerrymandering," where intersectional groups (e.g., disabled Hispanic women over 55) face discrimination despite individual protections.
• Pareto Curves: Kearns advocates presenting stakeholders with Pareto curves that visualize the numerical trade-off between system accuracy (error rates) and unfairness, allowing policymakers to choose a point on the curve rather than researchers deciding for society.
• Human Oversight: Algorithms should not autonomously define social norms; humans must inject values into the objective function, though computer scientists must increasingly engage in the philosophical debates surrounding these choices.

Privacy and Differential Privacy

• Anonymization Flaws: Traditional anonymization (removing names, coarsening data) is fundamentally flawed because combining datasets with external public data can re-identify individuals (e.g., the Netflix Prize re-identification).
• Differential Privacy Definition: A rigorous mathematical guarantee ensuring that the inclusion or exclusion of a single individual's data does not significantly change the output of an analysis.
• Mechanism: Differential privacy is achieved by adding calibrated noise (e.g., Gaussian or exponential) to computations, making it probabilistic rather than deterministic.
• Universality: Most standard statistical and machine learning algorithms (backpropagation, decision trees, support vector machines) can be modified to satisfy differential privacy.
• Obfuscation: Alternative privacy methods include "obfuscation" or "weed and chaff" (e.g., browser plugins sending irrelevant queries to mask user intent), though these do not offer the same mathematical guarantees as differential privacy.
• Future Economy: Future models may involve markets where users are compensated for their data, with differential privacy serving as a framework to value data based on its marginal utility.

Game Theory and Social Systems

• Algorithmic Game Theory: A 20-year-old field applying game theory to computer science, focusing on large-scale interactions and "no regret learning" where self-interested agents converge to equilibrium.
• Nash Equilibrium: Platforms like Waze and Facebook drive users toward Nash equilibria where individual optimization (e.g., shortest route, most engaging content) can lead to suboptimal collective outcomes (e.g., increased traffic, societal polarization).
• Stability: While equilibrium implies stability, it does not guarantee optimality for all participants; mechanisms exist to nudge systems toward better social outcomes.
• Financial Markets: Algorithms dominate high-frequency trading and execution optimization but struggle with long-term "Buffett-style" investing which requires understanding human nature, political cycles, and macroeconomic risks.
• Nuclear Deterrence: Game theory was successfully applied to Cold War detente, though Kearns notes modern contexts involve more players and complexities.

Social Media and Engagement Optimization

• Polarization: Social media platforms have inadvertently optimized for engagement, creating echo chambers and political polarization.
• Remedies: Engineers can tune algorithms to expose users to content they might disagree with (exploration vs. exploitation), though this may reduce short-term engagement and revenue.
• Economic Model Shift: Increasing user control over data and privacy would disrupt the current "users are the product" advertising model, necessitating new economic structures.

Personal and Philosophical Reflections

• Optimism on Human Nature: Kearns believes most people are fundamentally good, with "bad" behavior arising from group-specific social norms and professional cultures (e.g., finance, academia) rather than inherent evil.
• Watershed Moment: Kearns nearly dropped out of his PhD at Harvard but decided to persist after realizing the value of research independence and originality during a reflective day at Boston Common.
• Computer Science as a Universal Tool: The field is likely to become ambient across all disciplines, potentially marginalizing dedicated CS departments while increasing the need for data literacy in humanities and social sciences.
• Future of AI: Kearns does not believe we are closer to general AI than two decades ago; current deep learning successes are limited to narrow tasks.