Algorithmic Ethics, Fairness, and Privacy with Michael Kearns
The Intersection of Algorithms and Society
In this episode, Michael Kearns joins to discuss his book, The Ethical Algorithm. The conversation explores how technical constraints in machine learning can be leveraged to encode complex human values such as fairness, privacy, and algorithmic ethics directly into computational systems.
The Challenge of Algorithmic Fairness
Kearns emphasizes that we are in the early days of defining what it means for an algorithm to be "fair." He highlights several critical points regarding this field:
• Group vs. Individual Fairness: Many current definitions of fairness operate at the group level, which can miss nuances and create "fairness gerrymandering" where models are fair on average but discriminate against intersecting sub-groups.
• The Necessity of Trade-offs: Achieving fairness often requires sacrificing some level of accuracy. Kearns advocates for using Pareto curves to visualize the trade-off between error rates and unfairness, allowing stakeholders to make informed decisions rather than ignoring these inherent costs.
• Subjective Nature: Because fairness involves ethical and philosophical dilemmas, computer scientists must engage with the broader human context rather than treating metrics as the sole definition of justice.
Data Privacy and Differential Privacy
Addressing the failures of traditional anonymization, Kearns explains the power of differential privacy:
"Differential privacy basically says that any harms that might come to you from the analysis in which your data was included are essentially nearly identical to the harms that would have come to you if the same analysis had been done without your medical record included."
• By adding carefully calibrated noise to computations, systems can extract aggregate insights without risking the exposure of specific individual data points.
• This approach transforms privacy from a fuzzy concept into a quantifiable guarantee.
Game Theory and Market Interactions
Kearns bridges the gap between machine learning and game theory, exploring how algorithms influence the collective behavior of large populations:
• Nash Equilibria: He notes that modern platforms essentially nudge users toward competitive equilibria, which are not always optimal for the individuals or society at large.
• Optimization Limits: Simply optimizing for engagement has led to polarization. Future algorithmic design might require "exploration" triggers that expose users to diverse viewpoints to counteract echo chambers.