Interview
Michael Kearns: Algorithmic Fairness, Privacy & Ethics | Lex Fridman Podcast #50
- The field of algorithmic fairness is anticipated to be significantly more complex and immature than algorithmic privacy, with theoretical proofs confirming the impossibility of simultaneously satisfying three reasonable fairness notions, requiring a period of approximately 10 years for the discipline to mature through psychology and human subject experiments.
- Future advancements in algorithmic fairness and privacy are expected to rely on philosophical engagement rather than purely technical solutions, particularly as the field navigates "gray areas" and attempts to explain trade-offs on Pareto curves to non-technical policymakers in the near future.
- Social media platforms are currently described as being in a bad equilibrium driven by engagement optimization that fosters polarization, though it is deemed feasible to shift metrics toward minimizing divisiveness or maximizing intellectual growth, despite anticipated short-term revenue losses for companies.
- Implementing increased individual data control is predicted to necessitate a regulatory legal process rather than voluntary algorithmic fixes, likely causing extreme economic consequences that could upend the current internet business model where users are the product.
- Machine learning is expected to remain insufficient for macroeconomic prediction over long time scales (years) due to the requirement for understanding human nature and political landscapes, ensuring that high-level long-term investment roles remain safe from algorithmic replacement for the foreseeable future.
- The academic discipline of computer science may face long-term marginalization as its tools become ambient across all other fields, resembling a utility department rather than a distinct silo.
- Optimism exists for achieving a better compromise between individual privacy control and societally beneficial data uses than the current state of weak guarantees, though the speaker warns against allowing algorithms to define fairness or explore its definitions independently due to scientific ignorance and the topic's emotional weight.
- Improvements to current systems may involve showing users content they might disagree with or that is further from their interests, utilizing the same metric space models to identify such viable alternatives to the current engagement-focused trajectory.