Lex Fridman PodcastMichael Kearns: Algorithmic Fairness, Privacy & Ethics | Lex Fridman Podcast #50
CHAPTERS
- 0:00 – 1:00
Lex’s intro: Kearns’ background and why ethical algorithms matter
Lex introduces Michael Kearns and frames the conversation around the book Ethical Algorithm, focusing on fairness, privacy, and ethics. He also shares a personal connection from earlier academic collaboration and sets expectations for touching multiple research areas.
- •Michael Kearns’ roles: Penn professor, researcher, co-author of Ethical Algorithm
- •Core topics preview: algorithmic fairness, bias, privacy, ethics
- •Lex’s personal history with Kearns and algorithmic game theory
- •High-level scope beyond ethics: learning theory, game theory, finance, computational social science
- 1:00 – 2:31
Sponsor segment: fear of new technology and recurring social reactions
Lex describes the podcast Pessimists Archive and its theme: society repeatedly panics about new technologies that later become normal. The segment foreshadows how algorithmic systems provoke similar cultural fear and resistance today.
- •Pessimists Archive as a history of technology backlash
- •Examples of once-feared innovations (mirrors, bicycles, cars, etc.)
- •Recurring patterns of human psychology around new tools
- •Connection to modern tech anxiety and social media dynamics
- 2:31 – 3:40
Literary influences and the path from English to computer science
Kearns discusses formative non-technical reading and his early intention to be a writer. He explains how he started college as an English major before shifting toward math and computer science.
- •Infinite Jest as a major literary influence
- •Early attraction to journalism/literature in high school
- •Realization about what English curricula do (reading vs. writing)
- •Transition toward math/CS and eventual research career
- 3:40 – 7:33
From moral philosophy to implementable definitions of fairness
Lex and Kearns explore how technical work on fairness intersects with long-standing philosophical debates. Kearns notes the gap between philosophical nuance (“it depends”) and the need for precise, implementable definitions in algorithms.
- •Technical motivation: algorithms’ “antisocial behavior” and research opportunity
- •Consulting a moral philosopher (Kearns’ uncle) on Rawlsian-style notions
- •Philosophy vs. CS: ambiguity vs. operational definitions
- •Fairness as early, messy, and definition-dependent compared to privacy
- 7:33 – 13:14
Are people fundamentally good? Power, professional culture, and social norms
A philosophical detour: Kearns argues most people—including those in power—have good intentions, but environments shape behavior. He emphasizes how insular professional norms (finance, academia) can create outcomes misaligned with broader societal values.
- •Optimistic view: most people want to do right; circumstance drives failures
- •Power doesn’t imply moral corruption; culture and norms matter more
- •Finance and academia as examples of insider logic that outsiders find baffling
- •“Slippery slope” through normalization rather than explicit malicious intent
- 13:14 – 19:04
What is an “ethical algorithm”? Quantifying ethics and choosing what to protect
Kearns reframes ethical algorithms as ones optimizing measurable properties—once society selects specific definitions (e.g., a fairness metric). The discussion highlights that ethics becomes quantitative only after difficult human choices about protected groups and what counts as harm.
- •Analogy to algorithm trade-offs (time, memory) and “better vs. worse” algorithms
- •Ethical Algorithm focuses on specific, formalizable norms—not fuzzy morality
- •Example metric: disparity in false rejection rates across groups
- •Human choices required: which groups to protect, and what harms matter
- 19:04 – 25:43
Group fairness vs. individual fairness—and “subjective fairness”
They examine limitations of group-based fairness metrics and why they may not feel fair to individuals. Kearns introduces the notion of subjective fairness and argues the field needs more human-subject research to learn what people actually perceive as fair and understandable.
- •Group fairness can still leave individuals feeling wronged
- •Tension between group guarantees and individual-level promises
- •“Subjective fairness”: eliciting fairness judgments from ordinary people
- •Parallel critique in interpretability: researchers assume what humans understand without testing
- 25:43 – 33:36
Fairness gerrymandering and the combinatorial explosion of protected subgroups
Kearns explains how protecting only broad categories (race, gender, age) can still allow discrimination against intersections of attributes. He describes “fairness gerrymandering” and outlines algorithmic approaches that audit and mitigate discrimination across many subgroups.
- •Individuals can be viewed as intersections of many attributes
- •Marginal protections don’t ensure intersectional protections
- •Term: fairness gerrymandering (analogy to political gerrymandering)
- •Algorithmic auditing to find worst-off subgroups and improve guarantees
- 33:36 – 44:22
Fairness–accuracy trade-offs: Pareto frontiers and stakeholder decision-making
Kearns argues fairness should be treated like other engineering trade-offs: measurable and transparent. He proposes Pareto curves as an interface for policymakers and stakeholders to choose explicit operating points rather than making hidden, implicit trade-offs.
- •Auditing models for disparate error rates is straightforward once metrics are chosen
- •Fairness constraints typically trade off with prediction error (and/or other costs)
- •Pareto frontier: mapping error vs. unfairness and tuning enforcement strength
- •Decision authority should rest with stakeholders/regulators, not engineers alone
- 44:22 – 1:06:00
Divisive culture, social media algorithms, and escaping “bad equilibria”
They discuss how engagement-optimized personalization can polarize discourse and create unhealthy societal equilibria. Kearns suggests algorithmic and product interventions—like exploration sliders—to intentionally diversify content exposure, while acknowledging revenue trade-offs and measurement challenges.
- •Engineers didn’t fully anticipate platforms’ political and social impact
- •Platforms are social systems: user behavior co-evolves with algorithms
- •Idea: tunable “exploration” to show content further from a user’s preferences
- •Engagement optimization can create a bad equilibrium; changing objectives may cost revenue
- 1:06:00 – 1:22:32
Privacy basics: why anonymization fails and how differential privacy works
Kearns contrasts common anonymization/redaction with differential privacy, explaining why re-identification is easy when datasets can be joined. He then provides the core differential privacy idea—comparing worlds with and without one person’s data—and the main mechanism: carefully calibrated noise.
- •Anonymization: removing identifiers, coarsening demographics, “PII removal”
- •Failure mode: linkage attacks (Netflix Prize + IMDb; fingerprinting via sparse attributes)
- •Differential privacy guarantee: outcomes nearly unchanged with/without your record
- •Mechanism: probabilistic algorithms that add noise to computations (e.g., noisy averages)
- 1:22:32 – 1:27:49
The future of privacy: user control, regulation, and markets for data
Kearns is optimistic about balancing beneficial data use with stronger privacy and user control, but stresses it will require policy and regulation. They explore how giving users real control could disrupt today’s ad-driven internet economy and motivate new market structures where individuals can participate and be compensated.
- •Most ML/statistics can be made differentially private in practice
- •Obfuscation ideas (e.g., noisy search queries) and when personalization demands specificity
- •Need for regulatory/legal pressure to drive adoption of privacy norms
- •Data as product: shifting power to users could upend ad-based business models; potential for user-participatory data markets
- 1:27:49 – 1:36:55
Game theory, machine learning, and platform equilibria (traffic, feeds, markets)
Kearns defines game theory as a framework for collective outcomes in interacting systems, then connects it to modern ML-driven platforms. Examples like navigation apps show how personalized optimization can push society toward equilibria that may be stable yet inefficient or undesirable.
- •Game theory: predicting outcomes of interacting agents; classic prisoner’s dilemma framing
- •Key beauty: Nash equilibrium existence; stability is possible but not guaranteed by dynamics
- •ML link: personalization and optimization act like best-response dynamics in large games
- •Navigation apps may increase total travel time vs. coordinated alternatives; equilibrium can be worse than feasible improvements
- 1:36:55 – 1:44:08
Algorithms in finance: where bots dominate and where humans still matter
Kearns outlines how trading became algorithmic as exchanges went electronic, with early dominance in execution and high-frequency strategies. He argues longer-horizon investing remains harder to automate because it requires integrating heterogeneous data and forming world models about politics, cycles, and rare events.
- •Automation enabled algorithmic execution via APIs and electronic exchanges
- •Optimized execution: splitting large institutional orders across time/venues to reduce market impact
- •Stat arb and short-horizon directional prediction as a quant “sweet spot”
- •Long-horizon investing needs broader context (macro cycles, wars, politics); “robo Buffett” not soon
- 1:44:08 – 1:48:55
Closing reflections: a grad-school turning point and sticking with research
In the final personal segment, Kearns recalls an unhappy first year of grad school and a decisive moment in Boston Common when he chose to persist. He describes the transformation from coursework to genuine research independence as the watershed that shaped his career.
- •Feeling narrowed and isolated in early grad school compared to undergrad breadth
- •Considering leaving with a master’s and returning to industry
- •Decision to “stick it out” until research began to click
- •Doctoral transition: from learning past work to producing original ideas and autonomy
