Michael Kearns — A University of Pennsylvania computer science professor and co-author of The Ethical Algorithm. He is a world-class researcher in machine learning theory, algorithmic game theory, and quantitative finance.
The gist
Lex Fridman talks with Michael Kearns about his book The Ethical Algorithm and the project of building fairness and privacy directly into machine learning systems. Kearns explains how group-level fairness definitions can conflict, why individual fairness is far harder, and how trade-offs between accuracy and unfairness are unavoidable. He gives an accessible account of differential privacy, the failures of data anonymization, and how adding noise can protect individuals while still permitting useful analysis. The conversation broadens into game theory, social-media engagement optimization, algorithmic trading, and the growing societal responsibility of computer scientists.
Big reveals
- Kearns argues privacy is largely solved conceptually via differential privacy, but algorithmic fairness is a much messier, harder problem with provably incompatible definitions.
- Group fairness offers little comfort to a wronged individual: your 'compensation' is just knowing others are being denied at the same rate.
- He frames individual fairness as protecting all possible subgroups at once, naming the failure mode 'fairness gerrymandering.'
- Kearns insists on showing stakeholders an explicit accuracy-vs-unfairness Pareto curve rather than pretending no trade-off exists.
- He pushes back on blaming tech companies, arguing no one anticipated platforms like Facebook would warp political discourse.
- He floats giving users a 'slider' to deliberately see content they disagree with, since the same model that maximizes happiness can find what offends you.
- Just your Facebook likes can predict sexual orientation, drug use, and parents' divorce, making 'anonymized' data effectively a fingerprint.
- Navigation apps push everyone toward a Nash equilibrium that may make collective driving time worse than a coordinated solution.
Things worth remembering
- Kearns started college as an English major and once wanted to be a writer before turning to math and computer science.
- He has a moral-philosopher uncle he emailed early fairness definitions to, learning philosophers answer 'it depends' rather than yes or no.
- The Netflix Prize data was de-anonymized by joining it with public IMDb movie ratings.
- Differential privacy's smoking-cancer example: your record isn't crucial because the link is a fact discoverable from any large enough dataset.
- Nearly all of statistics and machine learning, including backprop, SVMs, and boosting, can be done in a differentially private way by adding noise.
- An NYU project built a browser plugin that hides your real Google searches in a torrent of ~999 fake queries.
- Game theory was taken seriously at the RAND Corporation in the 1960s for modeling US-Soviet nuclear strategy with two-by-two tables.
- Kearns argues Warren Buffett-style long-horizon investors are safest from automation because they need views on wars, recessions, and politics.
- Kearns comes from a deeply academic family: his father, uncle, and both grandfathers were professors.
Recommended in this episode
Books, products and media the guest or host genuinely endorsed here — with the buy link.
Affiliate link — we may earn a commission at no extra cost to you.
Infinite Jest
David Foster Wallace
“my favorite novel is Infinite Jest by David Foster Wallace which actually coincidentally much of it takes place in the halls of buildings right around us” — Michael Kearns 00:03:07Find it on Amazon
The Ethical Algorithm
Michael Kearns and Aaron Roth
“the title of the book is ethical algorithm by the way and they didn't think of that interpretation of the title” — Michael Kearns 00:45:43Find it on Amazon