fairness
The ad the algorithm never showed you
It is 2019, and civil-rights litigation has forced a reckoning: on a major platform, housing and job ads reached audiences skewed by race and gender even when advertisers targeted everyone equally — because the delivery optimizer, maximizing clicks per dollar, learned who engages and quietly reintroduced the very targeting the settlement forbade. Diagnose the mechanism: how objective functions over engagement recreate disparate delivery without any protected attribute in the targeting interface, how auction competition prices some audiences out of reach, and what a compliant delivery system would optimize instead — with the utility cost measured, not waved away. Get it wrong and fair-housing law is quietly repealed by a relevance model: opportunity itself distributed by an optimizer nobody audits, one impression at a time.
Who this problem belongs to
The two figures whose methods fit it best, out of 30 in contention.
Kleinberg's algorithmic-fairness research is built almost exactly for this scenario: his work formalizing how optimization for a proxy objective, here engagement or click-through, can reintroduce disparate outcomes even when the targeting interface never mentions a protected attribute, gives rigorous shape to the mechanism this problem asks the solver to diagnose. His broader analysis of how relevance and ranking systems interact with auction dynamics to price certain audiences out of reach directly addresses the delivery-optimizer half of the failure, and his impossibility-theorem instincts are well suited to specifying what a compliant delivery system would need to optimize instead, with the utility cost stated honestly rather than waved away. No other carrier built this exact formal apparatus.
Abebe's research program on algorithms and inequality, and her work on mechanism design for social good, is centered precisely on cases like this one: an optimizer maximizing a seemingly neutral objective, clicks per dollar, systematically reproducing discriminatory outcomes the advertiser explicitly tried to avoid, because the system learned who engages and reintroduced the very targeting civil-rights law forbids. Her framework for specifying what an algorithm should optimize instead, and at what measured utility cost, speaks directly to the compliant-delivery-system design this problem demands. Her contributions are recent and squarely in this exact subfield, giving her among the strongest possible claims on this scenario alongside Kleinberg. The throughline from the actual historical record to this exact failure mode is unusually direct for a carrier on this particular list.
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
30 figures are scored on this problem. Draw it in a battle to see where you land.