Rediet Abebe

Life and career

Rediet Abebe's career answers a question most of the field never asks precisely enough: what, specifically, is computer science *for* when the problem is poverty? Not "how do we make algorithms fairer" — a defensive question about systems that already exist — but the constructive one. If you had the full apparatus of algorithmic game theory, network science, and optimization, and you pointed it at inequality on purpose, what would you build?

She was born in Addis Ababa in 1991 and educated there before coming to the United States for university. She read mathematics at Harvard, then went to Cambridge for further mathematical study, and returned to Harvard for graduate work in applied mathematics before moving to Cornell for a PhD in computer science. Her doctoral advisor was Jon Kleinberg, which places her squarely in the algorithmic-social-networks lineage — the tradition that treats social structure as a formal object with theorems attached. She completed the PhD in 2019, the first Black woman to earn a computer science doctorate at Cornell.

Two organizations she helped found while still a graduate student have arguably shaped the field as much as her papers. In 2016, with Kira Goldner, she started Mechanism Design for Social Good (MD4SG), a research initiative and workshop series that grew into a large international community working on algorithmic and economic approaches to inequality — housing, healthcare access, credit, immigration, education. In 2017, with Timnit Gebru, she co-founded Black in AI, which changed the demographics of the major machine learning conferences by the direct method of organizing, funding travel, and building an actual community rather than issuing statements.

After the PhD she was elected a Junior Fellow of the Harvard Society of Fellows — an unusual placement for a computer scientist — and in 2022 joined the University of California, Berkeley as an assistant professor of computer science, the first Black woman on the EECS faculty there. She has also worked to build research capacity on the African continent rather than only recruiting from it, an emphasis that runs through her institutional work as consistently as it runs through her papers.

Her research group at Berkeley has continued along two tracks that she keeps deliberately entangled: formal work on allocation, matching, and network dynamics; and empirical work that uses computational tools to *measure* deprivation and access in settings where the data infrastructure is thin — including studies using search-query data to characterize health information needs across African countries, which is a good example of her habit of treating a data source as a measurement instrument for a population that official statistics undercount.

Key contributions

**Mechanism design for social good.** The framing contribution, and the one that organizes everything else. Classical mechanism design asks how to set rules so that self-interested agents, acting strategically, produce an outcome the designer wants — typically efficiency or revenue. Abebe's program asks what happens when the designer's objective is *distributional*: not total welfare, but welfare for the worst-off; not throughput, but access for those currently excluded. This is not a rhetorical substitution. Changing the objective changes the mathematics — incentive compatibility interacts differently with a maximin objective than with a utilitarian one, and the price of fairness becomes a quantity you can bound rather than a slogan.

**Subsidy allocation under income shocks.** A representative technical result, with Kleinberg and Matthew Weinberg. Model an individual not by a static income but by a distribution over income realizations, and define welfare in terms of the probability of falling below a threshold — the way poverty actually works, as exposure to volatility rather than a fixed level. Then ask how a fixed subsidy budget should be distributed to minimize expected deprivation. The answer is not proportional to need, and it is not obvious; the structure of the optimal allocation depends on the shape of the shock distributions, and the paper characterizes when the problem is tractable and when it is hard. The methodological point for a graduate audience: the *definition of welfare you write down* determines the algorithm you get, and most default definitions quietly encode a utilitarian choice nobody argued for.

**Roles for computing in social change.** With Solon Barocas, Kleinberg, Karen Levy, Manish Raghavan, and David Robinson, a paper that is now standard reading in the area. Its contribution is a taxonomy of what computational work can legitimately do about a social problem even when it cannot solve it: computing as a *diagnostic* that measures a harm precisely enough to make it undeniable; as a *formalizer* that forces vague policy commitments into specifications where their contradictions become visible; as a *rebuttal* that establishes the limits of what technical intervention can achieve, thereby shifting the argument back to politics; and as a *synecdoche*, where an algorithmic controversy makes a long-standing structural problem newly legible to the public. This is a genuinely useful epistemology for anyone deciding whether a technical project is worth doing.

**Network dynamics and diffusion.** Her more classically algorithmic work includes models of how opinions spread when individuals differ in their susceptibility to persuasion — and how an intervention budget spent on changing susceptibility compares to one spent on seeding — plus contagion and cascade models on social networks. This is Kleinberg-school network science, and it is a large part of why her range in the game extends well beyond ethics problems.

**Computational measurement of deprivation.** The empirical strand: using non-traditional data sources to characterize needs in populations that conventional instruments miss, with careful attention to what the data source can and cannot support.

In battle

Abebe has the highest mean of the modern ethics-adjacent carriers — 34.7 across 100 problems, with nine dominant cells (80+) and twelve strong ones. That is a meaningfully broader hand than Gebru's, and the reason is structural: Abebe carries two toolkits, not one. She has the fairness-and-accountability agenda, and she has algorithmic game theory and network science underneath it.

That second toolkit is where her single highest category sits: **games** at 74.2, her best column by a wide margin. **P055 — Design the auction** (84) and **P206 — The exchange with no prices** (85) are mechanism design and matching-market problems — auction design and barter exchange without money, the kidney-exchange family — and she scores on them as a working theorist, not as a critic. **Networks** at 45.1 over sixteen problems is her deepest bench, anchored by **P092 — Contagion on the network** (88), where diffusion and cascade modeling is her actual dissertation-adjacent territory.

The fairness column (63.0 over fifteen problems) produces her top cell: **P297 — The ad the algorithm never showed you** at 90 — an optimizer maximizing clicks per dollar reconstructing the very targeting civil-rights law forbids. The matrix's explanation is precise about why she owns it: her framework specifies what the system should optimize *instead*, and at what measured utility cost, which is the constructive half that a pure critique cannot supply. **P299 — The proxy that rationed care** (88) is the healthcare-cost-as-health-need failure, a canonical case in her literature. **P289 — The variable you removed is still there** (85) is proxy discrimination through correlated features. **P265 — A recognizer for a language of ten speakers** (85) and **P266 — What is in the training data?** (81) reflect her long insistence on underserved populations and thin data infrastructures.

The losses are clean and they are all mathematical rather than era-based. She scores 3 on **P211 — Roll the dice at Los Alamos** and 3 on **P216 — Calculus for a jagged path**, 6 on **P212 — Sample from the impossible posterior** and 6 on **P213 — The posterior at web scale**, 6 on **P122 — The lady and her teacups**, and 8 on **P217 — How high must the dike be?**. Monte Carlo variance reduction, Itô calculus, MCMC, scalable Bayesian inference, exact permutation testing, extreme-value theory. There is no bridge from computational social choice to a stochastic differential equation, and the matrix does not build one.

Two subtler numbers repay attention. Her **testing** average is 14.2 across sixteen problems — a large, uniformly weak column. Her work is allocative and structural; classical hypothesis testing and its machinery are simply not her instrument, and the game makes her pay for that sixteen times. And her **causality** score of 22.2 over eighteen problems is, like Gebru's, lower than a student would guess for a researcher who studies discrimination. The reason is the same: the game's causality problems are about *identification* — instruments, counterfactuals, do-calculus — and Abebe's method is design and allocation. She can tell you what a system should be optimizing and prove properties about the resulting allocation; she is not the person you send to disentangle a confounded observational estimate.

Practical read: Abebe is the most versatile of the modern ethics carriers because she can win on pure algorithmic-economics problems that have no fairness framing at all. Play her on allocation, matching, auctions, network diffusion, and any problem where the decisive question is *who benefits*. Keep her away from stochastic analysis, sampling, and classical inference.