Kleinberg and Raghavan proved algorithmic monoculture can hurt social welfare

What Kleinberg and Raghavan proved

Jon Kleinberg of Cornell and Manish Raghavan, then at Cornell and now at MIT, published "Algorithmic monoculture and social welfare" in the Proceedings of the National Academy of Sciences in early 2021.¹ Two firms compete to hire from a pool of candidates whose intrinsic values they cannot observe directly. Each firm ranks candidates through a noisy mechanism, then hires in random order: the first firm takes the top of its own ranking, the second takes the top of what remains in its own.

Under polyculture, each firm draws an independent ranking at accuracy \theta_H. Under monoculture, both firms use the same ranking from a more accurate algorithm at \theta_A > \theta_H. Social welfare is the sum of the intrinsic values of the two hires. Theorem 1 assumes two conditions on the candidate distribution and noise family. Under them, for any \theta_H there exists a \theta_A > \theta_H at which using the algorithm is strictly dominant for both firms, yet welfare would have been higher had both used the humans.¹

The mechanism inside the proof

The engine of the theorem is what Kleinberg and Raghavan call preference for independence.¹ Consider the firm that hires second. If the first firm used an independent ranker, the two rankers disagreed about the order, so the second firm's top-remaining candidate is close to its own ranker's top pick. If both firms use the same ranker, the second is left with the shared second-ranked candidate, systematically less valuable than an independent top-remaining draw would have been.

The paper gives a worked instance. Three candidates are drawn uniformly on [0,1] with Gaussian noise. At equilibrium both firms use the algorithm, and expected social welfare is "approximately 4% less than it would be had both firms used human evaluators instead."¹ The 4% is a stylized instance, not a general bound.

The noise family matters. Theorem 3 shows the argument's two conditions hold under the Mallows model for any candidate distribution. Theorem 2 shows they hold under random-utility models with Gaussian or Laplacian noise for three candidates. Section 3.1 then shows the effect vanishes under the Plackett-Luce model, another standard random-utility model: the optimal strategy is always to use the best available ranking regardless of what competitors do. A Laplacian case with fifteen candidates violates Definition 2. Generalization beyond three candidates for random-utility models is flagged as open.¹

What deployed measurements actually catch

A separate program measures, in real deployed systems, whether models agree on failure more than independence would predict. Rishi Bommasani, Kathleen Creel, Ananya Kumar, Dan Jurafsky, and Percy Liang defined the quantity.² Take the observed rate at which all models fail on the same subject, and divide it by the rate expected under independent per-model failures at the measured marginals. A ratio above one means more joint failures than independence predicts. On the German Credit, ACS PUMS, and LSAC fairness datasets, training the models on the same data reliably drove that ratio above one.

Connor Toups, Bommasani, Creel, Sarah Bana, Jurafsky, and Liang applied the metric to nine commercial APIs from Face++, Microsoft, Google, IBM, Amazon, and Baidu across eleven audit datasets in the 2020–2022 HAPI corpus. The observed rate of all-model failure exceeded the independence baseline on every one of the eleven datasets.³

On DIGIT the joint failure rate is nearly triple the worst individual model's rate; under independent failures the expected joint rate would have been on the order of two in a hundred thousand. Year-over-year improvements barely reach the subjects the ecosystem systemically fails. Amazon's sentiment API cut its error on the WAIMAI dataset by 2.5% between 2020 and 2021, and "precisely 0 out of the model's 303 improvements are on instances on which all other models had failed."³ Across the eleven datasets, about 10% of a system's gains landed on inputs every other model also failed on.

Elliot Kim, Avi Garg, Kenny Peng, and Nikhil Garg measured error correlation across 349 LLMs on the HuggingFace Open LLM Leaderboard and seventy-one on HELM.⁴ Their mean pairwise agreement rate on wrong answers was 0.423 on HuggingFace and 0.6 on HELM, against random-choice baselines of 0.127 and 0.33. In a resume-screening simulation where each of twenty firms picked a different LLM to score sixty candidates for thirty job openings, "around 20 percent of applicants continue to be systemically excluded."⁴

The pymetrics data shows partial monoculture

The strongest deployment evidence is Bommasani, Bana, Creel, Jurafsky, and Liang's 2026 FAccT paper.⁵ It analyzes real hiring outcomes mediated by the pymetrics assessment vendor from December 2018 through December 2022. The authors call it the first study to observe real algorithmic hiring outcomes for the same applicant across many employers.

Pymetrics asks applicants to play twelve or sixteen assessment games scoring cognitive traits such as processing speed, risk propensity, and planning ability, then returns a "recommended" or "not recommended" label per job. On average 41.8% of applications receive "not recommended."⁵ Of applicants who applied to at least four positions, 10% were recommended for rejection from every one; among those who applied to at least ten, 4% were. A chi-square goodness-of-fit test against a Poisson-Binomial independence baseline rejects independence at χ² = 18,481, p < 0.001. Run on Kline et al.'s 2022 correspondence audit of about 83,000 fictitious resumes sent to 108 Fortune 500 firms, the same test fits the baseline (χ² = 20.05, p = 0.69).⁵

The Equal Employment Opportunity Commission's four-fifths rule surfaces adverse impact only position by position. Aggregated across all 156 employers, the impact ratios for Asian, Black, Hispanic, and White applicants each exceed 0.8, the threshold at which prior pymetrics-authored analyses had reported no discrimination. Disaggregated to the 1,746 positions, 10.62% demonstrate adverse impact against Black applicants after Benjamini-Hochberg correction, and 30.70% of Black applicants applied to at least one such position.⁵ Computer and Mathematical positions carry the highest Black-adverse-impact rate by occupational group, at 16.7%.

Only 42 pymetrics models are shared across positions at different companies, and only 142 unique employer pairs share at least one model. Strict algorithmic monoculture, in which the same scorer is applied at every employer, is rare in the data. The measured excess correlation is produced by partial sharing: a common vendor, common feature construction, and a common training pipeline delivered through different fitted models. The authors report that most applicants are not scored by the same model twice, and that the harm operates through this more general form.⁵

Two illustrations often named in coverage do not extend this evidence. Amazon's internal résumé-screening system, built starting in 2014 and scrapped in 2017, was a single-firm training-data-bias case. The tool never scored a candidate at any other firm.⁶ The 2024 CrowdStrike outage was a shared kernel-driver update that crashed about 8.5 million Windows devices, with no machine-learning model involved.⁷ Both are correlated-failure analogies. Neither documents correlated hiring outcomes at scale.

Two follow-ups from inside the same program

Kenny Peng and Nikhil Garg of Cornell Tech extended Kleinberg-Raghavan into a matching-markets model with many firms rather than two (NeurIPS 2024).⁸ As the number of firms grows, firm welfare converges to first-best under polyculture and stays suboptimal under monoculture. Applicant welfare runs the other direction: for every applicant, the probability of matching to the top-choice firm is at least as high under monoculture as polyculture. In simulations on the 2018 Texas ACSIncome data, polyculture beats every monoculture on accuracy for 47 of 50 model subsets. Monoculture produces better average applicant outcomes on 50 of 50.⁸

The reversal is not total: the same theorem identifies a set of applicants whose outcome variance rises, more likely to match the top choice and less likely to match at all. But the naive reading that "being rejected once under monoculture is being rejected everywhere" does not survive the model.

Nathanael Jo, Nikhil Garg, and Manish Raghavan argued in a 2026 preprint that measured "excess" cross-model correlation depends on the null model the analyst chose.⁹ Under a rich enough null, all pairwise cross-model correlation can be re-explained by shared latent parameters. The authors reanalyze the HELM benchmark and the HuggingFace leaderboard with an item-response-theory null that accounts for question difficulty. Residual correlations reported by Kim et al. (2025) and Goel et al. (2025) partly vanish, in the paper's phrasing "even flipping from strongly positive to slightly negative."⁹ Raghavan co-authored the original PNAS paper and this critique, so the correction comes from inside the program.