The Correlated Decision Problem

The Evidence Chain, Part 6 of 10. Part 5 was about systems that fail together because they share infrastructure. This part is about something more unsettling: systems that keep working perfectly and still make the same mistake, about the same people, at the same time.

At 12:55 in the morning, Derek Mobley applied for a customer-service job. At 1:50 a.m., less than an hour later, he was rejected. That detail comes from a federal court order, which noted that the speed and number of his rejections supported a plausible inference that software, not a person, was turning him down (Mobley v. Workday, Inc., 2024). Mobley, who is over forty, says he applied to more than a hundred jobs at companies that used the same hiring platform, and was rejected every time (Mobley v. Workday, Inc., 2025a).

Put yourself in his place. After a hundred rejections from a hundred different employers, the natural conclusion is that the problem is you.

A hundred independent organizations looked and agreed. But what if they did not look independently? What if the same system scored, sorted and ranked you before any of them saw your name? Then you did not get a hundred opinions. You got one opinion, a hundred times.

That is the correlated decision problem. The employers, lenders, landlords and insurers involved are organizationally independent: different boards, different policies, different staff. What they increasingly lack is independent computation. Their reading of you is coming from the same place.

A billion rejections and one question

Mobley’s case is still being fought, and nothing in it has been proved. But one number from it deserves to sit in your head. In arguing about how many people should be notified of the lawsuit, Workday itself said that 1.1 billion applications had been rejected using its systems during the period at issue; the judge’s reply was that allegedly widespread discrimination is not a reason to deny notice (Mobley v. Workday, Inc., 2025a). A billion rejections passing through a shared layer of software is not a hypothetical about monoculture. It is the infrastructure of a job market.

Researchers have now measured what that looks like from the inside. In May 2026, Rishi Bommasani and colleagues published an analysis of three million applicants and four million applications screened by a single vendor’s algorithms, and found that among applicants who applied to ten positions, 4 percent were recommended for rejection from every one of them, more than chance would predict (Bommasani et al., 2026). A separate team tested twenty large language models on 1,800 résumé-and-job pairings and found that, even when twenty firms each picked a model at random, around a fifth of applicants were still systemically excluded; exclusion was worst when every firm used the same model (Kim et al., 2025). The same study found that, across dozens of models on a standard benchmark, when two models were both wrong they agreed on the wrong answer about 60 percent of the time, against a chance rate of 33 percent, and that more accurate models had more correlated errors. And a paper published in August 2026 found that the post-training step that makes models helpful and polite made them more alike in whom they excluded, with systemic exclusion rising from about 6 percent for base models to about 17 percent after post-training, driven mainly by age (Bone et al., 2026).

Why independence was doing so much work

Much of how we trust institutional decisions rests on an unspoken assumption of independence. We trust a second medical opinion because it comes from a different doctor. We trust a market price because many buyers and sellers disagree. We trust a scientific finding because several labs replicated it. In each case, agreement means something only because the parties could have disagreed. When they share a computational reading of the world, agreement tells you much less.

The theory was worked out before the evidence arrived. Kleinberg and Raghavan showed in 2021 that when many decision-makers converge on the same algorithm, overall decision quality can fall even if that algorithm is more accurate than the alternatives for any single user, because errors stop cancelling out (Kleinberg & Raghavan, 2021). Bommasani and colleagues named the human cost a year later: outcome homogenization, the tendency for the same individuals to receive the same negative outcome from every decision-maker (Bommasani et al., 2022). Being turned down by one system is bad luck. Being turned down by every system for the same hidden reason is a kind of exclusion that no single institution can see, because each one only sees its own decision.

Who this lands on

Correlated decisions do not fall evenly. In Mobley’s case, the alleged pattern concerns applicants over forty. In the August 2026 study, age was the main driver of exclusion (Bone et al., 2026). But the generation entering the job market now is the one that will meet shared screening systems at every door. Economists at Stanford reported in a revision of their study this summer that employment of 22-to-25-year-olds in the occupations most exposed to AI is about 19 percent below where it would have been had it kept pace with less-exposed peers (Brynjolfsson et al., 2026). That study measures employment, not screening, and I am not claiming one causes the other. I am saying that the people with the least work history to argue with are the ones most likely to be read by the same few systems, again and again.

The pattern is global. In Kenya, the rapid growth of app-based digital lending meant that, by 2017, an estimated 2.7 million digital borrowers had been reported to credit reference bureaus, and digital credit accounted for 90 percent of all blacklistings (Johnen et al., 2021). A shared bureau is a shared lens. A blacklisting by one lender becomes, in practice, a decision by all of them.

The power question

Who decides here? The vendor that builds the model and the employers or lenders that buy it, each believing they made their own call. Who can intervene? Each institution can override its own recommendation, but none can see the others’ decisions, and the applicant can see none of them. Who carries the consequences? The person who keeps getting the same answer and has no way to learn that it was always the same question. In the United States, a lender that rejects you must give you specific reasons (Regulation B, 2024), but nothing requires it to tell you that four other lenders used the same scoring vendor. The one fact that would tell you whether you got five opinions or one is exactly the fact the notices leave out.

So if you are on the receiving end, here is the question worth asking, out loud and in writing: which system read my application, and who else uses it? You may not get an answer. But a pattern of identical, instant rejections from organizations that share a platform is not proof that you are the problem. It may be evidence that you met the same gatekeeper many times, and in Mobley’s case it was exactly that pattern, written down and taken to court, that made the shared system visible at all.

The prediction

My prediction is that correlated decision-making will become a recognized category of AI risk, measured and supervised the way concentration risk is today, starting in hiring and credit, where the harms are most visible and the litigation is already underway. The research question for this part is one very few organizations can answer today: how much independence actually exists between decisions that look as if they were made independently?

For Mercury Intelligence members: the section below walks the Workday litigation through the evidence chain, maps it against the EU AI Act, US state and federal rules, sets out how to measure the independence of your own decisions, and tests the argument against the strongest recent defense of monoculture.

---

Support independent investigation. Following a case like this means reading court orders that most coverage skips, and asking where the evidence of correlated exclusion actually lives. Nobody who sells screening software will fund that question. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron: fairbydesign.org/patrons-2.

References

Bommasani, R., Bana, S. H., Creel, K. A., Jurafsky, D., & Liang, P. (2026). Algorithmic monocultures in hiring (arXiv:2605.27371) [Preprint]. arXiv. https://arxiv.org/abs/2605.27371

Bommasani, R., Creel, K. A., Kumar, A., Jurafsky, D., & Liang, P. (2022). Picking on the same person: Does algorithmic monoculture lead to outcome homogenization? In Advances in Neural Information Processing Systems 35 (NeurIPS 2022). https://arxiv.org/abs/2211.13972

Bone, M., Stephany, F., & del Rio-Chanona, M. (2026). Monocultural biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring (arXiv:2609.22169) [Preprint]. arXiv. https://arxiv.org/abs/2609.22169

Brynjolfsson, E., Chandar, B., & Chen, R. (2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence [Working paper, revised August 12, 2026]. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf

California Civil Rights Department. (2025, June 30). Civil Rights Council secures approval for regulations to protect against employment discrimination related to artificial intelligence [Press release]. https://calcivilrights.ca.gov/2025/06/30/civil-rights-council-secures-approval-for-regulations-to-protect-against-employment-discrimination-related-to-artificial-intelligence/

Colorado General Assembly. (2026). SB26-189: Automated decision-making technology [Bill page; signed May 14, 2026]. https://leg.colorado.gov/bills/sb26-189

Eigen, Z., Kelly, B., Tobey, D., Carr, A., Kloeppel, K., & Mansur, J. (2026, January 30). Critical audit of NYC’s AI hiring law signals increased risk for employers. DLA Piper. https://www.dlapiper.com/en-us/insights/publications/2026/01/critical-audit-of-nyc-ai-hiring-law-signals-increased-risk-for-employers

European Commission. (2026). AI Act: Regulatory framework for artificial intelligence [Policy page, updated August 3, 2026]. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

Exec. Order No. 14365. (2025, December 11). Ensuring a national policy framework for artificial intelligence. The White House. https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/

Financial Stability Board. (2024, November 14). The financial stability implications of artificial intelligence. https://www.fsb.org/2024/11/the-financial-stability-implications-of-artificial-intelligence/

Financial Stability Board. (2025, October 10). Monitoring adoption of artificial intelligence and related vulnerabilities in the financial sector. https://www.fsb.org/uploads/P101025.pdf

Hedden, B., & Raghavan, M. (2026). Algorithmic monoculture and its critics (arXiv:2604.06047) [Preprint]. arXiv. https://arxiv.org/abs/2604.06047

Johnen, C., Parlasca, M., & Mußhoff, O. (2021). Promises and pitfalls of digital credit: Empirical evidence from Kenya. PLoS ONE, 16(7), Article e0255215. https://doi.org/10.1371/journal.pone.0255215

Kim, E., Garg, A., Peng, K., & Garg, N. (2025). Correlated errors in large language models. In Proceedings of the 42nd International Conference on Machine Learning (ICML 2025). https://arxiv.org/abs/2506.07962

Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), Article e2018340118. https://doi.org/10.1073/pnas.2018340118

Mobley v. Workday, Inc., 740 F. Supp. 3d 796 (N.D. Cal. 2024). https://storage.courtlistener.com/recap/gov.uscourts.cand.408645/gov.uscourts.cand.408645.80.0_1.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. May 16, 2025a) (order granting preliminary collective certification). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-1.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. July 29, 2025b) (order re HiredScore dispute). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-3.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. Mar. 6, 2026a) (order granting in part and denying in part motion to dismiss). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-13.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. May 29, 2026b) (discovery order). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-16.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. June 22, 2026c) (order granting in part and denying in part motion to dismiss). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-17.pdf

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal. July 2, 2026d) (order denying certification under 28 U.S.C. § 1292(b)). https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_23-cv-00770/pdf/USCOURTS-cand-3_23-cv-00770-21.pdf

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj

Regulation B (Equal Credit Opportunity Act), Notifications, 12 C.F.R. § 1002.9 (2024). https://www.consumerfinance.gov/rules-policy/regulations/1002/9/

U.S. Department of Justice, Antitrust Division. (2025, November 24). Competitive impact statement, United States et al. v. RealPage, Inc. et al., No. 1:24-cv-00710 (M.D.N.C.). https://www.justice.gov/atr/media/1419471/dl