Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?

Filed correctly, and the id is a year out. proposals.md lists this under idea and the kind holds; what does not hold is the impression the id gives, which is that "algorithmic monoculture" is a 2022 idea. It is not. The term, the first formal model and the result everyone actually cites are Jon Kleinberg and Manish Raghavan, 2021. The 2022 paper named in the id is the second paper in the line, it says so on its own second page, and it contributes something different enough that the confusion matters. Section 1 keeps the two apart before anything else, because a reading that cites "the 2022 monoculture paper" for the welfare theorem has cited the wrong document.

On the kind. The test this canon has been applying — an idea is something you can implement, because there is a loop or a formalism in it — is passed twice over. Kleinberg and Raghavan supply a game-theoretic model with a proved theorem; Bommasani et al. supply two computable metrics and released the code that computes them. Both rivals are worth naming and declining:

• limit is the near miss, and it is tempting because Theorem 1 has the shape of an impossibility result: for any human accuracy, there exists an algorithmic accuracy at which every firm rationally adopts the algorithm and society is worse off for it. Decline it. A limit in this canon is godel-incompleteness-1931 or turing-halting-1936 — a statement that something cannot be done at all. This theorem is conditional (two named conditions must hold), it is about equilibrium welfare rather than computability, and as of April 2026 the loss it describes is bounded by a factor of two, which is the opposite of a barrier. Section 3 carries that number, and it is the single most under-cited fact about this idea.
• interpretation would fit the word, which is a metaphor borrowed from agriculture by way of computer security, and which does a great deal of unearned argumentative work. But the entry's core is machinery, not a reading of events, so the metaphor goes in section 4 where it can be checked rather than in the header where it would be endorsed.

On descends_from, which is thin and honestly so. Only one id is listed and it is an identification, not a citation.

• transformer-2017 is an identification, and a solid one. The paper's second half exists because of the foundation-model regime — its own words are "given the current regime in AI of foundation models" — and the systems it experiments on are CLIP with a ViT backbone and RoBERTa. All three are transformers, and the regime whose risk the paper is testing is that architecture's consequence. Vaswani et al. does not appear in the four pages of references I read, and the list may continue past them; what the paper cites are the descendants. The link is real and the route is indirect, and both facts belong on the record.

Three real ancestors have no id and cannot be listed, which is more unlisted parentage than an entry ought to carry.

• Kleinberg and Raghavan, "Algorithmic monoculture and social welfare," PNAS 118(22), 2021 is the direct parent — the source of the term as used here, and of the theorem. It deserves its own id and does not have one. A session that adds it should note that the 2022 paper's footnote 4 already states the relationship precisely, and that statement is quoted in section 1.
• Bommasani et al., "On the Opportunities and Risks of Foundation Models," 2021 — the Stanford CRFM report that coined "foundation model" and that named homogenization as the paradigm's central double edge. All five of this paper's authors are among its hundred-odd, Bommasani first on both. It is cited three times here and is the reason section 5 of the paper exists at all. No id, and it is a stronger candidate than most of what is on the proposals list.
• **Creel and Hellman, "The Algorithmic Leviathan: Arbitrariness, Fairness, and Opportunity in Algorithmic Decision-Making Systems," Canadian Journal of Philosophy 52(1), 2022** — the philosophical statement of why being rejected by everyone is a different wrong from being rejected by someone. Creel is a co-author of both.

Ancestors considered and declined. goodharts-law-1975 is thematically adjacent and mechanically unrelated: Goodhart is one measure corrupted by being optimised against; this is many decision-makers agreeing. Confusing them is common enough that section 4 separates them. frey-osborne-2013 shares the labour-market subject and nothing else — it estimates whether a job can be automated, while this asks what happens to a person once the screening is. Neither link would survive being asked what it descends from.

---

What it is

The word, and where it came from

Monoculture is a farming term: plant one cultivar across a whole region and the whole region shares one susceptibility. Computer security borrowed it — most famously in "CyberInsecurity: The Cost of Monopoly", the 2003 report by Dan Geer, Rebecca Bace, Peter Gutmann, Perry Metzger, Charles Pfleeger, John Quarterman and Bruce Schneier for the Computer and Communications Industry Association, arguing that near-universal Microsoft desktop deployment made one vulnerability into everyone's vulnerability. Geer was fired from the consultancy @stake the day it was published, which is worth recording as a fact about how the argument was received rather than as evidence it was right.

Both borrowings carry the same picture: a shock arrives, and the shock hits everything at once because everything is the same. Hold onto that, because the 2021 result's whole contribution is that the picture is not needed.

Kleinberg and Raghavan, 2021 — the theorem

The setup is deliberately small. Two firms hire one candidate each from a pool of n candidates with true, unknown values x₁ > x₂ > … > xₙ. Each firm must rank the pool using a noisy procedure, and has two options: its own private "human evaluator" ranking, with accuracy θ_H, drawn independently of the other firm's; or the algorithmic ranking, with accuracy θ_A, which is the identical permutation for every firm that chooses it — one vendor, one ordering. The firms then hire in random order, each taking the top remaining candidate on the ranking it chose. Social welfare is the sum of the values of the people actually hired.

Two conditions carry the result. Preference for the first position: if two independent rankings disagree about who is best, the top of your own ranking still beats the second on it, in expectation. Preference for weaker competition: you would rather the firm that picks before you picked using a worse ranking, because it is then more likely to have taken someone you did not want.

Theorem 1. Suppose that a given candidate distribution D and noisy permutation family F_θ satisfy Definition 2 (preference for the first position) and Definition 3 (preference for weaker competition). Then, for any θ_H, there exists θ_A > θ_H such that using the algorithmic ranking is a strictly dominant strategy for both firms, but social welfare would be higher if both firms used human evaluators.

That is the whole thing, and the mechanism is worth stating without the machinery, because it is not intuitive and it is what gets lost in citation. Under monoculture the firm that picks second gets the shared list's number two, always. Under independent evaluation it sometimes still gets its own number one, because the firm that picked first was reading a different list and took someone else. Sharing a better ranking buys each firm a better first choice and costs the loser of the coin flip its whole second draw, and for a sufficiently accurate algorithm the second effect is the larger one. The authors name the analogy themselves: "the introduction of a more accurate algorithm can drive the firms into a unique equilibrium that is worse for society than the one that was present before the algorithm existed" — a Braess' paradox for screening, after the 1968 result that adding a road can slow every driver.

The paper is careful about what it has shown, and the care is routinely stripped off in citation. From the conclusion: "Our results also make clear that algorithmic monoculture in decision-making doesn't always lead to adverse outcomes; rather, we given [sic] natural conditions under which such outcomes become possible, and show that these conditions hold in a wide range of standard models." And, on the shock picture: the effects "can persist at low levels even without a shock-like disruption to call our attention to them" — which is the real claim, and it is quieter than the agricultural metaphor, not louder.

The paper explicitly leaves the magnitude open: "it would be interesting to give more comprehensive quantitative bounds on the magnitude of monoculture's possible negative effects." It proves the effect exists and does not say how big it is. Section 3 gives the answer that arrived five years later.

Bommasani, Creel, Kumar, Jurafsky and Liang, 2022 — the measure

This is the paper the id names, presented at NeurIPS 2022 and posted 25 November 2022. It does something different, and its own footnote 4 says exactly how different: "The formal model of Kleinberg and Raghavan [2021] is related, but substantially distinct. Concretely, their formalism considers harms experienced by decision-makers, whereas we center decision-subjects."

That sentence is the reason both papers exist. 2021 asks whether the firms end up with worse hires. 2022 asks whether the same people end up rejected by everyone. They are different harms with different victims, and a reading that cites one for the other has swapped the subject of the sentence.

Its abstract, as printed at NeurIPS:

"As the scope of machine learning broadens, we observe a recurring theme of algorithmic monoculture: the same systems, or systems that share components (e.g. datasets, models), are deployed by multiple decision-makers. While sharing offers advantages like amortizing effort, it also has risks. We introduce and formalize one such risk, outcome homogenization: the extent to which particular individuals or groups experience the same outcomes across different deployments. If the same individuals or groups exclusively experience undesirable outcomes, this may institutionalize systemic exclusion and reinscribe social hierarchy."

Three moving parts:

1. A widened definition of monoculture. Kleinberg and Raghavan's monoculture is decision-makers running the same system. This paper extends it to decision-makers running similar systems in virtue of shared construction — the same training data, the same pretrained model. That is the component sharing hypothesis, stated in the paper's italics: "If deployed algorithmic systems share components, outcome homogenization will increase (i.e. there will be more systemic failures)."

2. A measure. An individual suffers a systemic failure when every deployed system fails them. The raw rate of that is not comparable across systems of different accuracy, so the metric normalises: H^individual is the observed rate of systemic failure divided by the product of the individual models' failure rates — that is, divided by the rate you would see if the systems failed independently. A value of 1 means observed matches independent; above 1 means the same people are being picked on. There is a group-level analogue with three weightings (average, uniform, worst).

Note what the denominator is: the whole quantity is defined against an assumption of independence, and section 4 is about what that costs.

3. Experiments, which came out half against the hypothesis. On tabular fairness benchmarks — German Credit, ACS PUMS census data covering 3.6 million individuals, and LSAC in the appendix — models trained on the exact same data were reliably more homogeneous than models trained on disjoint samples from the same distribution. Data sharing homogenises. But: "this relationship is not perfect: we do see several instances where the degree of homogeneity is similar or even sometimes greater for the disjoint setting."

The foundation-model half went the other way. Predicting that sharing more of a model would homogenise more — linear probing shares most, then BitFit, then finetuning, then training from scratch shares nothing — they found in vision that training from scratch was the most homogeneous of all: "This is the opposite of what we hypothesized." Their conjecture is that CLIP pretraining regularises away the idiosyncrasies of the small downstream dataset that scratch models otherwise memorise. In language the ordering was partly right and the mechanism was still wrong: BitFit and full finetuning homogenised about equally, "even though BitFit updates 0.08% of the parameters full finetuning does… suggesting the number of shared parameters is not the right lens for understanding model sharing." Their own summary: "sharing leads to homogeneous outcomes, but… it is incomplete explanation of homogeneity."

A fourth finding, and the most immediately usable one. Individual-level homogenization was consistently and substantially higher than group-level, with group values sitting near 1 — near independence. Systemic exclusion was happening to particular people while the racial-group aggregates looked unremarkable. The paper draws the consequence itself: work on algorithmic fairness that measures only social groups "may miss systemic failures for particular individuals that are obscured at the group-level… Even intersectional approaches may not suffice to surface these systemic failures, unless each intersectional group comprises a single individual."

The 2026 measurement, which is why this entry is worth writing now

The 2022 paper ran on benchmarks and said so, closing §5.3 with the exact question it could not answer: "This motivates investigation into real deployments of machine learning: for example, does sharing/monoculture lead to homogenization in algorithmic hiring?" The same group went and looked, and "Algorithmic Monocultures in Hiring" (Bommasani, Bana, Creel, Jurafsky and Liang, FAccT 2026, Montreal, 25–28 June 2026) is, in its own claim, "the first study observing deployed algorithmic hiring decisions across multiple employers from a single vendor."

The data is from the talent platform pymetrics, which builds sixteen online games measuring cognitive traits, has each applicant play twelve or sixteen of them, trains a per-client binary classifier on the gameplay features of at least fifty of that client's current employees, and returns "recommend" or "do not recommend": 4,197,168 applications from 3,372,132 applicants to 1,746 positions at 156 employers, across eleven industries with $225bn of cumulative annual revenue, December 2018 to December 2022, screened by 42 models. 41.8% of applications are "not recommended," and the paper treats that as rejection because a not-recommended applicant is unlikely to be seen by a human. The data-use agreement is stated: pymetrics "is unable to edit, restrict, or veto the research."

The mechanism is visible in the plumbing, which is why this dataset is the right one to have found. Twelve of the sixteen games are identical across every pymetrics position, and an applicant's stored features are reused for any further pymetrics-mediated application within 330 days. So the same measurements of the same person are fed to model after model. The paper's flat statement of the consequence: "42 pymetrics models screen applicants at multiple companies, which means a rejection at one company mechanically entails rejection from another company using the same model."

The findings, in the order that matters:

• Systemic rejection above chance, and not only from the mechanical route. "Of applicants that apply to ten positions, 4% are rejected from all positions." The rate falls exponentially with the number of applications, as independence would also predict, but it "decays more slowly than would be expected by chance" — and, crucially, "Even when applicants are evaluated by different pymetrics models, empirically some are systemically rejected." So the effect is not purely the mechanical one of the same model being asked the same question twice.
• The baseline was validated against a world without monoculture. This is the methodological move that makes the finding survive scrutiny, and section 3 leans on it. They took the largest prior correspondence study — 83,000 applications to 108 Fortune 500 firms — and found that its systemic rejection rates "are very accurately predicted by employers making statistically independent decisions. Our data systematically diverges from this baseline of independence."
• The cost, stated as a number an applicant can feel. Using the deterministic replayability of the models, they simulated every applicant against every position: everyone would be recommended by at least one model somewhere. Under realistic application behaviour, "To guarantee a systemic rejection rate below 0.1%, applicants would need to submit 25 applications compared to 10 under the baseline of independence."
• Adverse impact that aggregate auditing hid. pymetrics' own published analysis reported selection rates pooled across all applications and found nothing that would trip the four-fifths rule. Disaggregated per position — which is how U.S. guidelines actually operationalise discrimination — 10.62% of the 1,746 positions show adverse impact against Black applicants; 30.70% of Black applicants applied to at least one such position, and 25.87% of applications from Black applicants went to models with adverse impact against them. The Asian figure is 14.74%. And this is a vendor that "aims to pro-actively debias each model they produce," with outside researchers having found the seven models they inspected were constrained against adverse impact during training. Per-model debiasing did not produce per-position fairness.
• The scale of the seat. The paper's own framing: "Over 90% of U.S. employers rely on hiring algorithms to screen or rank job applicants," and "As of May 2023, over 60% of the Fortune 100 and eight of the ten largest US federal agencies use HireVue's algorithms." The HireVue figure is the vendor's claim about its own reach — the 2022 paper cited a HireVue press release for its version of it (700 companies, over 30% of the Fortune 100) — and should be cited that way. The 90% figure carries a citation I did not follow, so it is the paper's claim and not yet a checked one.

The limitation a reading must carry with all of this: there is no causal evidence in the paper. Homogeneity is observed; the mechanism is inferred. That phrasing is the authors' own, from their Q&A with the Stanford Digital Economy Lab rather than from the paper's text, and it is consistent with everything the four pages read here claim.

---

Why a reading would cite it

proposals.md states the trigger as: "Cite when: a judgment-seat finding shows the same model or same-data systems deciding across institutions, or independent deployments failing identically." That is right, and there are four occasions, one of which the proposal could not have anticipated.

First: it converts the concentration lens from a fact about owners into a fact about people

LENSES.md gives concentration a lens — "who holds the capability and on what terms" — and calls it the axis between the two poles this project is named for. That lens records ownership. This entry supplies the mechanism by which ownership reaches a person who does not own anything: if a handful of vendors mediate the screening decisions of most employers, then the structure of the market is also the structure of one applicant's opportunity set, and a rejection stops being a draw from a distribution and becomes a verdict. Concentration at the vendor layer is the same fact as correlated rejection at the applicant layer, seen from the other end. A reading covering vendor consolidation, API dependency, or a lab's share of enterprise deployment can cite this entry for why that is a finding about people and not only about companies.

Second: the judgment seat — the lens the owner has drafted and rule 5 keeps a job out of

LENSES.md does not currently carry a judgment-seat lens. A draft of one sits in queue/suggestions.md, written for the owner to paste or strike, and no job may install it. Meanwhile the findings keep arriving under other lenses: on 17 August 2026 this project filed Phillips v. Parlade under Policy and regulation because a court is a government, and the 3M expert-witness report under The public square because a lawsuit is what the public does.

This entry is the mechanism that lens would need, and it also disciplines it. Neither of those two findings is a monoculture finding, and saying so is the point. One judge allegedly leaning wholly on a model, and one expert generating 85–90% of a report with ChatGPT, are single-instance delegations. Monoculture is a claim about many decision-makers sharing one component, and it becomes checkable only when someone measures whether the same people lose everywhere. If a reading ever finds that many experts, many courts or many agencies are running the same model on the same kind of question, this entry says what would have to be measured before calling it a harm — and, just as importantly, says that nobody has measured it for LLMs in institutions yet.

Third: work and the economy, where the finding is a rate rather than a headcount

LENSES.md wants actual adoption and actual displacement, and ranks shipped deployments above surveys. Algorithmic screening is the largest deployed AI system that decides something about an ordinary person's life, it has been running for years, and it produces no headcount at all — nobody is laid off, the labour effect is a change in the rate at which applications convert. Twenty-five applications instead of ten is that effect with a number on it, from real deployment data. That is a work-and-economy finding of exactly the kind the lens says it prefers and rarely gets.

Fourth, and the one the proposal could not have known: the regulators got there first, and they are using the word

The concept has escaped the fairness literature into financial supervision, which is a citable event in its own right. Gary Gensler, then SEC chair, in January 2024: AI "could promote herding with all the individual actors, the downstream applications making similar decisions because they're getting some signal from the base model or data aggregator," which "could encourage, basically, a monoculture, and all study of biology, all study of financial markets — when you tend towards monocultures you tend towards fragility." The Financial Stability Board's The Financial Stability Implications of Artificial Intelligence (14 November 2024) and its follow-up Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector (October 2025) name third-party and pretrained-model concentration and correlated trading, lending and pricing as monitored vulnerabilities.

A reading covering an AI-and-financial-stability item can cite this entry for what the regulators are actually worried about, and for the fact that the version they use is the shock version. Gensler's fragility argument is the agricultural picture. Kleinberg and Raghavan's point is that you do not need a shock. The regulators are, so far, monitoring the easier half.

Where a reading must not cite it

It is not evidence and it cannot move the needle. A formal model and a metric are not findings about the world; the pymetrics study is a finding about the world, and if a reading has that study it should cite the study and use this entry to explain what its baseline means.

It also cannot be used to convert a delegation finding into a monoculture finding. The distance between "an institution used a model to decide something" and "many institutions used the same model and the same people lost everywhere" is the entire content of this idea, and it is exactly the distance a doom-shaped reading would be tempted to close for free.

---

What it got right, and what it got wrong

Not required for idea, and done under the full prediction discipline anyway, because this is a hypothesis with a date on it and four years of subsequent literature — much of it by the same people — that grades it. Each claim carries the date made.

Claim 1 — outcome homogenization is a real, measurable phenomenon in deployed systems. Made 25 November 2022 as a hypothesis about benchmarks. Confirmed at scale, May 2026, by the authors themselves.

The 2022 paper measured on German Credit and census data and said, in its own discussion, that the question was whether this happens in real deployments. The 2026 study is the answer and it is unusually clean: four million real applications, systemic rejection above the independence baseline, and — the part that makes it stand up — the baseline calibrated against a real non-monoculture dataset that fits it well. That is the right way to answer the objection that the null model is arbitrary, and it was done before the objection was published. Confirmed.

Claim 2 — the component sharing hypothesis: shared components ⇒ more homogenization. Made 25 November 2022. Half right, and the authors said so in the paper.

Data sharing: supported, with exceptions. Model sharing: not supported. Training from scratch was the most homogeneous condition in vision, the reverse of the prediction, and the parameter-count ordering failed in language. The paper's own verdict is that sharing "is incomplete explanation of homogeneity."

This is the most misquoted result in the entry, and the misquote runs in the doom direction. The paper is cited constantly for "foundation models homogenise outcomes." Its actual foundation-model experiment found that they sometimes homogenise less than independently trained models, and that how you adapt the model matters more than whether you share it. Anyone citing this paper for the strong claim is citing the section that failed.

Claim 3 — the 2021 welfare result, ungraded for magnitude by its authors. Made 14 January 2021, magnitude supplied 1 April 2026, and the answer is reassuring.

Kleinberg and Raghavan proved the effect and declined to size it. Robert Kleinberg, Erald Sinanaj and Éva Tardos, "Price of Anarchy of Algorithmic Monoculture" (arXiv 2604.00444, 1 April 2026), generalised the model and closed the open question: "We find that monoculture and more generally decentralized optimization is close to optimal: we show a tight constant bound of 2 on the price of anarchy… for the induced game."

A factor of two, tight, is not nothing and it is not a catastrophe, and it is a hard ceiling on the 2021 mechanism rather than an estimate: equilibrium welfare is never worse than half the socially optimal value, and there is a case where it is exactly half. Any reading tempted to use the 2021 theorem as a general argument that algorithmic hiring destroys social welfare should know that the best current answer to "how much?" is "by at most half, in the worst case the model permits" — and that the people who computed it chose to describe the result as monoculture being close to optimal.

Claim 4 — the implicit forward claim that monoculture is bad for the people subject to it. Made across both papers. Contested, and the strongest counterweight is not from a critic.

Kenny Peng and Nikhil Garg, "Monoculture in Matching Markets" (NeurIPS 2024, arXiv 2312.09841, 15 December 2023) added the market effects the earlier models omit — many applicants, many firms, preferences on both sides — and reported, verbatim: "All else equal, monoculture (1) selects less-preferred applicants when noise is well-behaved, (2) matches more applicants to their top choice, though individual applicants may be worse off depending on their value to decision-makers and risk tolerance, and (3) is more robust to disparities in the number of applications submitted."

Read all three clauses. (1) confirms Kleinberg and Raghavan on the firms' side. (2) and (3) run the other way for applicants: more people get their first choice, and the advantage that goes to whoever can fire off the most applications shrinks. The harm and the benefit fall on the same market from the same cause, and an entry that carried only the harm would be worth less than one that shows both.

The two grades, and they differ — with the same author on both sides

The canon carries predictions graded twice when the predictor grades themselves. This is not a prediction, and the two-grade discipline is what it needs anyway, because the divergence here is unusually direct.

The field's grade: confirmed and getting worse. The citing literature treats monoculture as an established and growing harm, and it has real evidence. Kim, Garg, Peng and Garg, "Correlated Errors in Large Language Models" (ICML 2025, arXiv 2506.07962, 9 June 2025) evaluated over 350 models on two leaderboards and a resume-screening task: "on one leaderboard dataset, models agree 60% of the time when both models err," and "crucially, however, larger and more accurate models have highly correlated errors, even with distinct architectures and providers." Provider diversity is not error diversity, and the correlation rises with capability. Ballestero, Hosseini, Khanna and Shorrer, "Strategic Algorithmic Monoculture" (arXiv 2604.09502, April 2026) ran a coordination game on humans and LLMs and found LLMs show high baseline similarity and, though they do adjust it in response to incentives, "lag behind humans in sustaining heterogeneity when divergence is rewarded" — convergence that survives being paid to diverge.

Manish Raghavan's grade, as of 2026: the critique is weaker than its users think — and its measurement is shakier than its users think. The co-author of the 2021 paper that named the concern has published twice this year qualifying the case it started.

• Nathanael Jo, Nikhil Garg and Manish Raghavan, "The Subjectivity of Monoculture" (arXiv 2602.24086, 27 February 2026): "we argue that this question is inherently subjective, relying on two key decisions. First, the analyst must specify a baseline null model for what 'independence' should look like… different null models result in dramatically different inferences about excess agreement." They report "drastically different inferences when using a null model with item difficulty compared to previous works that do not," and reframe monoculture measurement "not as an absolute property of model behavior, but as a context-dependent inference problem." This aims directly at the denominator of H^individual. If some applicants are genuinely weaker, every system will reject them without sharing a single component, and a metric normalised against independence will call that monoculture.
• Brian Hedden and Manish Raghavan, "Algorithmic Monoculture and its Critics" (arXiv 2604.06047, 7 April 2026): "We conclude that monoculture is less problematic than its critics have supposed: commonly cited objections fail, and while other objections have some force, they are not decisive against monoculture in general."

They differ, and the disagreement is about a baseline rather than about facts. Nobody disputes that outcomes correlate. The question is what number correlation should be compared against, and whether the excess is a harm or the signal working. The 2026 hiring paper has the best answer available — calibrate the null against a real market that has no monoculture in it, and show your data diverging from it — and that answer is an empirical argument, not a settled one.

One observation is owed, and it cuts toward the sceptics. The 2022 paper's own limitations section says: "direct optimization against our measure need not lead to desirable outcomes, potentially even contributing to ethics-washing: its interpretation must be contextual." Its footnote 3 concedes that "the notions of merit and ground truth in hiring are much more subjective" and that "some individuals can be justifiably rejected from all opportunities." The subjectivity objection of 2026 is in the 2022 paper, in the authors' own words, before anyone raised it. What the four years added was a demonstration that the choice changes the answer.

---

Commonly misused as

Not required for idea. Included because this concept is a metaphor with a formalism attached, the metaphor travels much faster than the formalism, and the two say different things.

1. As the security argument — one flaw, everything falls over at once.

This is what the word means in agriculture and in the 2003 Geer report, and it is what regulators mostly mean when they use it. It is not what the 2021 theorem says. That paper's contribution is precisely that welfare drops in the absence of shocks, during normal operation, invisibly — "these effects can persist at low levels even without a shock-like disruption to call our attention to them." Both concerns are real; they have different evidence, different remedies and different urgency, and a reading that reaches for the correlated-outage picture when the finding is about quiet correlated rejection has described the wrong risk.

2. As proof that agreement between models is a harm.

The measurement is a ratio against an assumed-independent baseline, and choosing that baseline is a judgment call — this is the standing 2026 objection, from one of the concept's own authors. Two accurate systems will agree about a clear case, and agreement about clear cases is what accuracy is. Excess agreement is only a harm if the excess is not explained by the difficulty of the decisions. Ask what the null model was before believing a homogenization number, including the ones in this entry.

3. Confused with model collapse.

Model collapse is a training-dynamics claim: models trained on the output of models degrade across generations. Monoculture is a deployment claim: many decision-makers using one system produce correlated decisions about the same people. Different mechanism, different timescale, different evidence, no dependence between them. They co-occur in commentary because both can be gestured at with the word "homogenization."

4. Confused with stylistic homogenization — "everything sounds like ChatGPT."

Also real, also not this. The complaint that AI-assisted writing is converging is about outputs looking alike. Outcome homogenization is about the same individuals being excluded by everyone. The em-dash discourse and a person rejected from twenty-five jobs by one vendor's forty-two models are not the same finding, and only the second one has H^individual in it.

5. Confused with Goodhart's law.

goodharts-law-1975 is one measure degrading because it is optimised against. This is many decisions correlating because they share a source. A gamed benchmark is Goodhart; identical rejections across employers are monoculture. The overlap is that both punish naive faith in a number, and that is not enough to make them the same idea.

6. As an argument that the foundation-model layer is the problem.

Tempting, and unsupported by the paper it would be cited from. §5 of the 2022 paper tested exactly this and found model sharing did not reliably increase homogenization. The 2026 empirical result is monoculture at the vendor layer — one talent platform, forty-two of its own models — not at the foundation-model layer. Whether a shared frontier LLM behind many institutions' decisions produces the same effect is an open question with suggestive evidence (the ICML 2025 correlated-errors work) and no deployment study. Say which layer a finding is about.

7. As a verdict on an individual decision.

Homogenization is a population statistic. It says nothing about whether any particular rejection was wrong, and the 2026 paper says outright that it establishes no causation. A reading that uses this entry to characterise one person's case has used an aggregate as a diagnosis.

8. The one this room has to watch in itself.

The tell bitter-lesson-2019 states and credits to moravecs-paradox-1988: a citation available to both sides of a reading decides nothing. From this material one can obtain, at zero evidentiary cost, "algorithmic hiring institutionalises exclusion — twenty-five applications where ten used to do" (AImageddon) and "consistent screening matches more applicants to their first choice and blunts the advantage of whoever can apply the most times" (AItopia). Both are published findings. Neither is a needle movement. The defence is that this entry is cited to say what would have to be true for a homogenization claim to hold, never as a reason to believe it already is — and that a reading with no measured deployment in its window has nothing here to explain.

---

Sources

Verification status, stated once rather than hedged throughout. Both primary papers and the 2026 follow-up were read directly as rendered pages, not through a summarising layer: the 2022 NeurIPS camera-ready in full (its ten-page body plus the reference list); the 2021 arXiv version's pages 1–6 and 11–16, which carry the introduction, the model, Theorem 1 and its proof, the multi-firm observation and the conclusion — its pages 7–10, which analyse specific ranking families, were not read; and the first four pages of the 2026 FAccT paper, covering its abstract, introduction and data section. Every quotation attributed to those three is transcribed from that reading and is verified, including the statement of Theorem 1, the component sharing hypothesis, footnote 4 on the relationship between the two formalisms, and every number in the pymetrics section. Everything else below arrived through a fetch layer that returns a summarising model's rendering of a page; where a quotation from those is load-bearing it was retrieved twice with consistent wording, and where I could not confirm something the text says so. PNAS refused direct retrieval (HTTP 403), so the published version was reached through the author-hosted arXiv copy, which carries a footnote saying it also appears in PNAS.

Primary — the two papers this entry is about.

• Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky and Percy Liang, "Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?", NeurIPS 2022 (36th Conference on Neural Information Processing Systems), arXiv:2211.13972, submitted 25 November 2022. arxiv.org/abs/2211.13972; camera-ready PDF. Read in full. Source of the abstract, the component sharing hypothesis, the metrics, the data-sharing and foundation-model experiments, the individual-versus-group finding, the societal-challenges section and the limitations. The arXiv abstract and the NeurIPS camera-ready abstract differ in their closing sentence; the version quoted above is the camera-ready.
• **Jon Kleinberg and Manish Raghavan, "Algorithmic Monoculture and Social Welfare," PNAS 118(22), e2018340118, 2021; arXiv:2101.05853, v1 14 January 2021, v2 1 June 2021. arxiv.org/abs/2101.05853. Model, Theorem 1, Definitions 1–3, the Braess' paradox framing and the conclusion read directly from the arXiv PDF. Not in this canon and it should be**; this entry is the second paper in the line and the first one carries the theorem.

Primary — the deployment study.

• Rishi Bommasani, Sarah H. Bana, Kathleen A. Creel, Dan Jurafsky and Percy Liang, "Algorithmic Monocultures in Hiring," FAccT '26, Montreal, 25–28 June 2026; arXiv:2605.27371, 26 May 2026; doi 10.1145/3805689.3812400; arxiv.org/abs/2605.27371; project page algorithmichiring.github.io. Pages 1–4 read directly: the pymetrics dataset, the systemic-rejection results, the independence baseline validated against the 83,000-application correspondence study, the per-position adverse-impact figures, and the HireVue and 90%-of-employers claims. The statement "there is no causal evidence in our paper" is from the authors' Q&A with the Stanford Digital Economy Lab, digitaleconomy.stanford.edu, read through the fetch layer, not verified against the paper's own limitations section. That Q&A also reports a "10% of applicants applying to four positions" figure that I did not find in the four pages I read; the 4%-at-ten-positions figure quoted above is from the paper itself.

The counterweights — all read as abstracts through the fetch layer.

• Kenny Peng and Nikhil Garg, "Monoculture in Matching Markets," NeurIPS 2024; arXiv:2312.09841, 15 December 2023. arxiv.org/abs/2312.09841. The three-clause finding is quoted verbatim from the paper's HTML abstract, retrieved directly.
• Robert Kleinberg, Erald Sinanaj and Éva Tardos, "Price of Anarchy of Algorithmic Monoculture," arXiv:2604.00444, 1 April 2026. arxiv.org/abs/2604.00444. The tight bound of 2 is quoted from the abstract, retrieved twice with identical wording. The proof was not read; the bound is reported as the authors state it, and a reading leaning on it should say so.
• Brian Hedden and Manish Raghavan, "Algorithmic Monoculture and its Critics," arXiv:2604.06047, 7 April 2026 (revised 20 April). arxiv.org/abs/2604.06047. Abstract only.
• Nathanael Jo, Nikhil Garg and Manish Raghavan, "The Subjectivity of Monoculture," arXiv:2602.24086, 27 February 2026. arxiv.org/abs/2602.24086. Abstract only.

The LLM-era evidence — abstracts only.

• Elliot Kim, Avi Garg, Kenny Peng and Nikhil Garg, "Correlated Errors in Large Language Models," ICML 2025; arXiv:2506.07962, 9 June 2025. arxiv.org/abs/2506.07962. The 350-model evaluation, the 60% error-agreement figure and the finding that correlation rises with accuracy.
• Gonzalo Ballestero, Hadi Hosseini, Samarth Khanna and Ran I. Shorrer, "Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games," arXiv:2604.09502, 10 April 2026. arxiv.org/abs/2604.09502.

Regulators and the metaphor's history — through search summaries and secondary reporting.

• Gary Gensler, then SEC chair, on AI monoculture and herding, in a Public Citizen fireside chat, January 2024, and in the SEC's Office Hours: Systemic Risk in Artificial Intelligence, 19 September 2024. The sec.gov transcript refused direct retrieval (HTTP 403); the quotations above are as returned by search summaries of secondary reporting and are not verified against a primary transcript. A reading that wants to quote this should get the transcript first.
• **Financial Stability Board, The Financial Stability Implications of Artificial Intelligence, 14 November 2024 (fsb.org) and Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector, October 2025 (fsb.org). Landing pages and summaries only; neither report was read.**
• Geer, Bace, Gutmann, Metzger, Pfleeger, Quarterman and Schneier, "CyberInsecurity: The Cost of Monopoly," CCIA, 24 September 2003. The report and Geer's dismissal from @stake on the day of publication are from search summaries and Wikipedia; the report itself was not read for this entry. Note that Kleinberg and Raghavan do not cite Geer: their security reference is Birman and Schneider, "The monoculture risk put into context," IEEE Security & Privacy 7(1), 2009, whose title suggests a qualification of the analogy rather than an endorsement of it. Not read here, and the one to read before leaning on the agricultural picture.

Named above and not read for this entry, listed so a later session knows what is owed: Bommasani et al., "On the Opportunities and Risks of Foundation Models" (2021); Creel and Hellman, "The Algorithmic Leviathan" (Canadian Journal of Philosophy 52(1), 2022); Engler, "Enrollment algorithms are contributing to the crises of higher education" (Brookings, September 2021); and Kline, Rose and Walters, "Systemic discrimination among large U.S. employers" (NBER w29053, 2021). That last is almost certainly the 83,000-application study the 2026 independence baseline is calibrated against — the counts of applications and of Fortune 500 firms match exactly — but the 2026 paper's reference [38] was not resolved, so the identification is mine and not the paper's. All four are ancestors or load-bearing citations of the papers above, and the first two are the strongest unlisted canon candidates this entry found.

In this canon. bitter-lesson-2019 was read in full, for the house's grading discipline and for the both-sides tell in section 4 — that tell is its formulation, and it credits moravecs-paradox-1988 for it. transformer-2017, scaling-laws-2020, goodharts-law-1975, frey-osborne-2013, godel-incompleteness-1931 and turing-halting-1936 were checked by header and by their entries in proposals.md rather than read through; the claims made about them here are about what kind of thing they are, which is all this file needs and less than a full reading would support. LENSES.md, README.md and queue/suggestions.md were read for the citation occasions in section 2, and digests/2026-08-17-12.md for the two judgment-seat findings named there. The judgment-seat lens draft is the owner's, sits unstruck in the suggestions box, and is not installed in LENSES.md.