Algorithmic Monoculture: One AI for Everyone, Risk or Remedy?
Two studies published just months apart seem to say opposite things. Looked at closely, they measure different things. Kyle Behm kept submitting job applications and receiving the exact same rejection. In Cathy O'Neil's account in Weapons of Math Destruction, picked up by the MIT paper, Behm only discovered at the end that all the companies were using the same screening system, which apparently penalized him for a mental health condition. They were not many independent judgments; it was a single judgment repeated everywhere.
This is the fear we call algorithmic monoculture: many decision-makers relying on the same algorithm, or on algorithms so similar that they fail together. The opposite paradox, however, is just as real. If every company had its own algorithm, the system wouldn't necessarily become fairer; we might end up with inconsistent decisions, impossible to audit, easier to manipulate, and with a person rejected everywhere due to an accumulation of quirks rather than a single error.
On September 29, 2026, MIT presented a study by Brian Hedden and Manish Raghavan that puts the danger in perspective. In May, Stanford had published the first major evaluation on real-world data, finding systemic rejections to be more frequent than expected. The guiding question is simple to state and uncomfortable to settle: is algorithm diversity always a good thing, or might a single transparent, audit-ready system, perhaps composed of many models, be preferable? I'll offer the cautious answer up front: it depends, and I'll try to explain on what.
Banana, a Metaphor with Limits
The term comes from agriculture, and the metaphor works: a field with a single variety yields well until a disease strikes that specific variety. Anyone who knows the history of the Gros Michel banana knows this well: decimated in the mid-20th century by Panama disease and replaced by the Cavendish we find in supermarkets today, cloned and now threatened by a new variant of the same fungus that reached the Americas in 2019. But Hedden and Raghavan issue a caveat: an analogy is not proof, and their work stems precisely from the suspicion that it is being applied automatically.
In the world of computing, the term covers different situations. There is strict monoculture, where everyone uses the exact same algorithm. There is market monoculture, where many organizations buy from the same vendor even if the models differ under the hood. There is component monoculture, where distinct systems share data, base models, or objectives. There is outcome monoculture, where different systems end up producing similar results. And there is generative monoculture, which we will discuss, where a single language model narrows the diversity of responses compared to the human texts it learned from.
The concept has a precise lineage. In 2021, Jon Kleinberg and Raghavan himself showed that a slightly more accurate algorithm, adopted by everyone for individual convenience, can leave society worse off than before.
Stanford, Numbers from the Field
The Stanford study, authored by Rishi Bommasani, Sarah Bana, Kathleen Creel, Dan Jurafsky, and Percy Liang, is the first to observe real decisions made by an automated hiring vendor across multiple companies. The paper, presented in June at the FAccT conference, analyzes pymetrics data between 2018 and 2022: approximately 4.2 million applications from 3.4 million people for 1,746 positions across 156 companies in 11 sectors. These are not resumes, but online games, after which a model for each position assigns a "recommended" or "not recommended" label. Those who are not recommended, as a rule, never meet a human being.
The first result concerns systemic rejections. Among those submitting four applications, 10% are rejected everywhere—more than would occur if each company decided statistically independently. The key point is the excess relative to that benchmark, not the percentage itself, and the gap persists at ten applications, where 4% are rejected by all positions. The scope must also be calibrated: 84% of people applied only once, and barely ~1% applied four times. The exact same model used by different companies exists (42 models, 142 pairs of companies) but is rare, and the correlation stems primarily from the fact that the same games feed different models.
As a sanity check, the authors use the study by Kline, Rose, and Walters, which sent 83,000 fake resumes to 108 large American companies: there, total rejections aligned with independent decisions. Intrigued, but not a twin experiment: populations, roles, and success rates differ, as the authors acknowledge. There is also a reassuring note: simulating evaluations of 939 applicants across 495 models, no one is rejected by all, though bringing total rejections below 0.1% would require applying to 25 positions instead of 10.
The second result is more technical and perhaps more instructive. In the United States, the four-fifths rule signals potential adverse impact when a group is recommended at less than 80% the rate of the most favored group. On aggregated data, the ratio for Black applicants is 0.839, making everything appear fine. But the law operates position by position, and evaluated that way, 10.6% of positions penalize Black applicants and 5.3% penalize Asian applicants. 25.9% of applications from Black individuals and 14.7% from Asian individuals end up in positions with this disparity. The study draws a clear comparison: if only men were recommended for doctor positions and only women for nurse positions, the average might balance out, but the problem would remain.
Limitations acknowledged by the authors themselves remain. There is only one vendor and one type of test, based on games. Measures of candidate quality and model validity are missing, and it is unknown how companies use the recommendations. Furthermore, a statistical disparity may justify an investigation, but it does not by itself establish that a practice is illegal.
MIT's Defense
The work by Hedden, a philosopher, and Raghavan, a computer scientist, published in a philosophy journal, does not argue that monoculture is good. It argues that the most common accusations do not hold up or are insufficient. It is a theoretical work with simulations, not a field study.
Take exclusion. In a market with a fixed number of positions and applicants, the authors argue, the same number of people remain unemployed whether there is a single algorithm or many. Who stays unemployed changes, not how many. Even when all companies rank applicants identically, the runner-up can still be hired by the second company because the first choice was already taken. And multiple algorithms do not guarantee fairness: if each prefers a different candidate, someone ranked second by everyone can still end up excluded. The paper adds a rarely discussed advantage: a single criterion makes it easier to discover and correct discrimination, whereas in a patchwork of algorithms, outcomes depend on a complex web that is difficult to untangle.
For Raghavan, there is even an effect on candidate power: if all companies pursue the same people, those individuals can secure better offers. However, the paper treats this idea with caution, noting that the economic literature on the topic focuses primarily on those who are already employed. On the front of individual autonomy, the authors observe that fear of being unable to improve one's score applies only to a naive monoculture based on immutable traits, whereas the opposite risk—learning to game the system—exists even with many different algorithms.
An attentive reader will notice where these conclusions are fragile. The model assumes hiring is conducted with a single, end-to-end algorithm, whereas Stanford observes only a preliminary screening phase: for MIT, getting an interview matters little if it doesn't change who gets the job, but one could object that for someone rejected upfront, that interview is everything. Moreover, the simulations idealize the real world, where job openings do not all appear simultaneously and markets are full of friction. And it is telling that the primary target of the critique, Creel, is also a co-author of the Stanford study: the two papers are in direct dialogue, almost a long-distance back-and-forth.
Ensembles, Exploration, and Doubts
The objection that the authors themselves consider most serious concerns discovery. If everyone trusts the same ranking, disadvantaged applicants are never tested; the system never learns that it was wrong about them. In the language of multi-armed bandit problems—those where one must decide between exploiting current knowledge and exploring—a population of decision-makers relying on the same information is more likely to pick the wrong arm and never realize it. The paper's simulations confirm this, provided the single algorithm is no better than the many it replaces.
But there is more. In hiring, no company can hire the top-ranked applicant if another company has already hired them, so some forced exploration occurs regardless. Real-world platforms are not pure exploiters either: they already introduce doses of randomness, and the paper even references a study where an algorithm accounting for uncertainty selected more diverse candidates than one ordering strictly by estimated score. Anyone familiar with Oblique Strategies—the deck of cards published in 1975 by Brian Eno and Peter Schmidt to help artists break through creative blocks—recognizes the principle: sometimes an explicit disruptive element is needed.
The most intriguing proposal is the ensemble of models: instead of many algorithms deciding separately, a single system combines many, for instance by averaging their scores. In simulations, according to the authors, this "multi-voiced" monoculture achieves results equal to or better than both simple monoculture and a patchwork of algorithms. Potential advantages: less noise, uniform decisions, centralized control. But questions arise here. Combining models built on the same data does not guarantee independence. Averaging can crush minority signals. An aggregated score can obscure who is responsible for an error. And a single point of failure remains. Hedden himself warns that it is unclear how practical this is. Is an ensemble of models pluralism, or just a more elaborate monoculture?
Finally, domain matters. The paper notes that hiring, credit, and university admissions have fixed and exclusive slots, where monoculture does not create more excluded individuals. Where such constraints do not exist, like film awards, it might exclude more. For medicine and justice, the authors point to future studies, and regarding scientific research or generated content, they warn that things might work differently and worse.
Same Ingredients, Different Labels
The fear becomes clearer when looking at what goes into applications. An app may have a different interface, prompts, and fine-tuning than its competitor while relying on the same base model, same data, and same safety filters. The Stanford authors state this explicitly: vendor concentration alone is insufficient to measure the phenomenon, because models from different companies can reject the same people when sharing data or architecture. A study presented at ICML 2025 across more than 350 models confirms this for language models: on a benchmark test, when two models make a mistake, they agree on the wrong answer 60% of the time, and larger, more accurate models fail in very similar ways even with different architectures and creators. The authors verify these effects in candidate screening and in LLMs used as judges.
The most insidious version is generative monoculture. Fan Wu, Emily Black, and Varun Chandrasekaran described it as a narrowing of output diversity relative to the training data: for example, book reviews being almost entirely positive, even for titles with mixed receptions, or code covering fewer strategies. The picture is ambivalent: reducing diversity sometimes improves quality—code can become more efficient—but it becomes a problem when non-conventional opinions vanish. The authors add a sobering detail: changing sampling strategies or prompt formulations is not enough to fix the phenomenon; the root causes appear embedded in model alignment.
Those working with local, quantized models or enterprise deployments should consider a distinction that serves here as a thought experiment rather than an established result: operational sovereignty over execution is not the same as model diversity. If many open projects derive from the same base model and same data, having the weights locally grants autonomy, but not necessarily distinct perspectives. The same applies to agents, where the question remains open: is a team of ten agents using the same model, same tools, and same evaluation criteria truly pluralistic, or is it a monoculture lined up ten deep? And if the model judging the outputs resembles the one producing them, who audits whom?
Journalism knows this well. Kyle Chayka, in his book Filterworld cited by the MIT paper, describes how recommendation algorithms flatten taste. If newsrooms, search engines, and AI assistants rely on similar models, minority sources can disappear without censorship, simply because they are statistically less probable.
Antidotes, Controls, and Contestability
What should be done, then, beyond slogans like "let's use different algorithms"? The Stanford paper proposes concrete tools. First: measure impact per position, not on a vendor's overall average—something that does not always happen today. Second: monitor the market to understand how widespread certain vendors are and how interconnected they are, given that current data collected by regulators is aggregated and anonymous, preventing the tracking of an individual across companies. Third: guarantee researchers access to data, as the European Digital Services Act does for large platforms. The authors themselves note they worked under an agreement with pymetrics that prevented the vendor from blocking publication, and they fear their research might discourage other vendors from sharing data.
Regarding metrics, beyond outcome correlation, error correlation is needed: do systems agree specifically when they are wrong? We need to understand the extent of data and model overlap, whether the system explores non-conventional candidates, and whether a systematic error can be attributed to a specific component. And above all, whether a rejected individual has a meaningful way to contest the outcome. Greater decision diversity is not inherently valuable if it stems from noise or incompetence, which only worsens outcomes. The goal is useful diversity that brings valuable information, backed by human accountability.
Europe does not have a regulation called "algorithmic monoculture," and it would be inaccurate to attribute one to it. However, recruitment falls under high-risk systems in the AI Act, with obligations covering risk management, data quality, documentation, and human oversight. One deadline needs updating: the August 2026 date cited in the Stanford study has been superseded. The simplification package known as Digital Omnibus, approved by the Council on June 29, 2026, became Regulation (EU) 2026/1744, published on July 24 and in force since July 27; according to a dedicated analysis, it shifts obligations for standalone high-risk AI systems to December 2, 2027. The most direct connection remains competition policy. A policy brief by the Commission's Directorate-General for Competition—which explicitly notes it does not represent the institution's official stance—identifies bottlenecks in generative AI such as data, chips, and compute, highlighting the difficulty of switching cloud providers. It points to open models and interoperability as remedies. A small number of dominant vendors is not illegal per se, but creates dependencies, and the Stanford study adds the risk that a vendor outage could halt hiring across thousands of companies.
Questions Still Open
Let's return to Behm and the initial question. The answer is not simply choosing between one algorithm and many. It is asking how many systems are truly independent, how much data and model architecture they share, whether their errors are correlated, who bears the harm, whether the system explores alternatives, and who can audit and correct it.
The two studies, in the end, do not contradict each other. MIT asks what logically follows from certain models and finds it is far less than assumed—a healthy antidote to slogans. Stanford measures what actually happened in a concrete ecosystem and shows that the risk is not purely hypothetical, even if limited to one vendor, one test format, and a small fraction of applicants. Consistency that makes errors more visible is an advantage, but only if someone is watching; diversity that multiplies errors with no one verifying them is merely confusion.
Unresolved questions remain that neither paper closes. What happens when generative models and agents replace ranking systems? Can an ensemble of models truly be pluralistic? How much randomness are we willing to accept in a life-changing decision? Algorithm diversity can help, but it cannot replace the right to be heard: asking how many AIs are deciding matters less than knowing how much independence exists among them and whether, after a rejection, there is still a human to answer to.