The Monoculture Trap: When AI Coordinates Too Well

AI agents are excellent at finding the obvious answer. That sounds like a compliment until the task is to avoid everyone else’s obvious answer.

Imagine three firms using AI assistants to screen applicants, forecast demand, or decide which customer segments deserve attention. If the goal is consistency, shared focal points are useful. Everyone reads the same policy, applies similar criteria, and avoids the usual mess of human improvisation. Lovely. The spreadsheet smiles.

Now change the environment. Suppose the goal is not to converge but to diversify: avoid crowding into the same trade, avoid interviewing the same narrow group of candidates, avoid recommending the same congested supplier, avoid all agents flagging the same “safe” option because it looks salient in training data. In that world, coordination means not stepping on one another’s shoes.

The paper “Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games” by Gonzalo Ballestero, Hadi Hosseini, Samarth Khanna, and Ran I. Shorrer gives us a clean experimental language for that difference.¹ Its core move is simple but important: algorithmic monoculture is not only about models being similar at baseline. It is also about whether models can adjust their similarity when incentives change.

That distinction matters because businesses are not merely deploying one chatbot in isolation. They are increasingly building multi-agent workflows: one agent screens, another ranks, another negotiates, another summarizes, another executes. The boring question is whether each agent is individually capable. The more interesting question is whether the system becomes too coordinated for its own good.

The paper separates two kinds of monoculture that are usually blended together

Most discussions of algorithmic monoculture begin with a familiar worry: if many people or firms rely on similar models, their decisions may become correlated. That can produce correlated errors, crowded strategies, or systemic blind spots. In a hiring market, similar screening tools may reject the same applicants for the same noisy reasons. In finance, similar decision aids may herd users toward the same trades. In operations, similar recommendation engines may all route attention to the same suppliers, channels, or risks.

The paper sharpens this concern by separating two mechanisms.

This distinction is the paper’s first real contribution. Without it, we might look at a set of similar LLM outputs and say, “These systems are homogeneous.” True, but incomplete. The more operational question is whether that homogeneity is passive or strategic.

Passive similarity means the model keeps choosing the same answers because its training, architecture, defaults, and decoding process push it there. Strategic similarity means the model understands the social environment: “I should choose what the other agent will choose” or “I should avoid what the other agent will choose.”

Those are different capabilities. Unfortunately, they can collide.

The experimental design is simple because the mechanism is the point

The authors adapt a classic coordination-game setup. Subjects answer open-ended prompts such as naming a city, color, flower, food item, month, positive number, or letter. The experimental arms differ only in incentives:

The main metric is the agreement rate: the probability that two independently sampled subjects provide the same answer. In the coordination arm, higher agreement is better. In the divergence arm, lower agreement is better. This is why calling the measure a “coordination index” would be misleading. In some environments, the best form of coordination is disagreement.

The human sample contains 301 U.S.-based Prolific participants. The model sample covers 16 open- and closed-source LLMs, with 50 independent responses per model, per question, per treatment arm. The paper reports 28,800 LLM responses in the main experiment. Responses are standardized so that spelling, pluralization, acronyms, and naming variants do not artificially split equivalent answers.

This design is not glamorous. It is better than glamorous: it is diagnostic. The prompts are deliberately open-ended, so the experiment is not just measuring whether models prefer the first multiple-choice option. The arms differ only in strategic incentives, so the comparison can isolate whether agents regulate similarity when matching or mismatching is rewarded.

LLMs are not merely similar; they are unusually good at converging

The headline numbers are stark.

In the picking arm, LLMs already agree far more often than humans: 58% versus 14%. That is primary monoculture. Independent instances of the same model, with no reward for matching, tend to land on the same answers.

In the coordination arm, the difference becomes useful. LLMs reach 72% agreement, compared with 31% for humans. They are much better at finding focal responses. If two model instances are both asked to name a city and rewarded for matching, they are quite good at converging on the same obvious city.

For many enterprise workflows, that sounds attractive. Standard operating procedures, compliance explanations, brand-safe responses, taxonomy mapping, structured extraction, report classification: in these settings, correlated behavior may reduce friction. When everyone needs the same canonical answer, monoculture is not necessarily a bug. Sometimes the whole point of automation is to remove the charming local improvisations that made last year’s process audit resemble archaeological fieldwork.

But the same mechanism flips sign in the divergence arm. Humans reduce agreement to roughly 4%. LLMs reduce it too, but only to 27%. They understand that they should avoid matching, yet they still collide far more often than humans.

This is the paper’s central asymmetry: LLMs are excellent at “finding each other” and weaker at “staying out of each other’s way.”

Divergence is not the same as randomness, and that is where the trap appears

The theory section explains why the divergence problem is not a small technical wrinkle. In a symmetric environment without communication, two agents can successfully diverge if they have either enough heterogeneity or reliable randomization.

For a simple game with $n$ valid actions, uniform randomization gives a match probability of $1/n$. In a divergence game, that means the expected success probability is $(n-1)/n$. With many possible valid answers, randomization should make divergence fairly easy.

So why do LLMs still collide?

One answer is primary monoculture. If model instances share the same internal salience map, “choose something obscure” may itself become a focal strategy. Both agents try to avoid Paris, red, rose, pizza, and January. Then they may both choose the same second-layer obscure answer. The model escapes the front door and meets itself at the side entrance. Very efficient. Also not the goal.

This is why the paper’s mechanism-first framing matters. The problem is not that LLMs cannot produce rare options. The supplemental experiments show they can generate large lists of valid answers. The problem is that, when asked to act strategically in the original divergence task, their attempts at obscurity remain correlated.

The random-list experiment is especially revealing. Models are instructed to generate a list of 100 valid answers, or as many as possible, and then choose one uniformly at random. Under this setup, agreement with the original divergence arm drops sharply: average agreement is about 4% when comparing the same model’s random-arm choices with its divergence-arm choices, and 3% for random pairs of models. When comparing pairs within the random arm, average self-agreement is 12% and cross-model agreement is 4%, still much lower than the 27% same-model agreement in the main divergence arm.

That test is not a second thesis. It is an ablation-style mechanism probe. It says the divergence failure is not simply “models lack enough answers.” Nor is it fully “models cannot ever randomize.” The issue is that the strategic prompt alone does not reliably produce the kind of independent diversification that the game rewards.

The reasoning text shows models often know the strategy but still fail the system-level task

The paper uses an advantage that human experiments rarely have: LLM subjects produce reasoning text. The authors analyze that text in two ways.

First, they use semantic similarity methods. They split model reasoning into sentences, screen for words related to salience, obscurity, and randomization, embed the sentences, and classify them against reference meanings such as choosing salience, avoiding salience, choosing obscurity, and choosing randomization.

Second, they use an LLM-as-judge approach. A judge model reviews the reasoning text and classifies the strategic logic into categories such as primary salience, secondary salience, Schelling salience, obscurity, higher-order obscurity, randomization, or not mentioned.

The result is important: the models’ reasoning broadly tracks the incentive structure. In the coordination arm, the judge frequently identifies secondary salience and Schelling salience. In the divergence arm, the most common categories become obscurity and higher-order obscurity. The models are not simply ignoring the task.

This creates a useful distinction for business readers:

That last row is the uncomfortable one. The paper reports that strategic reasoning categories shift substantially across arms, sometimes more dramatically than the agreement-rate changes visible in final choices. In other words, the model may articulate the right strategic concept while still producing correlated actions.

For enterprise AI systems, this matters because evaluation often stops at explanation quality. A model says, “I diversified the options,” and the answer looks plausible. But a multi-agent system is judged by joint outcomes, not by whether each agent can narrate a sensible private rationale. Correlated failure with a good explanation is still correlated failure. It just has better documentation.

Turning up temperature helps divergence but damages coordination

A tempting fix is to increase temperature. More stochasticity should reduce agreement, and therefore help divergence.

The paper tests this by replicating the experiment using the highest and lowest allowed temperatures for the models that support such settings. The result follows the theoretical tradeoff. Low temperature increases agreement across all treatments. High temperature lowers agreement across treatments, improving the direction of divergence but weakening coordination.

That is not a free lunch. It is not even a discounted lunch. It is a menu with one price printed in two columns.

This is a useful correction to a common misconception: “Just increase temperature” is not a governance strategy. It is a blunt instrument. If the workflow needs both standardization and diversification at different steps, a single global decoding setting cannot solve the design problem.

A procurement assistant may need low variance when mapping invoices to accounting categories. The same organization may need high diversity when generating independent supplier-risk hypotheses. Treating both as “AI output” and applying one setting across the workflow is how a configuration panel becomes a risk-management department. It should not.

Model diversity helps, but it does not make the problem disappear

The paper also compares agreement across different LLMs, not only independent copies of the same model. Cross-model agreement is lower, which is exactly what the theory predicts: heterogeneity helps coordinated divergence.

But the results still show substantial monoculture. Across different LLM pairs, average agreement is 37% in picking, 53% in coordination, and 9% in divergence. Those numbers are lower than same-model agreement in the relevant places, but still above the human benchmarks in the main comparisons.

This matters because many organizations treat “multi-model architecture” as a magic charm. Use one proprietary model, one open model, one small model, one reasoning model, and surely the system is diverse. Sometimes, yes. Enough? Not automatically.

The paper’s evidence suggests that different model families can reduce correlated behavior, especially in divergence tasks, but they may still share enough salience structure to converge more than humans. That should not surprise anyone. These models are trained on overlapping public text, optimized with related preferences, and deployed through similar instruction-following interfaces. Different logos do not guarantee independent cognition. The logo department will be devastated.

The practical lesson is to measure behavioral correlation directly. Do not infer diversity from vendor count, parameter count, open-versus-closed status, or whether a model has a charmingly different refusal style.

Personas move behavior toward humans, but not far enough

Another tempting intervention is personas. If the model is told to act as different kinds of people, perhaps it will diversify.

The authors test this by constructing personas from human-subject characteristics such as gender, age, occupation, race or ethnicity, and education level, then rerunning the experiment. Personas shift LLM behavior closer to human behavior. But the main qualitative results remain. In particular, persona-equipped LLMs still lag behind humans in the divergence arm. The paper also notes that, when assessing all LLM pairs, adding personas does not improve divergence performance and hurts coordination performance.

Again, the result is not “personas are useless.” It is more precise: personas introduce some heterogeneity, but they are not a reliable substitute for actual independent diversification.

That distinction matters for product design. Many AI-agent demos create five named agents: the analyst, the skeptic, the strategist, the compliance officer, the customer advocate. The interface looks plural. The reasoning may sound plural. But if all five agents are powered by the same model, prompted with the same managerial clichés, and drawing from the same salience map, the system may be performing a one-person panel discussion in costume.

Personas are useful when they encode genuine constraints, distinct information, different objective functions, or specialized evaluation criteria. They are weak when they merely decorate the same reasoning process with demographic or occupational labels.

Identity information changes reasoning more than outcomes

The paper also varies what LLM subjects are told about the other player. In one condition, the model is told it faces an identical copy. In another, it is told it faces another person.

The average outcome effects are small and not statistically significant. In the copy condition, agreement rises slightly in coordination and falls slightly in divergence; in the person condition, there is no statistically significant average effect. However, the textual analysis shows that models’ stated reasoning does respond to the identity information. Under the copy condition, discussions of randomization become more prevalent in divergence, and Schelling-salience reasoning becomes more prominent in coordination.

This is another place where the article should not overclaim. The identity manipulation is best read as a mechanism probe and robustness-style variation, not as a ready-made enterprise intervention. It shows that prompts can move the model’s representation of the strategic environment. It does not show that identity labels alone reliably fix joint behavior.

That gap between reasoning and outcomes is the recurring theme. The model can recognize the problem. It can discuss the problem. It can sometimes improve. But the system-level behavior remains more correlated than the business user may expect.

What the paper directly shows, and what Cognaptus infers for business use

The paper directly shows that, in controlled open-ended coordination and divergence games, LLM subjects exhibit high baseline agreement, strong same-answer coordination, and weaker-than-human coordinated divergence. It also shows that several plausible interventions—random-list prompting, temperature changes, co-player identity information, cross-model pairing, and personas—modify behavior but do not fully remove the asymmetry.

The business inference is not that every AI workflow will herd itself into disaster. That would be a theatrical reading, and theater has enough problems already.

The better inference is that businesses need to classify AI-agent tasks by whether they reward convergence or divergence.

This classification step is cheap and underused. Before deploying a multi-agent system, ask one boring question: Should these agents agree?

If yes, primary monoculture may be helpful, provided the focal answer is correct and the stakes of correlated error are managed. If no, then similarity becomes an operational hazard. The evaluation target must shift from individual answer quality to joint distribution.

The operational metric is not accuracy; it is correlation under incentives

Many AI evaluations ask whether the model gets the answer right. That remains necessary. It is not sufficient.

For multi-agent systems, the paper suggests a second layer of evaluation:

1. Baseline overlap: How often do agents produce the same answer with no coordination incentive?
2. Convergence ability: How often do agents match when matching is desirable?
3. Divergence ability: How often do agents avoid matching when difference is desirable?
4. Cross-model overlap: Does using different models materially reduce agreement?
5. Prompt sensitivity: Do personas, identity cues, or temperature changes alter final outcomes, not just reasoning text?
6. Portfolio effect: Does the set of outputs improve system-level performance, or merely repeat one strong-looking answer?

This is where the paper’s experimental simplicity becomes practically valuable. A company does not need to reproduce the whole academic design. It can create domain-specific variants.

For hiring, measure whether independent AI screeners reject the same candidates and whether their errors concentrate by background, school, keyword, career gap, or writing style. For procurement, measure whether independent agents recommend the same suppliers under supply-shock scenarios. For trading research, measure whether agents trained on different signals still produce overlapping entry points. For content strategy, measure whether “different” agents all propose the same three topics in different suits.

The point is not to worship diversity as a decorative value. The point is to know whether the production system needs variance, and then test whether the AI stack can actually supply it.

Boundaries: this is a clean experiment, not a direct enterprise field trial

The paper’s strength is also its boundary. The games are controlled, symmetric, and open-ended. They are designed to isolate coordination mechanisms, not to simulate a full hiring market, trading desk, admissions office, or procurement department.

Several limits matter for interpretation.

First, the human comparison uses a specific U.S.-based participant pool. Coordination behavior can depend on culture, shared references, and context. The authors intentionally use a relatively homogeneous human sample because salience is culturally sensitive, but that also means the human benchmark is not universal.

Second, LLMs do not receive real monetary incentives. They follow instructions. This is normal for LLM experiments, but it means the comparison is between paid humans and instruction-following models, not identical motivational systems.

Third, agreement rate is a clean proxy for similarity, but real enterprise outcomes have richer payoff structures. In hiring, two screeners rejecting the same candidate may be bad if the candidate is strong, good if the candidate is clearly unqualified, and ambiguous if downstream review catches the error. Agreement is a risk signal, not a complete welfare measure.

Fourth, the tested interventions are not exhaustive. A stronger divergence architecture might include explicit random seeds, private information partitions, calibrated sampling, independent retrieval corpora, causal debiasing, adversarial assignment, or post-generation portfolio optimization. The paper does not show these cannot work. It shows that common surface-level fixes do not automatically solve the problem.

Finally, the models tested are from a particular technological moment: November 2025 for the main data collection, with later supplemental exclusions where model availability changed. The broad mechanism is likely to remain relevant, but exact model rankings should not be treated as permanent personality traits. Models do not have personalities. Product teams keep trying to give them some, which is touching.

The management lesson: design for the direction of coordination

The paper’s central lesson is not “LLMs are homogeneous.” That is the shallow version. The sharper version is: LLMs can regulate similarity, but they regulate it asymmetrically. They are better at converging than at maintaining independent divergence.

That makes AI-agent design a coordination problem.

If the business process rewards sameness, exploit it carefully. Use shared prompts, shared retrieval, canonical schemas, and deterministic decoding. Then monitor correlated errors because sameness can still fail loudly.

If the process rewards difference, do not assume that five agents, five personas, or five vendors create five independent views. Build divergence into the architecture. Separate information sources. Use explicit randomization where appropriate. Assign agents different objective functions. Measure overlap. Penalize redundant recommendations. Evaluate the portfolio, not the prettiest individual response.

The monoculture trap is not that AI coordinates badly. It is that AI may coordinate too well in places where the business quietly needed disagreement.

That is a more subtle problem than model accuracy, and therefore more likely to be missed in a demo. Demos love convergence. Operations often need controlled friction.

A good AI workflow should know the difference.

Cognaptus: Automate the Present, Incubate the Future.