When Algorithms Agree Too Much: The Asymmetry of Strategic Monoculture

Introduction

There is a recurring anxiety in the literature on AI in markets: as more economic decisions get delegated to systems built on the same handful of foundation models, those systems will start behaving alike. The worry is not new in spirit. Economists have long studied herding, informational cascades, and correlated errors among human agents. What is new is the mechanism. Human herders at least bring different priors, different training, and different intuitions to a problem. Large language model (LLM) agents, by contrast, frequently share not only their priors but their actual weights.

A recent paper by Gonzalo Ballestero, Hadi Hosseini, Samarth Khanna, and Ran I. Shorrer, titled "Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games," brings empirical discipline to this concern. Rather than speculating about what homogeneous agents might do, the authors design a clean experiment that isolates two distinct forces and then runs it on both humans and 16 different LLMs. The result is one of the more useful conceptual refinements I have seen in this area, along with a finding that I think deserves more attention than the headline "LLMs are correlated" usually receives.

Two Kinds of Monoculture

The central contribution of the paper is conceptual before it is empirical. The authors distinguish between two phenomena that are usually lumped together under the term "algorithmic monoculture," a term introduced by Kleinberg and Raghavan (2021).

• Primary algorithmic monoculture is baseline action similarity. Two independent instances of the same model, asked the same open-ended question with no strategic incentive, tend to give the same answer. Ask several copies of a model to "name a letter of the alphabet" and you will see far more agreement than chance would predict.

• Strategic algorithmic monoculture is the adjustment of similarity in response to incentives. Here the agents are rewarded either for matching another agent or for differing from it, and the question is whether they can modulate their behavior accordingly.

This maps onto a well-established taxonomy from the human coordination literature, specifically Mehta, Starmer, and Sugden (1994), who distinguished primary salience (the tendency to give the same answer absent incentives), secondary salience (reasoning about what others find salient), and Schelling salience (identifying a unique focal point). The authors' contribution is to port this framework cleanly into the algorithmic setting and to design an experiment that separates the two.

The experimental design is deliberately simple, which is a virtue. Subjects, whether human or LLM, answer open-ended questions (name a letter, name a city) under one of three treatments:

• Picking: rewarded only for giving a valid answer. This isolates primary salience.
• Coordination: rewarded for matching another agent of the same type.
• Divergence: rewarded for differing from another agent of the same type.

The key metric is the agreement rate, the probability that two independent agents give the same answer. Comparing agreement rates across treatments lets the authors decompose baseline correlation from strategic response.

The Asymmetry That Matters

The headline results are clean. First, in the picking arm, where there is no incentive to match or mismatch, every one of the 16 evaluated LLMs is substantially more likely to agree with another instance of itself than humans are with each other. This confirms elevated primary algorithmic monoculture and is consistent with prior work showing homogenization in open-ended LLM generation (Wenger and Kenett, 2025; Jiang et al., 2025) and high correlation conditional on errors (Kim et al., 2025).

Second, and more interesting, LLMs do respond to incentives. They exhibit strategic monoculture: they increase agreement under the coordination treatment and decrease it under the divergence treatment. So they are not rigid. They can read the incentive structure and adjust.

But the adjustment is asymmetric. LLMs are exceptionally good at "finding each other" on a common focal response when coordination is rewarded, often outperforming humans. They are comparatively weak at "staying out of each other's way" when divergence is rewarded. Humans sustain heterogeneity more effectively when divergence pays.

This asymmetry is the finding I would emphasize most, and it has a plausible mechanical explanation that the paper points to: LLMs struggle to randomize. Van Koevering and Kleinberg (2024) documented that LLMs are poor at generating genuinely uniform random choices; they over-select certain options and exhibit predictable biases. Successful divergence in a symmetric game often requires something close to a mixed strategy, where each agent randomizes over a set of valid options so that two independent draws are unlikely to collide. If your sampling distribution is sharply peaked on a few focal answers, you will keep colliding no matter how strongly you are incentivized to differ. The same property that makes LLMs excellent focal-point coordinators, namely a concentrated output distribution, makes them poor diverging agents.

My Take: Coordinated Divergence as a Scarce Resource

I want to push on the economic implications, because I think the paper's framing understates how consequential this asymmetry could become.

In a wide class of economically important settings, the socially valuable outcome is coordinated divergence, not convergence. Consider screening. When many employers or admissions committees evaluate candidates using similar noisy signals, the same strong candidates get accepted everywhere and the same false negatives get rejected everywhere. Errors that would have been idiosyncratic and self-correcting in a heterogeneous market become systematic. Kleinberg and Raghavan (2021) made exactly this point about algorithmic hiring. The same logic applies to portfolio allocation, to price discovery, and to any setting where diversity of action makes aggregate signals more informative.

The paper's experiment suggests that the very systems we are deploying into these settings are structurally biased toward the failure mode. They converge easily and diverge poorly. If a thousand firms price using slightly fine-tuned versions of the same base model, and the model's output distribution under a "differentiate from competitors" instruction is still concentrated on a few focal price points, then the market loses the heterogeneity that keeps prices informative. As I argued in the short version of this analysis: we have spent decades worrying about correlated human herding, but algorithmic monoculture is worse because the agents share weights, not merely priors.

There is a second-order point worth raising. Humans achieve divergence partly through genuine heterogeneity of belief and partly through the social technology of conventions (you take the high road, I take the low road). LLM agents lack stable identities across instances and therefore cannot easily establish such conventions on their own. This suggests that coordinated divergence may need to be engineered externally, through assigned roles, injected randomness, or mechanism design at the platform level, rather than emerging from the agents' own reasoning.

Limitations

Several caveats temper the conclusions. The experimental tasks are abstract one-shot games over naming questions, which is exactly what makes the identification clean but also what limits external validity. Real markets involve repeated interaction, learning, and side information, all of which could either dampen or amplify the asymmetry. The agents also face each other only within type, so cross-model interaction, which is the more realistic ecology, is not tested. And "agreement rate" as a metric, while elegant, abstracts away from the magnitude of misalignment that matters in continuous decisions like pricing. Finally, the randomization deficit may be partly an artifact of inference-time settings (temperature, decoding strategy) that could be tuned, so it is worth asking how much of the asymmetry is fundamental versus an implementation choice.

Conclusion

The paper makes a clean conceptual cut between baseline similarity and strategically modulated similarity, and it delivers a memorable empirical result: LLMs are superb convergence coordinators and mediocre divergence coordinators. The asymmetry is not a curiosity. It tells us that as AI agents populate markets, the easy half of coordination will get easier and the hard, socially valuable half, the coordinated divergence that keeps signals informative, may get harder.

The open question I find most pressing is whether divergence can be restored through mechanism design rather than through the agents themselves. If individual LLM agents cannot reliably randomize or sustain heterogeneity, can platforms inject diversity at the system level, assigning correlated-error-breaking roles the way auction designers assign tie-breaking rules? The answer will determine whether AI-mediated economies inherit the informativeness of diverse human markets or collapse toward a brittle, weight-sharing monoculture.