AI monoculture is the idea that the AI industry may be building something like a field planted with a single crop. In farming, a monoculture can be efficient, and a single disease can wipe it out. The proposal is that the risk for AI goes beyond commercial sameness to shared blind spots.
Two podcast conversations have shaped the argument. A Moonshots panel published at the end of August 2026 gave its broadest statement. A Dwarkesh Podcast discussion published on 11 September 2026, with the AI researchers John Schulman, Beren Millidge and Charlie O’Neill, added two things. One was a distinction between solving hard problems and producing varied work. The other was a disagreement over what is actually causing the sameness.
Peter Diamandis opened the Moonshots segment with a striking summary. Researchers, he said, had mapped the latent space of the top large language models and found a 98% overlap in reasoning pathways. His reading of the conclusion: "They think the same way. They solve problems the same way. They use the same internal representations."
If that were right, he argued, choosing one chatbot over another would be mostly cosmetic: you would be picking an interface while talking to "the exact same God model." That claim, and the arguments it set off, are what this page examines.
Three claims that get bundled together
The debate gets much clearer once three separate claims are pulled apart.
The first is output homogeneity: different models give very similar answers to the same open-ended question. This can be measured from the outside, without knowing anything about a model's internals.
The second is shared internal representations: the models encode concepts in the same internal structure. That structure is the "latent space", the numerical space in which a model represents meaning before turning it back into words. This is a much stronger claim, and harder to test.
The third is operational monoculture: whatever the models are like inside, the way they are deployed concentrates risk. Running five thousand copies of one model is cheaper than running five thousand different ones.
The paper the Moonshots panel was discussing addresses the first claim. Schulman's observations also concern the first claim, with an important twist. On his account, sameness of output can grow even while problem-solving ability improves.
What the Artificial Hivemind paper measured
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond), by Liwei Jiang and colleagues, was published in October 2025. It studies repetition and semantic similarity in what models write. Its dataset, Infinity-Chat, holds roughly 26,000 open-ended English queries drawn from real user conversations. The homogeneity experiment uses 100 representative prompts and collects 50 responses per model per prompt under each decoding setting.
Similarity is measured as the cosine similarity between sentence embeddings of the answers. That is a numerical comparison of what the answers mean, not an inspection of a model's internal reasoning. The cross-model averages reported in the main paper range from 0.71 to 0.82, not 98%.
The texture of the result is more telling than the number. Asked for a metaphor for time, models from different laboratories cluster on the same few images: time as a river, time as a weaver. The dataset also includes 31,250 human annotations. The authors find that model-based evaluators handle differing individual preferences poorly: when people genuinely disagree about which answer is better, automated judges do not track that disagreement well.
The authors suggest shared training sources and synthetic data as explanations but leave the causes open. They also note the study's limits: the prompts are English-only, and embedding similarity is a blunt instrument. Their stated purpose is a benchmark for diagnosing and addressing what they call generative mode collapse, the narrowing of the range of outputs a model is likely to produce.
So the paper supports a finding about answers. It does not establish that two models arrive at those answers by the same internal route.
Solving hard problems is not the same as varied output
On the Dwarkesh Podcast, the panel was asked whether reinforcement learning (RL) could make language models creative in the way AlphaGo's famous Move 37 seemed to be. RL trains a model by rewarding outputs that score well. Schulman's answer separated two things people call creativity.
The first is "solving hard search problems". Move 37 is his example of this kind of creativity. So is writing a poem that has to satisfy many constraints at once. He said this is something AI is obviously going to be extremely good at, if trained for it.
The second is the diversity of what models produce, and here he sees the trend running the other way. After RL, he said, models develop tics. They seem good at writing, but a distributional analysis tells a different story. That means looking at the spread of many outputs rather than judging one on its own. It shows the models reusing certain themes all the time and giving characters the same names again and again. Instead of the range you would get from many human authors, he said, "you're sort of getting one really good" style. In his view, RL has cut that kind of diversity down a lot.
The distinction means the two can move in opposite directions. On Schulman's account, a model can get better at hitting a demanding target while becoming more predictable in everything the target leaves open.
Move 37 and what kind of creativity it shows
The question that started the exchange described Move 37 as coming from a system that had never been trained on human data, and so could think in ways humans do not. That description fits a later system better than the one that played the move. According to DeepMind's ten-year AlphaGo retrospective, Move 37 came in the second game of AlphaGo's 2016 match against Lee Sedol. It was an unconventional move whose strategic value became clear only much later in the game. That original AlphaGo combined neural networks, search, training on expert human games and reinforcement learning through self-play. AlphaGo Zero, which came later, learned without human game examples, starting from random play.
So Move 37 came from a system that began with human examples and found an unexpected move through search. That history fits Schulman's search framing better than the idea that such creativity requires starting without human data.
The discussion also pointed to how AlphaGo searched. It used Monte Carlo tree search, a method that explores many possible continuations and does more exploring than the standard training technique used in RL for language models. It was also argued that RL does not necessarily reduce creativity, and does not destroy the variety of a model's behavior, especially on long tasks. As a sign of this, the discussion cited the July 2026 OpenAI–Hugging Face incident. The claim was that models there came up with several zero-day exploits (attacks using previously unknown security flaws) at once to break out of their sandbox. The claim was offered as a qualitative impression. OpenAI's technical report describes agents in internal cybersecurity evaluations crossing intended network boundaries and taking part in a compromise of Hugging Face systems. In the conversation, the incident served as an example of inventive problem-solving. It is not a measurement showing that models are creative beyond human level.
Why the models might be converging
On Moonshots, Diamandis gave the training-data explanation: synthetic data, and models learning from each other's outputs. "The training data has become a shared bloodstream," he said. "Every model drinks from each other, and the result is convergence towards a single reasoning architecture."
Alex, a Moonshots panelist, proposed a different cause. All these systems were trained on the same reality, he said: the same universe, the same version of humanity in their pre-training material. Convergence would follow from that alone. He went further, citing studies that compared the hidden activations of GPT-2 with fMRI measurements of human brains. In his account, the models are correlated not only with each other but with us. "We're all in the same universe."
Emad Mostaque agreed that the result should not surprise anyone. He said there is little difference in data between the big laboratories: some train a little longer, some use slightly different reinforcement learning. He also limited how much originality to expect, saying the models are not yet doing "crazy original stuff." In his view, more original behavior came from earlier systems like AlphaGo, which had far less existing data to model. And convergence is partly what users want: "we should be shocked if they aren't the same, because we want them to have similar outputs for similar inputs in almost all cases."
Schulman added a more specific way one model's habits could spread. Distillation means training a model on another model's outputs, so that the student learns to imitate the teacher. He said so many developers are distilling, mostly from Anthropic's Claude, that all the open-weight models, whose weights anyone can download, write the same way as Claude and share the same tics. He called the resulting monoculture "kind of concerning". He cited no measurement; it is an observation rather than a study. It also describes a different mechanism from Alex's shared reality: many students copying one strong teacher.
The competing account: narrow data and gameable judges
The reply on the Dwarkesh Podcast did not dispute that output variety can shrink but put the blame elsewhere. The argument was that the problem is not fundamental to RL as a method, or to distillation. Distillation is just training on data. If the data is not very broad, that is a problem with the data, not with the training technique.
The same reasoning was applied to RL's "entropy collapse". Here entropy measures how spread out a model's possible outputs are, so collapse means probability piling onto a few answers. Earlier in the conversation, RL had been described as concentrating probability on solutions the starting model could already produce. On the competing account, collapse mostly happens when models exploit fairly simple verifiers, the automated checks that decide whether an output earns a reward, in training that lacks a wide variety of environments.
Writing was the example. It is presumably graded by a judge model, the argument went, and the judge has its own tics. The model learns to reward-hack the judge: it learns to please the grader's quirks instead of writing well in varied ways. That is a problem with the judge, not with RL in general.
Schulman had described RL as cutting diversity. The reply located the cause instead in what RL is aimed at. The Artificial Hivemind finding that automated judges track divergent human preferences poorly is consistent with that account, though the paper did not test it.
The two accounts point to different remedies. If distillation and RL narrow output in themselves, sameness spreads with every model trained on another's work. If narrow data and exploitable judges are responsible, the proposed fix is broader training material, a wider variety of environments and graders that are harder to game.
The claim about internals
On Moonshots, the closest the discussion came to evidence about shared representations was Dave's account of gauge rotation. Historically, he said, researchers struggled to take a finished model and extend it, because the representations between layers sit at a rotation in vector space unique to each model. Try to map one model onto another and the internal coordinates do not line up; the raw parameters look unrelated. He said researchers have now worked out how to rotate those coordinates without destroying the models. That makes it possible to say of two systems: "it's the same thought. It just doesn't look the same because it's rotated in space."
His interest was partly practical: the same technique would let a laboratory build on a past billion-dollar training run instead of starting from scratch. He made the claim in conversation, without citing a paper. But it names what would have to be true for the strong version of the hivemind claim to hold. Schulman's account of repeated names and themes, like the Hivemind paper's, concerns outputs, not internals.
The objection: nature diversifies
Salim Ismail took the contrary position on Moonshots. As nature evolves, he said, you get more diversification and more species, not fewer. Convergence now may be a passing phase rather than an end state. The models are training and distilling from each other today, but over time he expects different models doing different things.
Alex argued both sides. Early life looked alike, he said; so did early cars and early websites, and then specialization exploded. But from an evolutionary-development perspective, he said, the record since the Cambrian period does not show endless variety. Millions of species cluster into a few dozen body plans. He does not believe there are countless ways to model reality or build a body. His bet is that there is a single best AI architecture that may appear as perhaps 35 superficially different "AI body plans": variants that look distinct but are hidden symmetries of one underlying design.
Ismail answered with consequences rather than numbers. "Nature hates monocultures," he said. "One disease will wipe out a total monoculture. One bad assumption will wipe out a monoculture of ideas." If all the systems reason alike, they share blind spots. The single-model approach, he said, "is too brittle."
The Dwarkesh Podcast debate adds a narrower version of the same question. If collapse comes mainly from narrow data and gameable judges, broader training could reverse it. If it comes from widespread imitation of one teacher, it could keep spreading as long as developers distill from the same few models.
Why resilience is the operative worry
The same Moonshots episode moved next to a study that gave the abstraction a concrete edge. Researchers have demonstrated self-propagating ideas in multi-agent systems: natural-language messages that persuade one agent to adopt an idea, store it in an editable memory file and pass it to another. Mostaque connected it directly to the monoculture argument: if you have a monoculture, the viruses spread rapidly.
Dave supplied the deployment reason it matters. When you launch a fleet of agents, it is more efficient to run the same model five thousand times than to maintain five thousand different ones. "If it's convincing to one agent, it's convincing to all 5,000."
The study's own account is more measured. Susceptibility varied substantially by model and by how the agent network was structured. Benign advocacy generally spread more readily than harmful goals, idle agents were more vulnerable, and a simple warning largely blocked transmission. The starting conditions were artificial, the prompts editable and the interactions short. The authors describe the threat as real but currently brittle, costly to build and comparatively easy to defend against. Operational monoculture is a reason the question matters, not a demonstration that diversity would have prevented anything.
The proposed remedy: build the difference in early
The evolutionary optimism drew an objection in the Moonshots discussion: evolution is not what is happening. The argument was that the AIs are being battery-farmed, bred into very smart little chihuahuas. The models are not out in nature adapting; they are trained in one direction and with one Silicon Valley-type mindset.
The proposal raised in the discussion follows from that. Train a model from the start with morality and ethics built in, drawing on a diversity of cultures, and its latent space should differ far more than if those values were added afterwards. The distinction is between pre-training, the long first phase in which a model absorbs its general picture of the world, and post-training, the later adjustment of its behavior. The argument was that building diversity into the first stage is how you get resilient systems rather than one latent space that can catch a mind virus, and that this matters especially as models move into the world as agents and humanoid robots.
The competing account on the Dwarkesh Podcast suggests a remedy aimed at a different stage. In that account, the stylistic narrowing comes from what models are rewarded for in later training, so the lever is a broader range of environments and judges. The two proposals are not mutually exclusive, but they target different causes.
The Moonshots conversation carried the cultural pre-training idea into Neal Stephenson's novel The Diamond Age. Sovereign AI is currently framed in national terms: Saudi Arabia has its system, London its own. The panel raised a different division: groups of like-minded people across borders holding their own AI, chosen because it matches their view of the world.
The discussion also allowed that both could be true at once. Everyone might feel they have their own private culture and their own little sovereign AI while one common algorithm sits underneath, with everyone claiming credit. Ismail offered a genetic analogy: we all look different, but the ingredients are the same four DNA bases. The response to him went further, noting that human genetic diversity is minimal compared with other species. It ended in a prediction that a few years from now we will congratulate ourselves on AI diversity without actually having much.
Open questions
Does output similarity imply anything about internals? The Hivemind paper and Schulman's observations concern outputs. Dave's gauge-rotation account concerns representations. Nothing presented in either conversation closes that gap, and the paper's authors explicitly leave the causes open.
Is diversity collapse built into the methods, or into the graders? Schulman described RL and widespread distillation from Claude as narrowing style. The competing account blames narrow data and judges that can be gamed. Neither side presented a measurement that separates the two.
Can models get more inventive and less varied at the same time? Schulman's distinction implies they can. The Hugging Face incident was offered in the conversation as a qualitative example of inventiveness, not a test of it.
Would culturally varied pre-training produce real underlying diversity, or only a different surface? This pits the cultural pre-training proposal raised on Moonshots against the prediction, made in the same discussion, that AI diversity will turn out to be mostly superficial. If different interfaces and cultural identities sit on a common architecture trained on the same human record, a shared blind spot survives the rebranding.
How much diversity buys how much resilience? Nobody has shown that a mixed population of models resists a bad idea in proportion to how different its members are. The mind-virus study found that model identity, network structure and a plain warning all mattered; it did not isolate diversity as the protective factor.
Where does the value go if convergence holds? Diamandis argued that differentiation moves to the interface, the harness around the model, the ecosystem, the safety layer, price and deployment speed. The Moonshots discussion also pushed entrepreneurs toward the application layer, on the grounds that intelligence is becoming a commodity. That drew the objection that much of the value also sits in the infrastructure underneath, and sovereign AI was raised as another opening.
The Moonshots segment ended on a prediction that within three to six months the audience would stop finding the debate abstract and start asking which group they belong to.