The Model Monoculture.
A monoculture does not look dangerous at first. It looks efficient.
One crop. One method. One supply chain. One optimised system. The field is easier to manage, easier to scale, easier to predict. For a while, monoculture looks like progress. Then disease arrives. The same sameness that made the field efficient makes it fragile. Every plant shares the same weakness. The failure does not stay local. It spreads.
That is the risk AI is bringing to thought. Not because AI makes everyone stupid, which is too crude. The deeper danger is that AI makes everyone slightly more similar.
The same few models may help millions of people write, code, research, study, design, plan, argue, summarise, sell, hire, invest, and decide. They will not produce identical outputs every time. But they may push human work toward the same defaults: the same tone, the same assumptions, the same safe framing, the same reasonable answer, the same median strategy, the same polished-but-empty prose, the same code patterns, the same blind spots.
A hallucination is a local failure. A shared blind spot is a systemic one.
That is the Model Monoculture.
What happens when millions of people think, write, code, research, and decide through the same few AI models? The danger is not only wrong answers. It is correlated cognition.
I. The old fear was wrong answers. The new fear is same answers.
Most AI criticism still focuses on accuracy. Did the model hallucinate? Did it cite fake sources? Did it misunderstand the question? Did it produce biased output? Did it fail the benchmark? Those questions matter, but they miss something deeper.
The scariest failure may not be that the model gives one person a bad answer. It is that the model gives millions of people the same kind of answer. Not exactly the same words, not the same paragraph. The same shape of thought. The same way of making a business plan. The same way of explaining a political issue. The same way of structuring code. The same way of responding to criticism. The same way of sounding intelligent.
The model does not need to be wrong to create this problem. It only needs to be useful enough that people keep accepting its defaults. The dangerous output is not always false. Sometimes it is just average. And average, distributed at planetary scale, becomes a force.
II. The formal version: algorithmic monoculture
There is a formal version of this argument. Jon Kleinberg and Manish Raghavan studied what happens when many decision-makers rely on the same algorithm.1 Their result matters because they show that monoculture can reduce the overall quality of decisions even when the shared algorithm is more accurate for any one decision-maker in isolation. The problem is not only that a shared algorithm might face a rare shock. It can hurt the system under normal conditions because everyone’s errors become correlated.
That is the key word: correlated.
Human judgment is noisy. Different people make different mistakes. That diversity is frustrating, but it also has value. When one person misses something, another may see it. When one institution overweights a signal, another may discount it. When one style of thinking fails, another style may survive. Monoculture removes some noise. It can also remove some diversity.
There is a real counterargument. Algorithmic monoculture can sometimes improve consistency, reduce arbitrary variation, and outperform messy human diversity, especially in tasks where the human baseline is already noisy and biased. That challenge matters. But large language models are not just decision algorithms. They do not only choose between options. They generate the options. They generate the language around the options. They generate the reasons people use to justify the options.
That makes model monoculture deeper than algorithmic monoculture. It is not just the judge using the same scoring tool. It is the lawyer, the applicant, the consultant, the student, the manager, and the researcher all using the same machine to decide what is worth saying in the first place.
III. AI does not only answer. It standardises.
A search engine retrieves. A recommendation system ranks. A language model composes. That difference matters.
Google shapes what you find. TikTok shapes what you watch. Instagram shapes what you desire. But AI assistants increasingly shape the sentence before it exists, the plan before it is chosen, the argument before it is made. This is why the model monoculture is not only a media problem. It is a production problem. The model enters upstream. Before the email is sent. Before the essay is written. Before the code is committed. Before the strategy deck is made. Before the research question is framed. Before the job applicant explains themselves. Before the founder knows how to pitch.
The model becomes the first draft of thought.
The first draft matters. It anchors the rest. Even when the human edits, the model has already narrowed the path. Bommasani and colleagues at Stanford CRFM made the underlying point years ago: foundation models concentrate enormous influence into a small number of artefacts that get adapted into thousands of downstream tasks, and the homogenisation that follows is a feature of the architecture, not a side effect.2 A few base models become the substrate for almost everyone else’s applications. The same biases, blind spots, and stylistic defaults propagate through every system built on top.
IV. The writing monoculture
Writing is where the monoculture becomes easiest to see. A 2024 controlled experiment by Dhruv Agarwal, Mor Naaman, and Aditya Vashistha found that AI writing suggestions pushed Indian participants toward more Western writing styles, reducing cultural nuance.3 The study involved participants in India and the United States doing culturally grounded writing tasks with and without AI suggestions. The result was not just faster writing. It was stylistic movement toward the model’s dominant norms.
That is not a minor issue. Writing is not only communication. It carries rhythm, hierarchy, politeness, directness, humour, emotion, and worldview. If AI writing assistants quietly normalise a narrow style of professional English, the loss is not only aesthetic. It is epistemic. Different ways of writing often carry different ways of noticing.
The model does not simply fix grammar. It makes expression safer, smoother, more neutral, more legible, and less strange. That may be useful in a corporate email. It is dangerous if it becomes the default texture of public thought. The future may not be filled with bad writing. It may be filled with competent writing that sounds like nobody in particular.
V. The same pattern in code, strategy, and research
Software has the same shape, with different stakes. Coding agents are moving from novelty into workflow: agent-assisted pull requests, AI-generated tests, refactors, documentation, bug fixes. That is impressive. It also raises a question: what happens when thousands of teams ask the same models to solve similar problems? They may get similar architecture, similar libraries, similar error handling, similar auth patterns, similar tests, similar security mistakes, similar assumptions about what clean code looks like. If one team makes an unusual mistake, it is a bug. If thousands of teams inherit the same pattern from the same few coding agents, it becomes infrastructure risk. Software monoculture already exists in operating systems, cloud providers, package ecosystems, and frameworks. AI coding agents may add another layer: not only shared dependencies, but shared implementation instincts. The model becomes an invisible senior engineer whispering similar advice into every repository, and that advice is often good, which is precisely why it spreads.
Strategy follows the same logic. Ask a model for a startup idea and it will often produce something sensible. That is the problem. It will tell you to identify a niche, validate pain points, build an MVP, talk to customers, focus on distribution, create a wedge, find a moat, price based on value, and avoid building before selling. All good advice. But the more people use the same models to think about business, the more everyone receives a polished version of the same strategic common sense. The result is not stupidity. It is convergence. Everyone’s pitch deck gets cleaner. Everyone’s positioning sounds sharper. Everyone’s contrarian insight starts to sound strangely familiar.
A world where everyone can reach 7/10 quickly is powerful. It helps beginners, small teams, and outsiders. It also floods the world with competent sameness. The edge moves from being able to produce a decent answer to being able to reject the decent answer. That is why taste becomes more valuable. Not taste as decoration. Taste as resistance to the default.
The same risk reaches into science. Adoption of foundation models in research is growing rapidly across domains, with the same few widely available models doing much of the heavy lifting in literature search, drafting, code, and analysis. Research monoculture does not mean everyone publishes the same paper. It means the field begins to agree too quickly on what counts as progress. The same benchmarks. The same model families. The same leaderboard logic. The same hot directions. The same overlooked questions. A science that thinks with the same tools may become faster. It may also become easier to steer.
VI. The compute layer makes it physical
This is not only a cultural issue. It is physical. Frontier AI is expensive. It requires chips, data centres, power, talent, data pipelines, and distribution. The same-few-models problem is not accidental. It follows from economics.
The models most people use are not simply the best ideas. They are the ideas backed by enough compute, capital, distribution, data, and infrastructure to become default. Even if many models exist, only a few become embedded into operating systems, browsers, office software, phones, IDEs, search engines, enterprise workflows, and education platforms. The question is not whether alternatives exist. It is which models become the default layer through which people think.
This is what makes the monoculture sticky. It is not a fashion. It is infrastructure.
VII. Model collapse: the synthetic-data version
There is another kind of monoculture: models trained on the outputs of other models. The model-collapse literature warns that when future models are trained recursively on AI-generated content, diversity and fidelity can degrade over time. The most cited demonstration, by Shumailov and colleagues, found that successive generations of models trained on previous generations’ outputs eventually forget the tails of the original data distribution, with rare events and minority modes disappearing first.4
This is the data version of the same problem. Human culture produces strange, uneven, local, contradictory, rare material. AI systems tend to compress that material into more probable patterns. If those compressed patterns flood the internet and then become future training data, the long tail of human expression becomes easier to lose.
The danger is not only that the internet fills with slop. It is that the slop becomes the next model’s idea of reality. A society can survive bad content. It is harder to survive a feedback loop where bad content becomes training data, training data becomes model output, and model output becomes the default language of the next generation.
VIII. Correlated cognition and the case for cognitive biodiversity
The model monoculture is dangerous because errors become correlated. When one student lets AI flatten their essay, that is a local loss. When an entire education system teaches students to produce the same polished neutrality, that is a cultural one. When one founder uses AI to write a generic pitch, that is harmless. When thousands of founders use AI to discover the same contrarian market insight, competition becomes theatre. When one engineer accepts a bad AI-generated pattern, that is a bug. When whole ecosystems absorb similar generated patterns, the bug becomes systemic. When one researcher misses an obscure counterexample, that is an individual oversight. When everyone uses the same systems to map the frontier of knowledge, the blind spots become collective.
The case against monoculture can be overstated, and it is worth saying so. Shared models can raise the floor. They can help a student write better, a migrant navigate bureaucracy, a small business build software, a junior developer understand code, a patient parse medical terminology, a researcher explore literature, a founder test ideas. Diversity is not automatically wisdom. Human judgment can be biased, inconsistent, unfair, lazy, emotional, parochial, or simply wrong. In hiring, lending, medicine, law, and education, one consistent model may sometimes be better than thousands of arbitrary human decisions. The problem is not that shared models are always bad. It is that shared models become dangerous when people forget what diversity was doing.
A legal form should be standardised. A society’s imagination should not be.
The answer is not to stop using AI. That would be unserious. The answer is cognitive biodiversity. Use AI, but do not let one model become the first and last voice. Use multiple models, compare their disagreements, preserve your own first draft before asking for help, ask what the model did not consider, build teams where people bring different backgrounds, disciplines, languages, and local knowledge. Protect weirdness. Reward original framing. Teach students to think before prompting. Treat model output as a draft, not a worldview.
For organisations the question becomes operational. Are all teams using the same model for the same decisions? Are all analysts using the same AI-generated research summaries? Are all engineers accepting similar generated code patterns? Are all sales teams using the same AI-written outreach logic? Are all hiring managers using the same AI screening assumptions? Are all executives receiving strategy memos that passed through the same model? If yes, the organisation may be gaining efficiency while losing internal diversity. The strongest teams will not be the ones that reject AI. They will be the ones that design disagreement back into the system.
AI will make people faster. It will make beginners stronger. It will raise the floor for millions. That is real. But a raised floor is not the same as a higher ceiling. The future will not only be shaped by who has access to intelligence. It will be shaped by whether that intelligence makes people more independent or more alike.
The danger is not that AI makes everyone stupid. The danger is that it makes everyone fluent in the same borrowed intelligence. A monoculture field looks healthy until the disease arrives. A monoculture mind looks productive until the blind spot matters.
1 Kleinberg & Raghavan (2021). Algorithmic Monoculture and Social Welfare. Formalises the result that many decision-makers relying on the same algorithm can reduce aggregate decision quality, even when the shared algorithm is more accurate for any individual user, because errors across the system become correlated.
2 Bommasani et al., Stanford CRFM (2021). On the Opportunities and Risks of Foundation Models. Argues that the rise of a small number of base models adapted across thousands of downstream tasks creates an inherent risk of homogenisation: the same flaws, biases and stylistic defaults propagate everywhere the base model is reused.
3 Agarwal, Naaman & Vashistha (2024). AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Controlled experiment with participants in India and the United States: when AI writing suggestions were enabled, Indian participants’ writing shifted measurably toward dominant Western norms, reducing cultural nuance in the output.
4 Shumailov, Shumaylov, Zhao, Gal, Papernot & Anderson (2024). AI models collapse when trained on recursively generated data, Nature 631, 755–759. Shows that successive generations of models trained on data produced by previous generations forget the tails of the original distribution, with rare events and minority modes disappearing first.
This is Essay No. 008. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.