AI chatbots have become the default way millions of people look things up, but a new University of Copenhagen study finds that every large language model tested exposes users to meaningfully less varied information than a plain web search. Across 155 topics and 27 models, even the most diverse system in the study, OpenAI's GPT-5, produced at least 18.7 percent less diverse claims than Google search results. The finding matters because low diversity in model outputs is the precondition for what researchers call "knowledge collapse," a long-term scenario in which the models we increasingly depend on for information quietly narrow what the public knows.
The Copenhagen study that measured epistemic diversity
The research, led by the Department of Computer Science (DIKU) at the University of Copenhagen, is titled "What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models." It was submitted to arXiv in October 2025 and last revised on 31 August 2026, and it has been accepted to the EMNLP 2026 main conference, which takes place in October 2026.
The study introduces a way to measure something never measured directly before: epistemic diversity, the diversity of real-world claims a model produces in its answers. The authors tested 27 large language models from four families, spanning OpenAI, Meta, Google, and Alibaba. For each of 155 topics covering 12 countries, the team used 200 different prompt formulations based on questions from real users. That produced a dataset of roughly 1.7 million responses containing about 70 million individual claims.
The first author is Dustin Wright, a former postdoctoral researcher at DIKU who is now an assistant professor at Aalborg University. The senior author is Isabelle Augenstein, a professor at DIKU. The work was conducted with researchers from Aalborg University, Stanford University, the University of Colorado Boulder, and the University of Texas at Austin.
The 18.7 percent gap between chatbots and search
The headline result is that no language model in the study matched the diversity of a conventional web search. According to reporting on the paper, the top 40 Google results per topic scored an average of 3,110 on the researchers' diversity measure, while GPT-5, the strongest performer among the models, scored 2,621. The researchers describe the 18.7 percent gap as a conservative lower-bound estimate of the true difference.
"Every language model we tested provides users with more uniform information than a simple Google search across all the topics we looked at," Wright said. "In other words, people are to a large extent exposed to the same information over and over again. So AI chatbots are not just changing how we find knowledge, but also which knowledge we have access to."
The topics were deliberately broad and sensitive, ranging from nuclear weapons, marriage, pornography, racism, and genocide to country-specific subjects such as Marine Le Pen, the Falklands War, and K-pop.
Why model answers are less diverse
The researchers trace the narrowness to how large language models work. A model compresses the vast amount of text it was trained on and learns the patterns that occur most frequently. Information that deviates from those dominant patterns gets filtered out. The result is an answer that reflects the center of the training distribution rather than its edges.
The study surfaced two counterintuitive findings that challenge standard assumptions. First, smaller models generated more diverse content than larger models within the same family, reversing the "bigger is better" intuition. Second, newer models were more diverse than older ones, and epistemic diversity has increased substantially over the past three years. RAG, or retrieval-augmented generation, also improved diversity, because grounding an answer in retrieved documents pulls in material the model would not have produced on its own.
There is also a cultural skew. The paper reports that a model's parametric knowledge, the information baked into its weights, systematically reflects English over local-language knowledge for country-specific topics. In other words, when asked about a subject tied to a non-English-speaking country, the model tends to answer from an Anglophone frame.
The older warning: model collapse
The diversity gap matters because it connects to a well-documented failure mode called model collapse. In July 2024, Nature published "AI models collapse when trained on recursively generated data" by Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal, across the University of Oxford, University of Cambridge, Imperial College London, University of Toronto, and the Vector Institute. An earlier 2023 preprint circulated under the title "The Curse of Recursion."
The paper found that "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear." Rare events, unusual phrasings, and minority patterns are sampled less with each generation, until later models converge toward a narrow, low-variety approximation of the data and lose touch with the true distribution. The authors showed the effect in large language models as well as in variational autoencoders and Gaussian mixture models.
Crucially, collapse is not inevitable. A separate statistical analysis on arXiv (2404.05090) showed that collapse cannot be avoided when training solely on synthetic data, but that mixing real and synthetic data, below a computable maximum ratio of synthetic content, lets the system avoid collapse over time. The practical rule that follows is simple: accumulate real data alongside synthetic data rather than replacing it.
The Copenhagen researchers flag the same risk in their own terms. If models are increasingly trained on text produced by other models, they would learn from outputs that are already less diverse than the human-written texts they were first trained on. Repeated across generations, the range of available information narrows. That is the mechanism the researchers call knowledge collapse.
"It's a worrying thought. However, we can see that the more recent models produce slightly more diverse answers than older models, so knowledge collapse is not happening yet. But the mechanism that could trigger it in the longer term is already there. So it is something we should be aware of," Augenstein said.
Knowledge collapse: the economics of not learning
A second, independent line of research frames the same risk from the demand side rather than the supply side. In February 2026, Daron Acemoglu, a co-recipient of the 2024 Nobel Prize in Economic Sciences, published "AI, Human Cognition and Knowledge Collapse" with Dingwen Kong and Asuman Ozdaglar as NBER Working Paper 34910. The paper was updated in May 2026 and is a theoretical working paper that has not been peer-reviewed.
Their model starts from a distinction between two types of knowledge. Making good decisions requires general, community-level knowledge plus individual, context-specific information. Humans typically acquire both at the same time when they learn about a topic, and as they learn they add to the shared stock of general knowledge. The problem is that general knowledge is not static. It requires ongoing human learning effort to sustain it, because the frontier of knowledge keeps moving.
Agentic AI changes the incentive. If an AI can hand a person a context-specific recommendation without that person doing the learning, the AI becomes a substitute for effort. The better the recommendations, the less value a person gets from investing in their own understanding. Individual learning efforts are what build collective knowledge, so AI indirectly erodes the public stock.
The model identifies a tipping point. If a community's stock of general knowledge stays above a critical threshold, continued learning maintains it. If the stock dips below that threshold, learning incentives and knowledge can fall together, creating a downward spiral that the authors call a knowledge-collapse trap. The MIT Stone Center on Inequality and Shaping the Future of Work, which published a research insight on the paper in July 2026, notes a further twist: the risk can grow as AI improves, because a very accurate model discourages learning more sharply than a mediocre one. The authors describe a "sweet spot" of AI accuracy where net benefits peak, and suggest a two-phase policy: a temporary moratorium to rebuild public knowledge above the tipping point, followed by a permanent cap on model accuracy at an optimal level.
The evidence is already on the platforms
The theoretical models matter because there is suggestive real-world evidence that generative AI is already suppressing collective knowledge-building. The MIT Stone Center points to declining activity and human engagement on knowledge-sharing platforms such as Stack Overflow and Wikipedia.
Stack Overflow is the clearest case. Reporting on the platform's data shows that questions asked fell from roughly 87,000 in March 2023 to about 58,800 in March 2024, a drop of around 32.5 percent in a single year. DevClass covered the trend in January 2026 under the headline "Dramatic drop in Stack Overflow questions as devs look elsewhere for help." The irony is sharp: Stack Overflow's accumulated human answers became a key training source for the coding assistants that now answer developers directly.
This is the concrete mechanism behind the abstraction. When developers stop asking and answering questions on a public forum because a model answers them privately, the shared, aggregated, human-produced general knowledge that the model draws on stops growing. Remove enough of that ongoing contribution, and the supply side narrows at the same time demand concentrates on a handful of models.
What developers, users, and policymakers can do
The researchers are explicit that the situation is manageable rather than doomed. "We hope AI developers will build language models in a way that preserves the breadth of knowledge available to us. We have developed a method that they can use to measure diversity in models, which could help ensure that future language models do not become less diverse," Wright said. The epistemic diversity metric is offered as a tool that labs can track the way they track benchmark accuracy.
On the engineering side, three levers stand out. Retrieval-augmented generation demonstrably improves diversity by grounding answers in varied sources. Training pipelines that accumulate real human data alongside synthetic data, rather than replacing it, avoid the collapse dynamics the Nature paper identified. And tracking output diversity as a first-class evaluation metric gives teams an early warning system that today's benchmark dashboards do not provide.
For users, the guidance is behavioral. Augenstein's advice is to treat a chatbot summary as a starting point rather than an endpoint. "AI chatbot summaries can of course be useful, that is why so many people use them. But it is still important to seek out different sources in order to understand the nuances and get a broader picture, especially for the younger generation growing up with AI. We must not become so dependent on the technology that we stop understanding and thinking for ourselves," she said.
Augenstein compared the dynamic to globalization, which brought the same products and coffee chains almost everywhere: convenient, but a reduction in diversity. The analogy captures why the trend can be invisible. A homogeneous information ecosystem does not announce itself. It simply shows everyone the same answers, and the missing perspectives are the ones nobody notices are gone.
The bottom line
The knowledge-collapse debate in 2026 has two halves that reinforce each other. On the supply side, models compress training data into a narrower range of claims than search engines surface, and training on synthetic outputs can shrink that range further across generations. On the demand side, highly capable agentic AI removes the individual learning effort that replenishes collective knowledge. Neither finding predicts imminent collapse, and both describe mechanisms rather than certainties. Yet they point to the same conclusion: the diversity of what the public knows does not maintain itself. It has to be measured, funded, and deliberately preserved.
Sources: University of Copenhagen study announcement, What and Whose Knowledge? (arXiv 2510.04226), CopeNLU EMNLP 2026 papers, AI models collapse when trained on recursively generated data (Nature), How Bad is Training on Synthetic Data? (arXiv 2404.05090), AI, Human Cognition and Knowledge Collapse (NBER 34910), MIT Stone Center research insight, Stack Overflow question decline (DevClass).