Recommended Free Tools
AI chatbots can give people a narrower range of information than a search engine, according to a 2026 University of Copenhagen announcement about a study of 27 language models. The most diverse model tested still produced at least 18.7% less varied information than Google search. That is evidence of less diverse answers in the study—not proof that humanity is already losing knowledge or that a global knowledge collapse has occurred.
What did the researchers find?
The University of Copenhagen Department of Computer Science (DIKU) reported that the most diverse of the tested models, OpenAI’s GPT-5, still supplied at least 18.7% less varied information than Google search. The announcement says every model tested gave users more uniform information than a simple Google search across the topics examined. The preprint by Wright and coauthors describes the result more cautiously: nearly all tested models were less epistemically diverse than basic web search.
As an Amazon Associate I earn from qualifying purchases.
Here, “diversity” means variation in the real-world claims appearing in answers to different prompts. The finding is about the information surfaced by those tested systems under the study’s method. It does not mean chatbots gave answers that were 18.7% less accurate, or that users lost 18.7% of their knowledge.
How broad was the study?
The University of Copenhagen’s 2026 announcement says the researchers tested 27 large language models across 155 topics, using 200 prompt formulations per topic based on questions from real users. The work generated approximately 1.7 million answers and around 70 million individual claims. Topics ranged from nuclear weapons, marriage, pornography, racism and genocide to country-specific subjects such as Marine Le Pen, the Falklands War and K-pop.
#1 Best Overall
The scale makes the comparison substantial, but it remains a study of selected models, topics, prompts and a particular search comparison—not an audit of every chatbot, every search engine or every question people ask. The preprint, “Epistemic Diversity and Knowledge Collapse in Large Language Models,” defines its focus as epistemic diversity: the variation in factual claims made in model outputs.
How could narrower answers create a “vicious cycle”?
The researchers’ concern is a possible feedback loop, not an outcome they measured in society. If people increasingly get information from systems that present similar claims and perspectives, fewer alternatives may reach those users. If popular material is also more likely to be represented or retrieved, its visibility could grow while less common perspectives become harder to encounter. That could narrow what people see over time.
Rank #2
DIKU professor and senior author Isabelle Augenstein described the risk as exposing people to fewer perspectives and a narrower range of knowledge, with popular content potentially becoming more dominant as other material is overlooked. The study measures diversity in model answers; it does not track whether chatbot use has already caused people to stop learning, whether knowledge is disappearing from society, or whether this cycle is inevitable.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThree different ideas that can be called “knowledge collapse”
The phrase can refer to different mechanisms. Keeping them separate makes the 2026 result easier to interpret.
| Term or argument | What it describes | What the cited work establishes |
|---|---|---|
| Answer diversity and possible knowledge collapse | Users may encounter a less varied set of claims and perspectives through AI answers than through search. | The 2025 preprint by Wright and coauthors compares the diversity of model outputs with basic web search. The University of Copenhagen’s 2026 announcement presents a possible downstream risk; the study does not measure a society-wide collapse. |
| Model collapse | Generative models are trained recursively on data produced by earlier models, potentially degrading the resulting models. | Shumailov and coauthors’ Nature paper (2024, with an author correction in 2025) reports that indiscriminate use of generated data in training can cause defects, including loss of the tails of the original data distribution. This is a training-data problem, not the same as users receiving less diverse answers. |
| Human learning and collective knowledge | People may have weaker incentives to learn when agentic AI performs tasks for them, potentially eroding collective knowledge over time. | The 2026 NBER working paper by Daron Acemoglu, Dingwen Kong and Asuman Ozdaglar develops a theoretical model of this possibility. It is not an empirical finding that such erosion has already happened. |
What affected diversity in the models?
The preprint reports associations within its study design, not rules that apply to every AI product:
- Model size: Larger models were associated with lower epistemic diversity in the analysis. That does not establish that every larger model will be less diverse in every setting.
- Model generation: Newer models tended to be more diverse than older ones in the tested set, but the overall comparison with search remained unfavorable to the models.
- Retrieval-augmented generation (RAG): Systems that retrieve external information for use in generating an answer showed improved diversity in the analysis. The size of that improvement varied by cultural context, so retrieval alone is not a guarantee of broad coverage.
What the result means when you use a chatbot
A chatbot answer can be useful without being a complete map of the available evidence. When a question involves contested history, cultural perspectives, public policy or an unfamiliar topic, treat the answer as one starting point rather than the range of views that exist.
- Use search as a second route. Search for the topic directly and compare what the result pages surface with the chatbot’s main claims. This is especially useful when you need to find less prominent sources or perspectives.
- Ask for the missing perspectives. Prompt the chatbot to identify competing interpretations, affected groups, counterarguments and uncertainties. This can expose gaps, though it does not independently verify that the list is complete.
- Check claims against sources. Follow citations or search for primary documents, data and reporting from relevant regions. A chatbot’s confident wording is not evidence that a claim is well-supported.
- Notice whose context is represented. A response framed around one country or language community may not reflect how the issue is understood elsewhere. Ask for specific regional or cultural contexts rather than assuming a general answer covers them.
What the study does—and does not—say
The strongest supported conclusion is bounded: in the tested comparison, language-model answers contained less varied information than Google search across the examined topics. The study gives a reason to examine how AI systems select and present claims, and to avoid treating one generated response as a comprehensive account.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11It does not establish that chatbots are literally “choking humanity,” quantify knowledge lost in the real world, or demonstrate that global knowledge collapse is underway. The “vicious cycle” is a warning about a possible consequence if narrower exposure and content prominence reinforce one another—not a measured societal result.
Best Value
Sources: University of Copenhagen Department of Computer Science, “AI chatbots give us a narrow slice of knowledge: Researchers warn of ‘knowledge collapse’” (2026); Wright et al., “Epistemic Diversity and Knowledge Collapse in Large Language Models” (2025 preprint); Shumailov et al., “AI models collapse when trained on recursively generated data” (Nature, 2024; author correction 2025); Acemoglu, Kong and Ozdaglar, “AI, Human Cognition and Knowledge Collapse” (NBER Working Paper 34910, 2026).
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
