Yes. A University of Copenhagen study of 27 language models, 155 topics and 70 million claims, to be presented at EMNLP 2026 in Budapest on October 24-29, found even the most varied model, GPT-5, gave at least 18.7% less diverse information than Google. Kenneth Stanley's Why Greatness Cannot Be Planned names the cost: consensus averages away the odd stepping stones a founder needs.

A chatbot is the fastest way to find out what most people think about a decision. That is also its problem, and this week the problem got a number.

The study comes from the Department of Computer Science at the University of Copenhagen, led by Dustin Wright, now an assistant professor at Aalborg University, with Isabelle Augenstein as senior author. It will be presented at EMNLP 2026 in Budapest on October 24-29; TechXplore wrote it up on September 28, 2026. The team tested 27 language models on 155 topics covering 12 countries, with 200 prompt templates per topic drawn from real user questions, which produced 1.7 million responses and about 70 million individual claims. They grouped claims that mean the same thing and measured how many distinct claims each model produces. Every model was less diverse than a plain search baseline, and even the most varied one, OpenAI's GPT-5, gave at least 18.7% less varied information than Google.

Three findings in the paper matter more to an operator than the headline. Diversity has risen substantially over the past three years, so this is not knowledge collapse yet, and the authors say so. Retrieval-augmented generation, where the model pulls in search results before answering, has a statistically significant positive effect on diversity. And larger models were less diverse than smaller ones, which the authors suggest may be because bigger models memorise more. The model you pay most for is not automatically the one that shows you the most perspectives. There is a geography effect too: for country-specific topics, the models' built-in knowledge reflected English sources more than local-language ones, which matters for any founder selling into a market that does not write in English.

Kenneth Stanley and Joel Lehman's Why Greatness Cannot Be Planned gives the cost a name. Their argument, which grew out of AI research, is that ambitious results come from stepping stones that rarely resemble the destination, and that consensus is the enemy of finding them: it blends sharp contrasts into what they call a faded gray. A chatbot answer is a consensus engine by construction, because it compresses the most frequent patterns in its training data. For routine questions that is exactly what you want. For a strategic bet, the claim you need is often the odd one the average dropped, and the Copenhagen study now puts a lower bound on how much gets dropped.

The move this suggests is to ask a model for the disagreement before the recommendation, and for a way to leave the model and check. I ran this prompt on October 5, 2026 with Claude Opus 5.5:

Tested October 5, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

I'm a founder deciding whether to raise a seed round now or keep bootstrapping for another 12 months. We have $18k in monthly revenue growing about 8% a month and 9 months of runway.

Before you give me any recommendation:
1. State the consensus view most sources would give, in two sentences. Label it CONSENSUS.
2. Give me three positions that serious people hold that DISAGREE with that consensus, each from a different vantage point (an operator, an investor, a researcher or critic). Describe who holds each view in general terms. If you are not sure a view is actually held by real people, say so instead of inventing it.
3. For each dissenting position, give me one exact search query I can paste into a web search to check whether it holds up.
4. Only then tell me which view you lean toward and how confident you are, from 0 to 100%.

Do not blend the positions into a balanced average. Keep […]

The start of what came back, unedited:

# Seed Now vs. Bootstrap 12 More Months

**A constraint first:** With 9 months of runway, bootstrapping for 12 more months only works if your burn falls as revenue grows. At 8% monthly growth, revenue doubles roughly every 9 months, so you'd reach about **$36k MRR at month 9** and about **$45k MRR at month 12**. Whether you can bootstrap depends on one number you didn't give me: **your monthly expenses.** If they're below about $30–35k, growth alone could get you close to break-even before the cash runs out. If they're well above that, the bootstrap option partly depends on cuts.

---

## 1. CONSENSUS

Start raising now, because fundraising takes 3–6 months and you want to negotiate with 6+ months of runway, not 2. $18k MRR growing 8% monthly is real traction, so raise while it's a story of momentum rather than a rescue.

---

## 2. Three Dissenting Positions

### A. The Operator: "Get to default-alive first. Don't raise from a position you can't walk away from."
**Who holds it:** Founders of profitable bootstrapped software companies, the "calm company" crowd, and indie SaaS […]

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

What it did well: before answering, it flagged the number I had not given it, monthly burn, and showed its arithmetic (8% a month takes $18k to about $45k in twelve months, which checks out). It kept three dissenting positions separate, described who holds each, and admitted it was unsure how widespread one of them is instead of inventing support. The search queries were the most useful part, because they are the exit from the model to the wider search baseline the study measured. Its 55% confidence was honest about depending on facts it did not have.

What it did badly, measured against the study: the three dissents were the well-worn ones, bootstrapper doctrine, revenue-based financing, and Noam Wasserman's rich-versus-king trade-off. They are real positions, but they are the canon of startup advice, not the strange stepping stone Stanley means. Asking a model for dissent widens the answer inside its own distribution; it does not get you outside it. That is why step three matters more than step two: paste the queries, read the actual pages, and treat the model as a map of the mainstream rather than the territory.

The limits, plainly. The 18.7% figure measures the diversity of factual claims on broad topics, from nuclear weapons to K-pop, not decision quality, and it does not say ten search results are better reasoned than one chatbot answer. One run of one prompt is an anecdote, and the same prompt will answer differently next time. And the fix costs time: the dissent step adds minutes to every big question, which is precisely the friction the chatbot was meant to remove. Spend those minutes on the decisions you cannot easily reverse.

Sources: AI chatbots give us a narrow slice of knowledge: Researchers warn of 'knowledge collapse', TechXplore (University of Copenhagen), September 28, 2026; What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models, Wright et al., arXiv, v7 camera-ready August 31, 2026; full paper (HTML), findings on RAG, model size and local-language knowledge, arXiv, 2026; Kenneth O. Stanley and Joel Lehman, Why Greatness Cannot Be Planned: The Myth of the Objective, Springer, May 2015, chapters on deception and consensus.


Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours