The One Oracle: What a Civilization Loses When a Billion People Ask the Same AI

THE THESIS. The danger of the one oracle is not that it lies: it is that it averages. We have already analyzed what AI does to each user's brain; today we are talking about something else. We are talking about what it does to everyone's thinking at once. Our thesis is simple: when hundreds of millions of people consult three or four models trained on similar data and tuned with similar methods, the answers converge toward the center of the distribution. And a civilization that converges loses its tails: the odd ideas, the useful dissent and the minority knowledge that are the raw material of all innovation. Let us be clear from the start: this is not deliberate censorship or a conspiracy. It is a measurable statistical effect, documented in peer-reviewed studies. And precisely because it can be measured, it can be governed.

THE EVIDENCE. The most cited experiment is by Anil Doshi (UCL) and Oliver Hauser (Exeter), published in Science Advances: 293 people wrote short stories, some unaided and some with AI-generated ideas, judged by 600 evaluators. The result cuts both ways. AI improved individual creativity, especially for the least creative writers. But the AI-assisted stories were more similar to each other than the unaided ones. The authors call it a social dilemma: each person writes better, and the collective produces less variety. A 2026 review in Trends in Cognitive Sciences by Sourati, Ziabari and Dehghani (University of Southern California) widens the picture: homogenization shows up in style, in perspectives and in reasoning patterns, models align with a Western, university-educated worldview (what the literature calls WEIRD bias), and the effect persists even when you raise the model's 'temperature' or ask it to role-play different personas. Another recent study measured 'epistemic diversity' — how many distinct sources and perspectives appear in answers — and found that chatbots offer less diversity than a traditional web search, and that making models bigger does not fix it. The theory predates the chatbots: Kleinberg and Raghavan showed in 2021 that when many decision-makers use the same algorithm, collective welfare can fall even if that algorithm is the best one available, because errors become correlated. Bommasani and colleagues verified it in 2022: when several systems share the same base model, the same people get rejected by all of them at once.

WHY IT CONVERGES. There are three mechanisms, and none of them is malicious. First, the basic mechanics: a language model predicts the most probable next word, so it naturally drifts toward the most frequent — toward the center of what it has read. Andrew Peterson (University of Poitiers) calls the consequence 'knowledge collapse': if the average answer is nearly free, fewer and fewer people pay the cost of visiting rare sources, and the tails of knowledge lose their visitors. In his simulation, AI-generated content being just 20% cheaper than seeking original sources was enough to leave public beliefs 2.3 times further from the truth. Second, the human tuning. RLHF — the process where humans rate model answers to teach it what we prefer — rewards what we like, not necessarily what is true. Anthropic itself documented that both human raters and preference models choose the well-written, flattering answer over the correct one a non-negligible fraction of the time. Sycophancy — the model's tendency to agree with you — is not an occasional bug: it is friendly consensus turned into a product. Third, the feedback loop. A Cornell experiment with 1,506 participants showed that writing with a biased AI assistant shifts the user's opinion, not just their text: they called it 'latent persuasion.' And those influenced texts go back onto the internet, which is the training material for the next generation of models. The average feeds on itself.

THE PRECEDENT. Humanity has been here before, but never exactly like this. The printing press standardized spelling and canonical texts, yet it multiplied diversity, because every press was a new voice: without it there is no Reformation and no modern science. The Encyclopédie tried to put all knowledge in a single work, but it was a starting point: it forced you to keep reading elsewhere. Google concentrated the doorway to knowledge, but it pointed outward: its results page sent you to millions of different websites. The chatbot breaks that pattern. It does not point you to the sources: it replaces them with a single synthesis. And the doorway is extremely narrow. OpenAI claims more than 800 million weekly ChatGPT users, and traffic aggregators based on Similarweb estimate that in May 2026 ChatGPT held around 54% of chatbot web traffic, Gemini 28% and Claude 9%: three companies account for over 90% of queries. No knowledge technology had ever concentrated not just access, but the answer itself, to this degree.

THE SERIOUS COUNTERARGUMENTS. The first is personalization: if the model adapts to each user — memory, context, the 'pluralistic alignment' that recent research explores — a billion people would not get the same answer but a billion variants. It is the most promising path, but today it is more promise than reality: the Trends in Cognitive Sciences review finds that homogenization persists even with simulated personas and temperature tweaks. The second is the explosion of open-weight models (models whose weights can be downloaded and run on your own hardware). GLM, Qwen, DeepSeek or Kimi, plus their thousands of local fine-tunes, mean the oracle is not literally unique. We welcome this — it is our long-standing thesis about democratization — but with one caveat: many models do not equal many distributions if almost all of them are trained on the same web data with similar recipes. The third is technical and hopeful: the epistemic diversity study found that retrieval-augmented generation (RAG: when the model searches external documents before answering) improves the diversity of responses. In other words, monoculture is not a law of physics: it is a design choice.

OUR READ. In the short term, the risk is real and underpriced precisely because it is invisible: nobody notices that their idea is the same as ten million other users'. That is why our proposal is to measure and publish output diversity the way we measure accuracy or safety today: it should be a standard benchmark, with incentives for labs that reward rare-but-correct answers, not just pleasant ones. Users have homework too: use AI after thinking, not before — the Doshi and Hauser study suggests order matters — and treat a chatbot's answer as a starting point that cites sources, not as a verdict. In the long term, we keep our tempered optimism, for two reasons. First: for most of humanity, the one oracle does not replace a diverse library — it replaces nothing; hundreds of millions of people are getting their first competent tutor in their own language, and that, if well designed, is a net gain in global cognitive diversity, not a loss. Second: unlike recommendation-algorithm bias, which took us fifteen years to take seriously, here the science arrived before the catastrophe: we have the metrics, the identified mechanisms and the levers (RAG, pluralism, open models, local and sovereign ecosystems). The first century of the printing press was decided by printers and readers, not by the machine. So will this one. The question is not whether AI homogenizes — it does, and it has been measured — but whether we will treat diversity of thought as what it is: critical infrastructure for civilization, as worthy of investment as computing power.