---
title: "The AI paradox: More successful scientists, narrower science"
author: Martin Ebner
date: 2026-11-19
url: https://postgenerative.at/indications/2026-11-19/the-ai-paradox-more-successful-scientists-narrower-science
lang: en
type: Indication
summary: "A Nature study reveals a paradox: AI use is associated with more publications, citations and faster careers, while collective research focus contracts. Individual success does not automatically translate into broader scientific exploration."
tags:
  - Research
  - Science
sources:
  - title: "Hao, Q., Xu, F., Li, Y. & Evans, J. (2026): Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature 649, 1237–1243. DOI: 10.1038/s41586-025-09922-y"
    url: https://www.nature.com/articles/s41586-025-09922-y
  - title: "Hao et al. (2026): Abbildung 3 / Figure 3 – thematische Reichweite und Konzentration der Forschung"
    url: https://www.nature.com/articles/s41586-025-09922-y/figures/3
  - title: "Hao et al.: frei zugängliche akzeptierte Manuskriptfassung, arXiv v4 vom 29.11.2025 (CC BY 4.0)"
    url: https://arxiv.org/html/2412.07727v4
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
alternate:
  de: https://postgenerativ.at/indizien/2026-11-19/das-ki-paradoxon-erfolgreichere-forschende-engere
---

# The AI paradox: More successful scientists, narrower science

AI can expand individual scientists’ influence while narrowing the knowledge space explored collectively. This **paradox between individual gains and collective progress** is central to the study by Qianyue Hao, Fengli Xu, Yong Li and James Evans, published in *Nature* on 14 January 2026. [1]

### Greater individual success

The result highlighted in the German version translates as follows: On average, scientists using AI published 3.02 times more papers, received 4.84 times more citations and became research project leaders 1.37 years earlier. Source: Hao et al., *Nature* 649, 1237–1243 (2026), abstract and discussion. [1][3]

**How to read these figures:** They compare researchers with and without identified AI adoption. They should not be treated as conclusive proof that AI alone caused the differences. The authors explicitly acknowledge that they cannot fully identify the causal linkage. Earlier career transition is modelled; “team leadership” refers to the study’s operationalized transition to leading a research project. [3]

### Why collective exploration can still contract

The analysis covers approximately **41.3 million natural-science papers from 1980 to 2025**, across biology, medicine, chemistry, physics, materials science and geology. A trained language model identifies papers using AI research methods from their titles and abstracts. This concerns much more than current chatbots or AI writing assistance. [3]

AI-augmented research concentrates particularly on data-rich problems. Individual researchers become more productive and visible while many focus on the same well-established areas. The measured median collective knowledge extent is **4.63 percent smaller** for AI research, and measured follow-on engagement is **22 percent lower**. Engagement here means how frequently papers citing the same original work also cite one another, rather than a general reduction in personal collaboration. [3]

The authors suggest that individual incentives can encourage collective optimization of known problems rather than exploration of new questions. More papers and attention can therefore coexist with greater concentration. Data-poor, potentially foundational questions risk being left behind. [3]

### Figure 3: What is contracting?

[View the original Figure 3](https://www.nature.com/articles/s41586-025-09922-y/figures/3)

**Figure attribution:** Hao et al., *Nature* (2026), Fig. 3. The figure is described and linked here. [2]

- **Panel a** explains the method: Titles and abstracts are embedded in a 768-dimensional semantic space.
- **Panel b** visualizes 10,000 papers, half AI and half non-AI. AI papers appear more clustered; their red-marked extent is smaller than the blue-marked comparison group’s.
- **Panels c and d** compare knowledge extent and topic distribution across the six disciplines. AI research has lower extent and entropy, indicating stronger concentration. [2][3]

The two-dimensional map is a visualization; extent is calculated in the original high-dimensional space. A narrower topic space is a text-based measure and does not establish that every AI paper is less original or scientifically weaker. [2][3]

### Implications

The authors advocate deploying AI to expand sensing, data gathering and experimentation in previously inaccessible domains. It should enable new observations rather than merely analyse existing data more efficiently. The study is observational, focuses on six natural sciences and can miss unmentioned AI use. Its findings on recent generative AI are more preliminary than its evidence spanning decades of older methods. [3]
