Research opportunity · Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?
Does language-model use narrow what whole fields study, including the social sciences?
Question
Since language models became widely used in 2023, has the range of topics covered by papers from adopting researchers shrunk relative to non-adopters, at the level of whole fields, and does the answer differ between the natural and the social sciences?
Not yet measuredA PhD projectLow to medium costPartly addressed
Why it matters
The strongest field-level evidence of narrowing comes from AI used as a research method in six natural sciences, mostly before chat assistants, and its authors exclude the social sciences. Studies of the language-model era find adopters drawing on more diverse prior work and scientists in general spreading across more fields, but they measure individuals, not the space a field covers. Without the field-level measure for language-model use, the debate cannot say whether the narrowing position (P1) or the concentration position (P2) describes the present.
The debate lists this among what we still do not know: Does language-model use narrow what fields study, measured as section 1 measured AI methods in the natural sciences?
Evidence that motivates it
- In the same natural-science corpus, AI-augmented research was associated with a 4.63% smaller median collective knowledge extent and 22% less follow-on engagement among papers citing the same work, while each AI paper’s own family of citing papers was 3.46% more spread out; the authors link the contraction to AI work clustering where data are abundant.
AI-augmented natural-science research covered a 4.63% smaller knowledge extent, associated with data availability; the authors exclude the social sciences. - Across arXiv, bioRxiv and SSRN preprints (2018–2024), authors whose abstracts first showed signs of language-model writing then posted 36.2%, 52.9% and 59.8% more papers respectively; for their manuscripts, more complex writing went with lower odds of publication, the reverse of the pattern for human-written ones; and adopters drew on “more diverse prior work”, being 11.9% more likely to cite books and citing younger documents.
Language-model adopters posted more preprints, were more likely to cite books and cited younger documents. - Among 775,323 scientists with full-text papers in PubMed Central, individual research portfolios spread across more fields after 2022: field entropy rose from 0.55 in 2022 to 0.71 in 2025 and the share entering a new field from 30.4% to 35.5%, more so for scientists in non-English-speaking low- and middle-income countries; the authors call the design descriptive.
Individual portfolios spread across more fields after 2022, in a descriptive design. - By February 2024 an estimated 17.5% of sentences in computer-science abstracts on arXiv were substantially modified by language models; the share was higher in papers closer to their nearest peer, a proxy for more crowded research areas (22.2% against 14.7% for less similar papers).
Language-model-modified sentences were more common in crowded research areas.
A possible design
Classify papers in open full-text and abstract corpora (arXiv, bioRxiv, SSRN, PubMed Central and social-science repositories) by estimated language-model use from 2018 on, using a distribution-level estimator rather than per-paper detection. Apply the knowledge-extent and follow-on-engagement measures of the natural-science study to adopters and non-adopters within field and year, adding economics, sociology, psychology and political science. Exploit differences in the timing of institutional access to language models (for example national blocks or licence roll-outs) for a difference-in-differences design, and test whether data availability explains any contraction.
Data needed
Open abstract and full-text corpora with field codes; a validated distribution-level estimator of language-model text; dates of institutional or national access to chat assistants; text embeddings frozen at one version.
Feasibility
A PhD project. Low to medium; the corpora are public and the measures need modest compute, but validating the estimator for social-science writing needs hand-labelled samples. Per-paper detection is unreliable, so adoption must be estimated at group level; dating adoption by first flagged paper creates artefacts (LD09-SL-C06) and should be avoided.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 15 October 2026; not yet checked in depth).
No study was found that applies field-level breadth measures to language-model use or to the social sciences. Nearest work measures article-level novelty after 2022 in one field without observing model use, the distinctiveness of individual grant proposals, the prevalence of model-modified text by field, or individual scientists' field-switching.
How strong this evidence is
- Who searched and checked: .
- What was read: 4 of the 5 nearest works below were judged from an abstract, summary or search listing only; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search (3 queries) and arXiv queries on the date shown; Scopus, Web of Science and Dimensions were not searched, and science-of-science conference proceedings were not checked. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Qian, Wen, Furnas, Bai, Shao & Wang, "The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding," arXiv:2601.15485 (2026); published as PNAS 123(33), e2601439123 (2026), per the Crossref record Judged from an abstract, summary or search listing only
Abstract read. Higher language-model involvement in NSF and NIH proposals and awards is associated with lower semantic distinctiveness from recently funded work; measures proposals, not the topical breadth of a field's publications. - Kousha & Thelwall, "How much are LLMs changing the language of academic papers after ChatGPT? A multi-database and full text analysis," arXiv:2509.09596 (2025) Judged from an abstract, summary or search listing only
Abstract read. Tracks model-associated vocabulary across databases, with much higher growth in STEM than in the social sciences; vocabulary, not topics. - Cao, "Visualizing the adoption of large language models across sociology subfields," Socius (2026), doi 10.1177/23780231261460657 Judged from an abstract, summary or search listing only
Seen in search results only; the article was not read. Reported to estimate model-modified sentences in sociology abstracts by subfield; adoption, not narrowing. - Zheng et al., PubMed Central study of field-switching (LD09-SL-C07)
Full text read for this debate. Individual portfolios, descriptive; no field-level breadth measure. - Safari & Babaei, "Research Novelty in Information Systems Journals After ChatGPT: Differences Across Institutional Language Contexts," arXiv:2603.22510 (2026) Judged from an abstract, summary or search listing only
Abstract read. 13,847 articles in 44 information-systems journals, 2020–2025; relative semantic novelty fell more after 2022 for first authors in non-English-dominant countries; article-level novelty in one social-science-adjacent field, with individual model use not observed, not field-level breadth.
Search queries used
large language models ChatGPT effect on topic diversity of scientific research field-level narrowing 2025 2026 studyLLM adoption social science papers topic concentration narrowing bibliometric 2026 arXiv "social sciences"PNAS 2026 LLM involvement research proposals NSF NIH less semantically distinct funded arXivarXiv: abs:"large language models" AND abs:science AND abs:diversity AND abs:topics
Provenance
RO-LD09-01 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-09 debate on 2026-10-15. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.