Research opportunity · Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?
Does the convergence caused by writing with AI persist or compound over months of use?
Question
When the same people write or brainstorm with a language model repeatedly over a term or a year, does the similarity of their work to each other's grow, stay level or fade, and does it carry over into what they write without the model?
Not yet measuredA PhD projectNeeds a budget (medium or higher)Partly addressed
Why it matters
Nearly all evidence of homogenization comes from single sessions. One experiment found that creativity gains vanished once the model was taken away while coached participants' ideas stayed more alike, which suggests the collective cost could outlast the individual gain. If convergence compounds over repeated use, single-session effects understate it; if it fades, the narrowing position overstates the cost.
The debate lists this among what we still do not know: Does convergence persist or compound over months and years of use, and over generations of students trained with AI?
Evidence that motivates it
- In two preregistered experiments, language-model assistance gave short-term boosts during assisted tasks only in some measures (in the divergent-thinking task, originality during exposure was similar across conditions), did not raise originality or fluency in later unassisted tasks, and participants who had received strategy coaching from the model produced more similar ideas afterwards.
Help from a model did not raise later unassisted originality, and strategy coaching left ideas more similar afterwards. - In a preregistered online experiment, writers given one or up to five story ideas from GPT-4 wrote stories rated 5.4% and 8.1% more novel and 3.7% and 9.0% more useful than writers without AI, with the largest gains for less creative writers; but their stories were 5.2% and 5.0% more similar to the AI ideas and more similar to each other.
Stories written with AI ideas were rated better and were more similar to each other, in one session. - In argumentative essays co-written with a language model, only the instruction-tuned model (InstructGPT) made essays more alike; the base model (GPT-3) did not, and the reduction in diversity came from the text the model contributed, while the users’ own text remained unaffected.
Homogenization came from the model's contributed text, not the users' own text, within a session.
A possible design
A preregistered longitudinal experiment in writing courses. Sections are randomized to a model-assisted condition, a plural-model or diverse-prompt condition, or no model, for a term. Students complete matched writing and ideation tasks with and without assistance at baseline, mid-term, end of term and three months later. Measure pairwise similarity within section at the level of ideas (coded by blind raters) and of style (embeddings), separately, and each student's distance from model outputs.
Data needed
Access to course sections willing to be randomized, with consent; rater coding of ideas; model logs for the assisted conditions.
Feasibility
A PhD project. Medium; mainly rater time and participant incentives. Contamination across conditions is likely when students use models outside class; measure out-of-class use and analyse by intention to treat.
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 randomized panel was found that follows the same writers over months of model use and measures convergence with and without the model. Nearest work is observational population trends, single-session essay experiments, and one experiment reported to follow students for about two weeks.
How strong this evidence is
- Who searched and checked: .
- What was read: 6 of the 6 nearest works below were judged from an abstract, summary or search listing only, and 1 has details that could not be confirmed; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by two web searches and arXiv record lookups on the date shown; education databases (ERIC) and learning-sciences conferences were not searched. 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
- Lee, Borchers, Alvero, Joachims & Kizilcec, "The Digital Divide in Generative AI: Evidence from Large Language Model Use in College Admissions Essays," arXiv:2602.17791 (2026) Judged from an abstract, summary or search listing only
Abstract read. 81,663 applications from 2020 to 2024 show post-2023 convergence in surface linguistic features, largest among lower-SES applicants; repeated cross-sections, not a panel of the same writers. - Inoshita, Omura, Yamanaka, Maeda & Tsuji, "Does AI Homogenize Student Thinking? A Multi-Dimensional Analysis of Structural Convergence in AI-Augmented Essays," arXiv:2603.21228 (2026) Judged from an abstract, summary or search listing only
Abstract read. 6,875 essays across five conditions; quality gains co-occur with structural homogenization, reversed on argument depth by specific prompts; one occasion per essay. - Sourati, Karimi-Malekabadi, Ozcan, McDaniel, Ziabari, Trager et al., "The Shrinking Landscape of Linguistic Diversity in the Age of Large Language Models," arXiv:2502.11266 (2025), reported published in Nature Human Behaviour Judged from an abstract, summary or search listing only
Abstract read. Writing-complexity variance falls by 21–50% across datasets when models rewrite text; population-level and style-level, not a panel. - Susanto, Woo, Yeung, Lo-Philip & Yeung, "The Crutch or the Ceiling? How Different Generations of LLMs Shape EFL Student Writings," arXiv:2604.15460 (2026) Judged from an abstract, summary or search listing only
Abstract read. Compares student compositions assisted by models before and after ChatGPT; quality and lexical diversity, not convergence across students over time. - Christenson, de Langis, Hayati & Kang, "Effects of Varying LLM Access on Essay Writing Behavior," arXiv:2606.00250 (2026), BEA Workshop 2026 Judged from an abstract, summary or search listing only
Abstract read. A pilot with 24 students randomized to no, limited or unlimited model access for one short essay; unlimited access produced essays more similar to model output; one session. - Randomized experiment with 210 undergraduates allowed or forbidden to use AI, with a retest about a week later, NBER conference paper f236743 (authors and year not confirmed) Judged from an abstract, summary or search listing onlySome details not confirmed
Seen in search results only; the paper was not read. Reported to measure essay similarity over about two weeks, not months.
Search queries used
longitudinal study persistent homogenization writing after months of LLM use students diversity panelrandomized semester-long experiment ChatGPT writing course students essays similarity diversity follow-uparXiv id_list: 2602.17791, 2603.21228, 2502.11266, 2604.15460, 2606.00250
Provenance
RO-LD09-02 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.