Research opportunity · Levelling or hollowing? Skill compression and the reproduction of expertise
Does handing work to the model cause both skill loss and sameness?
In short: has this been done?
Related work: Partly addressed. Related studies cover parts of this question; what they leave open is set out below. Only a first-pass search has been done, so it is not yet known whether a fuller search would find a study that does the whole design.
How far checked: First-pass screen only. A first-pass AI web search on 16 October 2026. No deeper check yet. Every search and check so far was done by AI; no person or domain expert has checked them yet.
The searches, the nearest work found and what the search could not reach
Follow this opportunity: Atom feed of RO-LD04-07 updates (new versions, and revisions of the claims it rests on).
Across debates: it also draws on Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?, and is listed on each of their pages.
Question
When people learning a writing or analysis task are randomly assigned to use a language model that does the work for them, a model that only explains while they do the work, or no model, do the same participants who lose more skill once the model is removed also produce work more similar to each other's and to the model's, and does the way the model is used account for both?
Gap type: Causal mechanism not isolatedOne PhD chapter or paperLow to medium costRelated work: Partly addressed
Why it matters
LD-04 and LD-09 each point to the way people use the model, without measuring the other debate's outcome. In LD-04, learners who delegated code to an assistant scored lowest on a later quiz, and students who asked for explanations understood more than those who asked for solutions. In LD-09, the loss of diversity in co-written essays came from the text the model contributed, not from the users' own text, and assisted gains did not carry over once the model was taken away. If delegation drives both skill loss and convergence, the two debates describe one phenomenon and one safeguard (explanation instead of answers) could address both; if not, they need different safeguards.
Evidence that motivates it
- In a randomized experiment with 52 Python programmers learning an unfamiliar library, those given an AI assistant scored 17% lower on a later quiz of concepts, code reading and debugging, with no significant speed gain; participants who delegated code to the AI scored lowest.
How this study uses it: Programmers learning a new library with an AI assistant scored 17% lower on a later quiz, and those who delegated code scored lowest; the authors were Anthropic Fellows, a conflict the claim records. - In two preregistered lab experiments and a field study, LLM access had no overall effect on learning: students who used it to generate solutions covered more topics but understood each less, students who asked it for explanations understood more, and the gap between students with low and high prior knowledge widened.
How this study uses it: Students who used a model to generate solutions covered more but understood less, while those who asked for explanations understood more. - In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48% but left students 17% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127% and left exam scores indistinguishable from controls.
How this study uses it: A standard chat interface left students worse on a later closed-book exam, while a tutor version with safeguards did not. - 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.
How this study uses it: The reduction in essay diversity came from the text the model contributed; users' own text was unaffected. - 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.
How this study uses it: Assistance did not raise originality in later unassisted tasks, and those coached by the model produced more similar ideas 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.
How this study uses it: Writers given AI ideas wrote stories rated more novel but more similar to each other and to the AI ideas.
A possible design
Run a preregistered experiment over several sessions with three arms (delegation mode, explanation mode, no model) on a writing or data-analysis task new to participants. Measure, for each person, unaided performance on a held-out task at the end and some weeks later, and the similarity of their work to other participants' work and to the model's own output, with semantic and idea-level measures validated against human judgement. Log actual use to measure how much each person delegated within each arm. Test whether the mode of use, and the logged share of delegated text, predict both outcomes, and whether the two outcomes are correlated within persons.
Data needed
Participants (students or trainees); a task battery with unaided tests; interaction logs; similarity measures validated by human raters.
Feasibility
One PhD chapter or paper. Low to medium; a multi-session online or classroom experiment with participant payments and human rating of similarity. Participants in the explanation arm may still copy text, so the logged share of delegated text is needed alongside the assigned arm. The delayed test needs good retention of participants.
Proposal pack: this page and its references on one printable page, with BibTeX.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 16 October 2026; not yet checked in depth).
One small multi-session study measured essay homogeneity together with recall and ownership among people writing with a language model, and in a fourth session moved some model users to writing without a tool, which comes close to a measure after removal; several studies measure either retained learning or diversity alone. No study was found that manipulates delegation against explanation and measures both retained skill and convergence in the same participants. The search was web-based and partial.
How strong this evidence is
- Who searched and checked: no checker is recorded beyond the first-pass screen, which was part of the AI drafting (Claude (SocioLens Project Lead), AI-drafted from the LD-04 and LD-09 debates). Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 2 of the 2 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 four web searches and arXiv and Crossref records on the date shown; education and psychology databases were not searched, and this is a fast-moving area. 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
- Kosmyna and others, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", arXiv:2506.08872 (2025) Judged from an abstract, summary or search listing only
Abstract read. Fifty-four participants wrote essays with a language model, a search engine or neither; the model group's essays were more homogeneous and its members recalled their own text less well. In a fourth session, taken by 18 of the 54, model users wrote without any tool, which measures after removal. Mode of use was not manipulated and the paper is a preprint. - Barcaui, "ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention", Social Sciences & Humanities Open 12 (2025) Judged from an abstract, summary or search listing only
Seen in search results only. With about 120 students, those who studied with ChatGPT scored lower on a surprise retention test 45 days later; it did not measure diversity.
Search queries used
Kosmyna "Your Brain on ChatGPT" essay homogeneity recall ownership cognitive debtexperiment delegation versus explanation LLM use both learning outcomes and homogenization of student writing diversity same participantscognitive offloading generative AI skill retention and idea diversity jointly measured longitudinal randomizedBarcaui "ChatGPT as a cognitive crutch" Social Sciences & Humanities Open 2025
Cite this page
SocioLens project. (2026, October 16). Does handing work to the model cause both skill loss and sameness? [Research opportunity RO-LD04-07, version v1]. SocioLens. https://sociolens-web.pages.dev/opportunities/ro-ld04-07/v1/
@misc{sociolens_ro_ld04_07_v1,
author = {{SocioLens project}},
title = {Does handing work to the model cause both skill loss and sameness?},
year = {2026},
date = {2026-10-16},
howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld04-07/v1/}},
note = {Research opportunity RO-LD04-07, version v1}
}The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.
Provenance
RO-LD04-07 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-04 and LD-09 debates on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.