Research opportunity · Levelling or hollowing? Skill compression and the reproduction of expertise
Do newcomers get less feedback and real work once AI coding tools arrive?
Question
In open-source projects, after AI coding tools are adopted, do first-time contributors receive less substantive review feedback, less mentoring and fewer tasks handed to them by experienced maintainers than in comparable projects that have not adopted them?
Not yet measuredOne PhD chapter or paperLow costPartly addressed
Why it matters
The apprenticeship-gap position rests on a mechanism, that experts stop handing novices the work and feedback through which skill is built. It is documented for surgical robots and formalised in theory, but not measured for generative AI. Open-source projects record every review comment and assignment publicly, so they allow a direct test of the feedback novices actually receive, which existing studies of newcomer counts and onboarding speed do not measure.
The debate lists this among what we still do not know: Does AI reduce the real work, feedback and supervised struggle that seniors hand to juniors?
Evidence that motivates it
- A two-year comparative ethnography of robotic and open surgery, with interviews at 13 US teaching hospitals, found that approved routes for learning failed in robotic surgery; only a minority of trainees came to competence, through tolerated but unapproved practices, and these practices led to hyperspecialisation and a shrinking supply of experts relative to demand.
In robotic surgery, approved learning routes failed when the technology removed trainees from hands-on work, and the supply of experts shrank: the mechanism this study would test for generative AI. - A theoretical model in which novices acquire tacit knowledge by working with experts shows that automating entry-level work can raise output yet lower long-run growth, even without cutting entry-level employment, by moving novices away from the most productive experts.
A model shows that automating entry-level work can lower long-run growth by moving novices away from the best experts, even without cutting entry-level jobs; who novices work with is the variable to measure. - In US résumé data covering 65 million workers, firms that began integrating generative AI saw junior employment fall by about 9% relative to senior employment after six quarters, compared with non-adopters, mainly because they hired fewer juniors, while junior promotion rates rose slightly.
US firms that adopt generative AI hire fewer juniors while promotion rates of those hired rise slightly; what happens to juniors' learning inside adopting organisations is unmeasured.
A possible design
Staggered difference-in-differences on public repositories. Date each project's adoption of AI coding tools (assistant configuration files, AI co-authored commits or pull requests, or tool-specific markers) and match adopters to non-adopters on size, language, activity and newcomer inflow before adoption. For each first-time contributor, measure the number and length of review comments received, whether comments explain (classified with a validated coding scheme, checked by hand on a sample), whether maintainers assign issues to them, and whether they return and become regular contributors. Cover both the 2021–2023 assistant wave and the 2024–2026 agent wave. Interviews with a sample of maintainers can test whether any change is deliberate.
Data needed
GitHub event data (GH Archive or the GitHub API), review-comment text, issue-assignment events; a coding scheme for explanatory feedback with inter-rater reliability; markers of AI-tool adoption, validated against a hand-checked sample.
Feasibility
One PhD chapter or paper. Low; public data and free tools; the main cost is hand-coding a validation sample. Adoption markers are noisy and partly invisible (private assistant use), so the treatment is "visible adoption" and must be described as such. Open source is not employment; the study would show whether the mechanism exists where it can be seen, not its size in firms.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 14 October 2026; not yet checked in depth).
Open-source studies already estimate newcomer inflow, first pull-request acceptance, time to first response, retention and the supply of beginner-labelled issues after AI coding-agent adoption, with conflicting results (one finds no decline; another finds newcomers' share of participation down 3.7 points). None found measures the content or amount of review feedback directed at first-time contributors, explicit mentoring, or issue assignment from maintainers to newcomers, and both adoption studies cover the agent era from 2023 rather than the earlier assistant wave. In firms, the delegation channel rests on interview studies; no study using staffing or assignment data was found.
How strong this evidence is
- Who searched and checked: an independent AI reviewer with fresh context (not a person), 14 October 2026 (review of PR #72): verdicts, nearest-work reading labels and limits checked; query lists corrected to the 8 queries run where they differed. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 6 of the 6 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 (5 queries), arXiv abstract pages and a keyword search of two full texts on the date shown. ACM Digital Library, IEEE Xplore and the 2026 MSR and ICSE proceedings were not searched systematically. A study of about 23,000 AI-assisted pull requests was seen only through a bibliography page and could not be identified. Legal-industry evidence found was surveys and firm announcements, not staffing data. 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
- Xu, Cui, Ye & Zhou, "Decoupling Code Complexity from Newcomer Participation: A Causal Study of AI Coding Agent Adoption in OSS," arXiv:2607.01810 (2026) Judged from an abstract, summary or search listing only
Abstract read and the full text searched by keyword. Difference-in-differences of AI coding-agent adoption against matched projects: no significant decline in newcomer inflow, first pull-request acceptance, time to merge, time to first response, retention or beginner-issue supply. Does not measure review content or mentoring, according to the text searched. Partly does the proposed design. - Zhang, Jiang & Koziolek, "Augmentation with Dilution: A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption," arXiv:2606.26289 (2026) Judged from an abstract, summary or search listing only
Abstract read and the full text searched by keyword. Staggered difference-in-differences on 11,097 repositories, January 2023 to May 2026: no change in the number of human contributors, newcomers' share down 3.7 points, comments per review up 5.3%. Review depth is measured per repository, not as feedback received by newcomers. - Gao, Banyongrakkul, Guan, Zahedi & Treude, "On Autopilot? An Empirical Study of Human-AI Teaming and Review Practices in Open Source," arXiv:2601.13754 (2026) Judged from an abstract, summary or search listing only
Abstract read. AI co-authored pull requests from contributors without prior code ownership get the least feedback, and about 80% are merged with no explicit review. Cross-sectional comparison of pull-request types, not before and after adoption. - Xu, Medappa, Tunc, Vroegindeweij & Fransoo, "AI-Assisted Programming Decreases the Productivity of Experienced Developers by Increasing the Technical Debt and Maintenance Burden," arXiv:2510.10165 (2026) Judged from an abstract, summary or search listing only
Abstract read. After Copilot's introduction, gains are driven by peripheral developers while core developers review 6.5% more code and write 19% less of their own. Shows maintainers' review load rising; does not measure what newcomers receive. - Song, Agarwal & Wen, "The Impact of Generative AI on Collaborative Open-Source Software Development: Evidence from GitHub Copilot," arXiv:2410.02091 Judged from an abstract, summary or search listing only
Abstract read. Copilot usage data and project data: more contributions, longer coordination time and more code discussion; peripheral developers see smaller gains and larger coordination costs. Measured by core or peripheral role, not for first-time contributors; no mentoring measure. - Yu & Moon, "Who Will Become the Next Senior? How Generative AI Erodes the Development Pathway in Software Engineering," arXiv:2607.17067 (2026) Judged from an abstract, summary or search listing only
Abstract read. Fourteen interviews with junior and senior software engineers in South Korea describe AI absorbing entry-level work into senior workflows. Qualitative; supports the mechanism without measuring it.
Search queries used
open source newcomers first-time contributors after ChatGPT Copilot adoption code review mentoring empirical GitHub studygenerative AI senior workers delegate less to juniors mentoring feedback empirical study "shadow learning" BeaneGitHub Copilot effect on open source collaboration project-level code review participation knowledge sharing difference-in-differenceslaw firms generative AI junior associates fewer tasks training hours staffing empirical data study 2026LLM era pull request review comments newcomers first-time contributors feedback mentoring decline empirical MSR 2026
Provenance
RO-LD04-03 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-04 debate on 2026-10-14. Checked by an independent AI reviewer with fresh context (not a person), 14 October 2026 (review of PR #72): verdicts, nearest-work reading labels and limits checked; query lists corrected to the 8 queries run where they differed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.