LD-04 — LEVELLING OR HOLLOWING? SKILL COMPRESSION AND THE REPRODUCTION OF EXPERTISE
Living Debate v0.2 — PUBLISHED
STATUS
Second version, repository-authored. It adds one experiment that v0.1 left out (section 1), narrows two statements that were worded more strongly than the evidence (the short answer and sections 4, 5, 6 and 10), and says which version of one study a result comes from (section 3). Revised by Claude (SocioLens Project Lead), an AI model made by Anthropic; not reviewed by a human domain expert. Version 0.1 stays available unchanged.
EVIDENCE CUT
14 October 2026
LAST UPDATED
18 October 2026
WHAT CHANGED IN v0.2
Added: the 2022 GitHub Copilot experiment by Peng, Kalliamvakou, Cihon and Demirer (LD04-SL-C21), in section 1. v0.1 did not cite it. It strengthens the evidence for gains while the tool is in use on a codified task; its finding that less experienced developers gained more is not statistically significant at the conventional 5% level.
Narrowed: v0.1 said of the evidence once the tool is removed: "Gains persisted in routine customer support." The evidence is narrower: during software outages, when the assistant gave no suggestions, agents who had used it still beat their own pre-AI baseline. The short answer now says that. Section 3 also says which version of that study the outage result comes from.
Narrowed: v0.1 said of young workers hired into exposed work: "Nobody has yet measured whether those who are hired still become experts." Section 4 said the same of the supply of future experts ("has not been measured"), the heading of section 5 called the mechanism behind hollowing "untested for generative AI", and in section 6 position P2 called direct evidence of deskilling "missing" and position P5 called its mechanism "unmeasured for generative AI". SocioLens can only say that no study it found measures these things. The short answer and sections 4, 5 and 6 now say so, and a note in section 10 says what such words mean throughout this debate.
Why: AI-simulated reader tests of the site on 17 October 2026 pointed to the experiment v0.1 left out and to these statements; SocioLens then checked them against the sources. No claim changed for this version, and the current SocioLens position is unchanged.
The evidence cut is unchanged: the added study was published in February 2023, and no source published after 14 October 2026 was added.
QUESTION
When generative AI helps the least experienced workers most, does it level expertise, or does it hollow out the way expertise is reproduced?
Is the pattern best described as:
• levelling: AI lets novices do expert-level work and narrows gaps;
• hollowing: AI takes over the work through which novices used to become experts;
• complementarity: AI pays most to those who already know the domain or know how to use the tool;
• or some combination that depends on the task, the horizon and whether the tool is still in use?
SHORT ANSWER
The task-level evidence is strong for codified tasks and mostly points one way. The career-level evidence is thin and points another.
• On well-specified, codified tasks, many experiments find that AI raises average output and helps less experienced or lower-performing workers most while they use the tool.
• On open-ended or judgement-heavy tasks, several studies find no narrowing, or widening.
• When the tool is removed, the picture splits. In routine customer support, agents who had used the assistant still beat their own pre-AI baseline during software outages. In patent drafting, three months of assisted work improved seniors' unaided judgement but not juniors'. Novices learning something new without guardrails did worse once AI was taken away.
• At career level, early evidence shows fewer young workers hired into exposed occupations and at adopting firms. We found no study that measures whether those who are hired still become experts.
The central question is therefore not:
"Does AI help novices?"
It is:
DOES HELP WHILE THE TOOL IS IN USE BECOME SKILL THAT REMAINS WHEN IT IS NOT, AND DO NOVICES STILL GET THE WORK THROUGH WHICH EXPERTS ARE MADE?
EVIDENCE LANGUAGE
SocioLens distinguishes:
FACT — directly supported observation or study fact.
ESTIMATE — measured quantity whose design, sample and period matter.
INTERPRETATION — synthesis across evidence, or a theoretical model.
UNKNOWN — unresolved empirical question.
THREE THINGS THAT GET CALLED "SKILL"
This debate stays confused unless three outcomes are kept apart:
ASSISTED PERFORMANCE
How well someone does a task while the AI is available.
RETAINED SKILL
How well they do it when the AI is taken away.
THE REPRODUCTION OF EXPERTISE
Whether novices still get the work, feedback and practice through which they become experts over years.
A result on one does not settle the others. Most of the evidence is about the first.
1. WHILE THE TOOL IS IN USE, THE LEAST EXPERIENCED OFTEN GAIN MOST
ESTIMATE
Several experiments and field studies agree on well-specified tasks.
In one large customer-support firm, an AI assistant raised issues resolved per hour by 15% on average, with much larger gains for less experienced and lower-skilled agents (LD02-HQ-C12). In a preregistered writing experiment, ChatGPT cut time by 40%, raised quality by 18% and reduced inequality between workers (LD04-SL-C01). In three company field experiments with software developers, less experienced developers adopted an AI coding assistant more and gained more (LD02-HQ-C13). In an earlier, smaller experiment with freelance developers, run by Microsoft and GitHub researchers, those given GitHub Copilot finished a standard programming task 55.8% faster; the authors report larger gains for less experienced developers, but that difference is not statistically significant at the 5% level (LD04-SL-C21). Among BCG consultants, GPT-4 users did better on tasks inside the tool's frontier, with the bottom half of baseline performers gaining most, but did worse on a task outside it (LD02-HQ-C14).
This is the evidence behind the levelling position. It is strongest for short, codified tasks with clear quality standards.
2. ON OPEN-ENDED WORK, THE PATTERN BREAKS
ESTIMATE
Compression does not hold everywhere.
Among Kenyan small-business owners, a GPT-4 adviser had no average effect, but did worse for those who started below the median than for those above it; both groups got similar advice, and the difference lay in what they did with it (LD04-SL-C02). In a UK experiment with a sample of the working-age population, ChatGPT raised productivity on every task, but it did not change productivity differences between gender, age, education or occupation groups (LD04-SL-C04). In creative tasks, AI raised the returns to general ability and education and reduced the value of domain-specific expertise (LD04-SL-C05). In a study-with-AI experiment, gains followed skill at eliciting and checking AI output rather than grades or prior knowledge (LD04-SL-C06).
Among highly experienced open-source developers working on their own projects, AI tools made tasks slower in a 2025 trial, while the developers believed they were faster (LD04-SL-C07). A 2026 meta-analysis of programming studies found large gains in controlled experiments that were not distinguishable from zero in enterprise and open-source settings (LD04-SL-C08).
THERE IS NO VALIDATED RULE YET FOR PREDICTING, BEFORE A STUDY, WHETHER AI WILL NARROW OR WIDEN A GAP ON A GIVEN TASK.
3. WHEN THE TOOL IS TAKEN AWAY, THE EVIDENCE SPLITS
ESTIMATE
This is the evidence that decides the debate, and there is little of it.
Retained gains. In the customer-support firm, agents who had used the assistant still beat their own pre-AI baseline during software outages, more so the longer and more closely they had followed it (LD04-SL-C09). This outage result is cited from the November 2024 preprint of the study; the published version in the Quarterly Journal of Economics, which LD02-HQ-C12 cites for the main result, was not checked for it.
Retained gains for seniors only. In a three-month trial with patent lawyers, AI assistance improved graded drafting. On a later task done without AI, treated seniors improved and treated juniors did not, and the difference was significant (LD04-SL-C10). The study was conducted and funded by Google.
Losses for novices learning something new. Turkish high-school students who practised with a standard GPT-4 interface did 17% worse on a closed-book exam later in the same session, while a tutor version with safeguards caused no detectable loss; the narrower spread of practice scores did not survive removal of the tool (LD04-SL-C11). Programmers learning an unfamiliar library with an AI assistant scored 17% lower on a later quiz, and those who delegated code scored lowest (LD04-SL-C12). In classroom experiments, students who used an LLM to generate solutions understood less, students who asked for explanations understood more, and the gap by prior knowledge widened (LD04-SL-C13).
No loss straight afterwards. In an Argentine experiment, a module done without AI immediately after the task showed treated participants no worse than controls; the authors call it a diagnostic, not a test of skill formation (LD04-SL-C03).
Incumbents, in a diagnostic-AI analogue. Experienced Polish endoscopists detected fewer adenomas in colonoscopies done without AI after an AI system was introduced. The study is observational, and its authors list differences between the cohorts and a near doubling of colonoscopy volume among its limitations (LD04-SL-C14).
INTERPRETATION
A plausible reading, not yet tested: whether skill is retained depends on how routine the task is, on the worker's prior expertise, and on whether they use the tool to understand or to delegate. None of the tool-removed studies in this debate combines a baseline skill test with a horizon longer than three months.
4. AT CAREER LEVEL, FEWER JUNIORS ARE HIRED INTO EXPOSED WORK; WHETHER AI IS THE CAUSE IS CONTESTED
ESTIMATE
The career-level evidence is about who gets hired, not yet about who gets trained.
US payroll data show workers aged 22–25 in highly AI-exposed occupations about 19% below the employment path of less-exposed peers by June 2026, mainly through reduced hiring; the authors stress this is descriptive (LD02-HQ-C04). In US résumé data, firms that began integrating generative AI saw junior employment fall about 9% relative to senior employment, mainly through fewer junior hires, while junior promotion rates rose slightly (LD04-SL-C15). In Sweden, the fall in job postings tracks monetary tightening, but young workers in exposed occupations are 5.5% lower within the same employers (LD04-SL-C16). Danish registers show precise nulls on incumbent earnings and hours (LD02-HQ-C02), and exposed online freelancers lost work and earnings, with past high quality not protecting them fully (LD02-HQ-C15). Whether AI causes the fall in junior hiring is contested: in Sweden the overall fall in postings tracks interest rates (LD04-SL-C16), and Norwegian registers show near-zero employment differences by exposure and an imprecise zero for young workers (LD02-HQ-C09).
Across US occupations, those more exposed to AI already pay more for experience, and since 2022 exposure has gone with weaker wage growth only where experience pays little (LD04-SL-C17). That is a level difference across occupations, not evidence that the junior–senior premium is widening.
A FALLING NUMBER OF JUNIORS IS NOT THE SAME AS A FALLING SUPPLY OF FUTURE EXPERTS. NO STUDY WE FOUND MEASURES THE SECOND.
5. THE MECHANISM BEHIND HOLLOWING IS OLD; WE FOUND NO TEST OF IT FOR GENERATIVE AI
FACT + INTERPRETATION
When surgical robots removed trainees from hands-on work, approved routes for learning failed; only a minority of trainees came to competence, through tolerated but unapproved practice, and the supply of experts shrank relative to demand (LD04-SL-C18). A recent model formalises the same worry: automating entry-level work can raise output while lowering long-run growth, even without cutting entry-level jobs, by moving novices away from the best experts (LD04-SL-C19). Another model predicts that the direction of compression can reverse as AI moves from assisting to acting autonomously (LD04-SL-C20).
These are mechanisms and models, not measurements of generative AI.
6. THE FIVE POSITIONS
P1 — LEVELLING
Claim:
AI lets workers with less training do expert-level work and narrows performance gaps.
Current status:
Supported for assisted performance on codified tasks. Not shown for retained skill or careers.
P2 — HOLLOWING (DESKILLING)
Claim:
Expertise is coded into tools; novices do less of the work that built skill, and discretion and pay can fall even as output rises. This position draws on labour process theory.
Current status:
Consistent with fewer junior hires in exposed work and with losses among novices learning without guardrails. We found no direct evidence on whether generative AI deskills workers over time.
P3 — EXPERTISE AS A COMPLEMENT
Claim:
Gains depend on, or widen with, prior expertise or skill at using the tool.
Current status:
Supported on open-ended tasks and in retained skill in one professional trial. Theory predicts it grows as AI becomes autonomous.
P4 — PROFESSIONS AND JURISDICTION
Claim:
AI reopens contests over who may do and certify expert work; the outcome is delegation and boundary work rather than uniform deskilling. This position draws on the sociology of professions.
Current status:
Not yet tested with generative-AI evidence in this debate.
P5 — APPRENTICESHIP GAP
Claim:
AI raises productivity now and reduces the supply of experts later, because novices lose the practice through which expertise is reproduced.
Current status:
A strong mechanism from earlier technology and from theory; we found no study that measures it for generative AI.
7. WHY THE CAMPS TALK PAST EACH OTHER
INTERPRETATION
• "Skill" means assisted output to most experiments, discretion and control to labour process theory, the right to do and certify work to the sociology of professions, and capacity built through practice to studies of apprenticeship.
• "Levelling" can mean a narrower output distribution, lower wage premiums or easier entry. Lower expertise requirements can mean more access and lower pay at the same time; calling that levelling or hollowing is a normative judgement.
• The unit of analysis differs: task, worker, occupation, profession or career.
8. WHAT WE STILL DO NOT KNOW
UNKNOWN
1. Do people who learn a task with AI become less able to do it without AI over months or years ("never-skilling")?
2. Does retained skill depend on how routine the task is and on prior expertise, as the split in tool-removed studies suggests?
3. Does AI reduce the real work, feedback and supervised struggle that seniors hand to juniors?
4. Is the junior–senior wage premium changing within occupations?
5. What do employers actually deploy for juniors: guardrails, learning modes, or unrestricted tools?
6. Will compression reverse as AI agents act autonomously rather than assist?
7. Does routine AI use deskill experienced professionals in the skills they no longer exercise?
9. WHAT EVIDENCE WOULD CHANGE THE PICTURE?
• Longitudinal studies with a baseline skill test that follow AI-trained novices without AI for a year or more;
• tool-removed experiments that cross routine and judgement-heavy tasks with novice and experienced workers;
• measures of task allocation, feedback and mentoring for juniors before and after adoption;
• individual wage–tenure panels in exposed occupations;
• trials with planned AI-off sessions in clinical or professional settings;
• the same experiment run with copilot and autonomous-agent tools.
10. CURRENT SOCIOLENS POSITION
INTERPRETATION
The best-supported description of the evidence to October 2026 is:
AI COMPRESSES PERFORMANCE WHILE THE TOOL IS IN USE ON CODIFIED TASKS. WHETHER IT COMPRESSES SKILL IS UNKNOWN, AND WHETHER IT PRESERVES THE PATH TO EXPERTISE IS UNMEASURED.
Levelling is established only for assisted performance. Hollowing is a well-founded worry with a clear mechanism and almost no direct generative-AI evidence. The most useful research now measures retained skill and junior work over time, not more assisted-performance experiments.
In this debate, words such as "unmeasured", "untested", "unknown" and "missing" mean that SocioLens found no study that measures the point, not that none exists.
GEOGRAPHIC AND EVIDENCE SCOPE
The evidence is strongest in the United States, with studies in Kenya, Argentina, the United Kingdom, Türkiye, Sweden, Denmark and Poland. It is mostly from 2023–2026, mostly working papers and preprints, and mostly about tools from 2023–2025. Several key studies were run or funded by AI developers (Google for the patent-drafting trial; Anthropic Fellows for the programming-library experiment; GitHub or Microsoft co-authors on several coding studies), and are marked as such. Medical studies concern diagnostic AI and are analogues, not generative-AI evidence.
PROVENANCE
Research build:
Source notes prepared 13–14 October 2026 by Claude with AI research agents, then checked against official abstracts and, where the sites allowed, full texts (some PDFs supplied by the Owner).
Claims:
Each claim rests on a source checked in full text or against its official abstract; the claim's verification state says which. Seven claims rest on the abstract only (C01, C04, C05, C06, C16, C18 and C20; for C16 the front matter too). For C01, C04 and C18 the abstract was read in the publisher's DOI record on 14 October 2026; for C04 that abstract differs from an earlier one summarised in the source notes, and the claim follows the abstract read. Sources found only in press coverage, unverified figures and a withdrawn study were left out.
Shared with LD-02:
Seven claims cited here belong to LD-02 and are not repeated.
Review:
Checked on 14 October 2026 by an independent AI reviewer with fresh context (not a person), and corrected after that review. Not reviewed by a human domain expert.
Version 0.2:
Revised in the repository on 18 October 2026 by Claude (SocioLens Project Lead), an AI model. It cites every claim v0.1 cites, at the same version except LD02-HQ-C02, which it cites at v1.1 (revised on 14 October 2026; the revision does not change what this debate cites it for), plus LD04-SL-C21, which rests on the full text of its source (arXiv, read on 18 October 2026).
Previous version:
Living Debate v0.1 (14 October 2026), still available unchanged.
Version:
Living Debate v0.2 — PUBLISHED
18 October 2026