When the Ladder Moves: Learning, Apprenticeship and the First Rung of Careers When AI Does the Entry-Level Work
Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-11-22). "When the Ladder Moves: Learning, Apprenticeship and the First Rung of Careers When AI Does the Entry-Level Work". SomaSoft Research. https://somasoft.ai/papers/when-the-ladder-moves. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
When the Ladder Moves
Learning, apprenticeship and the first rung of careers when AI does the entry-level work
Draft, 22 November 2026. Not peer reviewed. Part of a series with "The Power to Say No" and "Beyond the Price."
Reading conventions. [I] marks our own inference or argument: not sourced, and offered to be disagreed with. Every other sentence that states a figure, a finding or a named source is a reported fact, taken from the publication cited, and is listed with its source in the facts file that accompanies this paper. Where a study is a working paper, observational or small, we say so in the text, because early evidence quoted without its caveats is how a finding becomes a slogan.
1. The first rung was never only a job
The bottom of a career ladder has always done two things at once. It produced something an employer would pay for: first drafts, routine tickets, document review, simple code, the first pass on a spreadsheet. And it was where people learned the job, through repetition, small mistakes that were caught, and watching how someone more experienced fixed them.
[I] Those two functions were bundled by economics rather than by design. Entry-level work was worth paying for, so employers hired juniors, and the juniors learned as a side effect. Generative AI can now do much of that routine work quickly and cheaply. If the paid half of the bundle goes to a machine, the learning half does not automatically survive. This paper asks what the evidence says about that risk, and how the first rung could be rebuilt on purpose.
The evidence points in two directions at once, and both are real. AI assistance helps novices most, often by a wide margin. But practice that a tool does for you is practice you do not get, and skills that are not exercised can weaken.
2. AI helps novices most, inside its frontier
Three studies are now the standard evidence, and they agree on direction.
Customer support. Brynjolfsson, Li and Raymond studied a conversational assistant introduced in stages to more than 5,000 customer-support agents. In the NBER working paper, issues resolved per hour rose 14 per cent on average and 34 per cent for novice and low-skilled workers, with minimal impact on the most experienced and skilled. The published version in the Quarterly Journal of Economics (2025) is summarised by its publisher as a 15 per cent average gain and a 30 per cent gain for less experienced, lower-skilled agents. Agents with two months of tenure who had the tool performed as well as agents without it who had more than six months.
During software outages, when the assistant gave no suggestions, agents who had used it still performed better than their pre-AI baseline. The authors read this as durable learning: the tool's suggestions taught the agents something. The outage estimates are less precise than the main results.
Writing tasks. Noy and Zhang (Science, 2023) gave 453 college-educated professionals occupation-specific writing tasks. Access to ChatGPT cut average completion time by 40 per cent and raised rated quality by 18 per cent. Performance inequality between workers fell, because those who did worse on the first task gained most.
Consulting. Dell'Acqua and colleagues ran a pre-registered experiment with 758 consultants at Boston Consulting Group. On 18 realistic tasks inside the frontier of what GPT-4 could do, consultants with AI completed 12.2 per cent more tasks, finished 25.1 per cent faster and produced work rated more than 40 per cent higher in quality. Those below the average performance threshold improved by 43 per cent, those above by 17 per cent. On one task deliberately chosen to sit outside the frontier, consultants using AI were 19 percentage points less likely to reach a correct answer than those without it. The authors called this uneven boundary a "jagged technological frontier".
[I] Read together, these studies say AI compresses the distance between novice and expert on tasks the tool can do. They also say that a novice cannot easily tell which tasks those are. The task outside the frontier looked like the others. Knowing where the frontier lies is itself expertise, and it is the kind that comes from having done the work.
3. Practice that is skipped is not learned
The counter-evidence comes from settings where a person's own unaided performance was measured after a period of AI help.
Students. Bastani and colleagues (PNAS, 2025) ran a randomised trial with nearly 1,000 high-school mathematics students in Turkey. One group used "GPT Base", a standard chat interface; another used "GPT Tutor", with prompts designed to protect learning by guiding rather than handing over answers. During practice with the tools, GPT Base students did 48 per cent better than the control group and GPT Tutor students 127 per cent better. On a later exam taken without AI, GPT Base students scored 17 per cent lower than students who never had access. GPT Tutor students scored about the same as the control group. The authors conclude that unrestricted access to GPT-4 can harm educational outcomes, and that the safeguards largely prevented the harm.
Clinicians. Budzyń and colleagues (The Lancet Gastroenterology & Hepatology, 2025) studied 19 experienced endoscopists, each with more than 2,000 prior colonoscopies, at four centres in Poland where AI polyp detection was introduced at the end of 2021. In colonoscopies done without AI, the adenoma detection rate fell from 28.4 per cent before routine AI use to 22.4 per cent after, an absolute change of −6.0 percentage points (95% CI −10.5 to −1.6). The study is observational, covers 1,443 non-AI procedures, and is open to confounding, as the authors and commentators note. It is still the clearest clinical signal so far that skills can weaken quickly once a tool takes over part of the looking.
Operators and knowledge workers. [I] None of this is new to human-factors research. Parasuraman and Manzey (Human Factors, 2010) reviewed decades of studies of automation complacency and automation bias and concluded that both are attention effects, found in naive and expert users alike, and that automation bias could not be prevented by training or instructions. In aviation, the US Federal Aviation Administration's Safety Alert for Operators 13002 (4 January 2013) cited an increase in manual handling errors and encouraged airlines to give pilots opportunities to fly by hand. A 2025 survey of 319 knowledge workers by Lee and colleagues at Microsoft Research, covering 936 examples of AI use at work, found that higher confidence in the AI was associated with less critical thinking, and higher self-confidence with more.
[I] The aviation response is the important part. The FAA did not ban the autopilot. It asked airlines to budget manual flying so that the skill would be there on the day the automation failed. That is the model we propose for juniors in section 8.
4. What is happening to the first rung
Payroll data. Brynjolfsson, Chandar and Chen used ADP payroll records covering millions of US workers. The August 2025 version, "Canaries in the Coal Mine?", reported that workers aged 22 to 25 in the most AI-exposed occupations had experienced a 13 per cent relative decline in employment, after controlling for firm-level shocks, while more experienced workers in the same occupations did not. The decline was concentrated where AI is more likely to automate tasks than to augment them. A revision dated 12 August 2026, with data through June 2026, reports that young workers in AI-exposed occupations now stand 19 per cent below where they would be had they kept pace with less-exposed peers, with no comparable gap for experienced workers. The authors describe these as early indicators, not causal estimates.
Résumé data. Hosseini and Lichtinger, using résumé and job-posting data on nearly 62 million workers in 285,000 US firms from 2015 to 2025, find that from the first quarter of 2023 junior employment in firms adopting generative AI declined sharply relative to non-adopters, driven primarily by slower hiring rather than more separations.
The cautious reading. The Budget Lab at Yale, writing with Brookings on 1 October 2025, found the broader US labour market "characterized broadly by stability, rather than disruption" in the 33 months after ChatGPT's release, while noting that this was consistent with emerging evidence of effects on early-career workers and that a weakening labour market could also be hurting them.
[I] Our reading is that the aggregate is calm and the entry point is not. If firms that adopt AI hire fewer juniors rather than dismissing anyone, nobody is laid off and the ladder still loses its bottom rung. Ide (2025) formalises the long-run worry: in his model, novices learn tacit knowledge by working alongside experts, and entry-level automation can raise output on adoption while reducing growth and welfare, because fewer novices learn from the best experts. That is a model, not a measurement.
5. What apprenticeship systems teach
Germany and Switzerland run "dual" systems in which most training happens inside firms, alongside part-time vocational school. They are the largest working examples of a deliberately designed first rung.
Scale. Around two-thirds of young people in Switzerland enter vocational education and training after compulsory school, a share the Swiss government says has been steady for many years. In Germany, the Federal Institute for Vocational Education and Training (BIBB) counted about 476,000 new dual training contracts as of 30 September 2025, 2.1 per cent fewer than a year earlier, with about 54,400 places unfilled and about 84,400 young people still searching.
Why firms train. Cost-benefit surveys explain why Swiss firms take on apprentices. Dionisius and colleagues (IZA, 2008) found apprenticeship training on average profitable for Swiss firms during the training period itself, unlike in Germany, and attributed most of the difference to Swiss apprentices being given a higher share of productive tasks, and to relative pay. Moretti and colleagues (2019) estimated that a Swiss firm gains on average about 3,400 euros per apprentice per year of training, while a comparable Austrian firm bears net costs of about 4,200 euros and recovers them mainly by keeping apprentices on as skilled workers.
The trade-off. Hanushek, Schwerdt, Woessmann and Zhang (Journal of Human Resources, 2017) found, using adult-skills data from 11 countries, that the early employment advantage of vocational education is offset later in life by lower adaptability and lower employment, especially in countries that emphasise apprenticeships.
[I] Two lessons follow. First, the Swiss model works because apprentices do real, valuable work from early on. That is exactly the work AI is now absorbing, so the arithmetic that made training pay for itself will weaken unless the productive tasks are kept for learners on purpose. Second, training narrowly for today's tasks has a cost that arrives decades later. When tasks change fast, a first rung has to teach judgement and general skill, not only the current procedure.
6. Tutoring works when it guides
Human tutoring. Nickow, Oreopoulos and Quan's meta-analysis of randomised trials reported a pooled effect of 0.37 standard deviations in its 2020 working-paper version and 0.288 across 89 studies in its later published version, with larger effects for programmes run during the school day. In Chicago, Guryan and colleagues (American Economic Review, 2023) found that intensive two-on-one maths tutoring by trained paraprofessionals raised maths scores by 0.18 to 0.40 standard deviations in two randomised trials.
AI tutoring. A World Bank pilot in Edo State, Nigeria, gave about 800 first-year senior secondary students six weeks of after-school English sessions in June and July 2024, using Microsoft Copilot with GPT-4. Teachers introduced each topic, suggested prompts, mentored students and led reflection. The World Bank reports an effect of about 0.3 standard deviations, equivalent to nearly two years of typical learning, better than 80 per cent of interventions in its database of randomised trials in developing countries. At Harvard, Kestin and colleagues (Scientific Reports, 2025) randomised 194 physics students, in a crossover design, between an AI tutor and an active-learning class. Median learning gains with the AI tutor were more than double. That tutor was built to reveal one step at a time and never the full solution, and was given correct solutions in advance.
[I] The successful tutors share the feature that separated GPT Tutor from GPT Base: they make the learner do the thinking, and a person stays in the loop. [I] The sessions in Nigeria were teacher-led; the Harvard tutor withheld answers. The evidence does not show that AI teaches. It shows that AI designed and supervised as a tutor can teach, and AI used as an answer machine can stop people learning.
7. Redesigning the ladder
[I] Everything in this section is our argument from the evidence above. The old ladder assumed that production and training were the same activity at the bottom. That assumption is failing, so the two have to be separated and each paid for explicitly.
- Production at the entry level will increasingly be done by people supervising AI. That is real work, and the studies in section 2 suggest juniors can do it well on tasks inside the frontier.
- Training has to become a protected activity: tasks done without the tool, or with a tool set to guide rather than answer, then checked by someone more experienced. This is how the FAA treats manual flying and how Swiss firms treat productive apprentice work. It is a budget, not a by-product.
- Judgement about the frontier is the scarce skill. It is learned by having done the work, getting it wrong under supervision and seeing why. A junior who has only ever reviewed AI output has not had that chance.
We think the measurable question for any organisation is simple. Can its juniors, after two years, do the core task unaided to the standard its seniors expect? If not, it is spending down a stock of expertise it is not replacing.
8. Rules
[I] These are proposals, not tested policy.
For employers
- Keep a protected practice budget for juniors. Reserve a fixed share of junior time for doing core tasks without AI, or with AI in tutor mode, and review that work. Treat it like manual-flying time: scheduled, not left to spare moments.
- Use AI as a tutor, not an answer machine, for learners. Configure tools for new staff to explain, question and reveal steps, as in the GPT Tutor and Harvard designs, rather than produce finished work.
- Give credit for supervised work. Count reviewed practice, corrected drafts and supervised review of machine output in performance reviews and promotion, so that learning is not penalised against output.
- Keep hiring at the bottom. If AI raises senior productivity, keep some of the gain as junior hiring and mentoring time. A firm that stops hiring juniors is drawing on a pool others have to fill.
- Test the frontier openly. Show juniors tasks where the tool fails, and why. Knowing the boundary is part of the job.
- Measure unaided skill. Periodically assess core skills without AI, as the endoscopy and classroom studies did, and act on a decline.
For schools and training systems
- Separate practice from assessment of practice. Allow AI where it guides, restrict it where the point is to build a skill, and assess that skill without it.
- Prefer guided designs. Choose tools that withhold answers and require attempts, and keep a teacher in the loop.
- Fund tutoring as the main use of AI money. The strongest evidence is for structured, frequent tutoring, human or AI-assisted.
- Build work-based routes with real tasks. Learn from the dual systems, but protect general skills so that training does not expire when tasks change.
For AI builders
- Ship a learning mode that is easy to turn on and hard to bypass, so that an employer or school can set it for learners.
- Show the frontier. Say when a task is likely outside what the tool does reliably, rather than answering with the same confidence.
- Support supervision. Make it easy for a senior to see what a junior asked, what the tool produced and what the junior changed.
- Measure your users' unaided skill, not only their output, and publish the result, including when it is unflattering.
9. Limits
- Much of the evidence is early. The employment studies are working papers using data that end in 2025 or mid-2026, and their authors describe them as early indicators. Effects could grow, shrink or prove to be cyclical.
- The deskilling evidence is thin. The endoscopy study is observational and small; the classroom study is one school in one country. Neither shows that skills cannot be rebuilt.
- The novice-gain studies are short. They measure weeks or months on specific tasks. None follows a cohort of AI-assisted juniors through to seniority.
- The tutoring pilots are small. The result from Nigeria is reported by the World Bank in a working paper and blog; the Harvard trial is one course at one university.
- Apprenticeship evidence comes from a few rich countries with long traditions, strong employer associations and stable occupations. It may not transfer.
- Our rules are untested. We have not applied them in a workplace or school. The project that wrote this paper has not evaluated its own system with real users.
10. Conclusion
[I] The ladder is not disappearing, but it is moving. AI makes novices more productive on tasks it can do, and the same tool, used without care, takes away the practice that makes a novice an expert. Whether the next generation of experts exists depends on choices employers, schools and builders make now: to keep a protected place for practice, to use AI to guide learners rather than answer for them, and to reward supervised learning as real work. That costs little next to discovering, a decade from now, that nobody below a certain age can do the job without the tool.
References
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R. (2025). Generative AI without guardrails can harm learning: evidence from high school mathematics. Proceedings of the National Academy of Sciences 122(26), e2422633122. doi:10.1073/pnas.2422633122
- Brynjolfsson, E., Chandar, B. & Chen, R. (2025; revised 12 August 2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab working paper.
- Brynjolfsson, E., Li, D. & Raymond, L. (2023). Generative AI at work. NBER Working Paper 31161; Quarterly Journal of Economics 140(2), 889–942 (2025).
- Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology 10(10), 896–903. doi:10.1016/S2468-1253(25)00133-5
- Bundesinstitut für Berufsbildung (BIBB) (2025). Press release 40/2025 on new training contracts as of 30 September 2025, 10 December 2025.
- De Simone, M. E. et al. (2025). From chalkboards to chatbots: evaluating the impact of generative AI on learning outcomes in Nigeria. World Bank; and World Bank blog posts of 18 September 2024 and 9 January 2025.
- Dell'Acqua, F. et al. (2023). Navigating the jagged technological frontier: field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013; Organization Science 37(2), 403–423.
- Dionisius, R. et al. (2008). Costs and benefits of apprenticeship training: a comparison of Germany and Switzerland. IZA Discussion Paper 3465.
- Federal Aviation Administration (2013). Safety Alert for Operators 13002: Manual Flight Operations, 4 January 2013.
- Guryan, J. et al. (2023). Not too late: improving academic outcomes among adolescents. American Economic Review 113(3), 738–765.
- Hanushek, E. A., Schwerdt, G., Woessmann, L. & Zhang, L. (2017). General education, vocational education, and labor-market outcomes over the lifecycle. Journal of Human Resources 52(1), 48–87.
- Hosseini, S. M. & Lichtinger, G. (2025). Generative AI as seniority-biased technological change: evidence from U.S. résumé and job posting data. SSRN working paper.
- Ide, E. (2025). Automation, AI, and the intergenerational transmission of knowledge. arXiv:2507.16078.
- Kestin, G., Miller, K. et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 3 June 2025.
- Kinder, M., Gimbel, M., Kendall, J. & Lee, M. (2025). New data show no AI jobs apocalypse, for now. Brookings and The Budget Lab at Yale, 1 October 2025.
- Lee, H.-P. et al. (2025). The impact of generative AI on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. CHI 2025, Microsoft Research.
- Moretti, L., Mayerl, M., Mühlemann, S., Schlögl, P. & Wolter, S. C. (2019). So similar and yet so different: a firm's net costs and post-training benefits from apprenticeship training in Austria and Switzerland. Evidence-based HRM 7(2), 229–246.
- Nickow, A., Oreopoulos, P. & Quan, V. (2020; published 2024). The promise of tutoring for PreK–12 learning: a systematic review and meta-analysis of the experimental evidence. NBER Working Paper 27476; American Educational Research Journal.
- Noy, S. & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science 381, 187–192. doi:10.1126/science.adh2586
- Parasuraman, R. & Manzey, D. H. (2010). Complacency and bias in human use of automation: an attentional integration. Human Factors 52(3), 381–410. doi:10.1177/0018720810376055
- State Secretariat for Education, Research and Innovation (SERI). Vocational and Professional Education and Training in Switzerland: Facts and Figures.