AI Lets Apprentices Work Beyond Training. Not Mastery

An IT apprentice who has never formally studied machine learning can now attempt a machine-learning exercise with an AI assistant. That sounds like a story about the disappearance of entry-level work. A randomized field experiment in German vocational schools points to a second possibility: AI may let early-career workers reach into jobs that used to sit above their qualification. The harder question is what they have actually learned when the work is done.

A working paper released on 7 October 2026 by researchers affiliated with the Stanford Digital Economy Lab, Germany’s Federal Institute for Vocational Education and Training and other institutions tested that idea rather than relying on forecasts. In a field experiment with 673 final-year IT apprentices at 12 German vocational schools, access to ChatGPT and the internet raised scores on tasks deliberately set beyond the apprenticeship curriculum by 26 to 31 percentage points.¹ That is an increase in the share of available points earned, not a 26% rise in employment, wages or long-term skill.

The result complicates a debate often framed as a choice between AI replacing juniors and AI helping seniors. It suggests a third route: someone with a solid vocational foundation may be able to take on more demanding work sooner. A tool-assisted result is still not proof that the worker could repeat the task independently, recognize every mistake or retain the knowledge months later.

The experiment measured a bigger task, not a bigger job title

The apprentices came from two established German training tracks: application development and system integration. Everyone first completed a baseline exercise without AI. Participants were then randomly assigned either to ChatGPT Pro plus wider internet access or to a control group without internet access. They completed one task within their curriculum and two beyond it. The advanced exercises resembled entry-level software and network engineering work, including machine-learning problems and hybrid cloud migration.¹

Researchers scored each answer with task-specific rubrics. Across the three post-treatment exercises, AI access raised scores by 26 to 44 percentage points. The largest gain came on a within-curriculum task, where prior knowledge could combine with the tool. On the two deliberately more advanced tasks, gains were still 26 to 31 percentage points. The assisted group was also more likely to submit a substantive answer to tasks that unaided apprentices often left blank.¹

That matters because much AI-and-employment talk starts with a fixed task list and asks which items can be automated. This experiment asks whether the boundary of a worker’s task list can move. For these IT exercises under these conditions, the answer is yes. It does not show that employers created new positions, that qualifications became unnecessary, or that the apprentices could perform unsupervised production work at the same standard.

Encouraging comprehension results, with a large caveat

The researchers did not stop at scoring finished work. Immediately after each exercise, apprentices answered four questions intended to test understanding without AI. In the pooled sample, the AI-access group scored 6.7 percentage points higher on within-curriculum comprehension, 9.7 points higher on the first advanced task and 4.4 points higher on the second. The paper found no average decline on this immediate comprehension measure.¹

That is useful evidence against the assumption that every assisted answer is empty copying. It is not evidence of durable learning. The questions came right after the task; the study did not retest weeks later or watch unexpected incidents in a live workplace. The treatment group also had general internet access while the control group had none, so the design does not isolate the incremental benefit of AI over ordinary online search. The authors acknowledge both limits.

There is also reason to resist a simple “AI improves learning” headline. In a separate randomized study published in early 2026, Judy Hanwen Shen and Alex Tamkin examined programmers learning an unfamiliar software library. They found that AI assistance could weaken conceptual understanding, code reading and debugging, especially when users delegated the thinking rather than engaging with the material.² That smaller study and the German apprenticeship experiment tested different tasks, participants and outcomes. They are not direct contradictions. Together they show why immediate output, immediate comprehension and retained independent competence must be measured separately.

The new divide may be between access and effective use

The German paper reports that 84% of participating apprentices already used AI regularly at school or work, yet only 16% said they had learned how to use it at vocational school.¹ Apprentices who rated their own AI-use ability above the sample median generally benefited more from access, especially on the within-curriculum task. The association was stronger for practical abilities - using AI to support decisions and complete work efficiently - than for professed knowledge of how the technology works.

Two cautions remain. First, the AI-literacy measure was self-reported and collected after the experimental tasks, so it should not be treated as a clean causal estimate of what a training course would achieve. Second, the sample is unusually specific: final-year German IT apprentices, average age about 23.5, only 9% women, in occupations highly exposed to software-based AI. A retail trainee, care worker or charity administrator should not be assumed to receive the same gains.

Even so, the institutional gap is hard to miss. The tool is already in the workflow; the formal learning pathway has not caught up. An August 2026 joint report coordinated by the International Labour Organization and European partner institutions argues that AI literacy is becoming a foundational capability, alongside judgment and the ability to use AI safely.³ That is a policy recommendation, not a guarantee that a short course creates expertise. The German findings make the question concrete: who teaches people to turn assisted performance into reliable, explainable work?

What a small employer can test without redesigning every job

Imagine a small nonprofit that needs to redesign a donor database workflow, or a 30-person company whose junior IT administrator has been asked to plan a more complex network migration. These are hypothetical situations, not cases documented in the experiment. In both, AI might allow a capable employee to draft a first solution to a problem previously sent to an outside specialist. The decision is not whether to trust the first answer. It is how to test whether the employee can judge it.

A sensible pilot would distinguish three stages. First, define a stretch task with an experienced reviewer and explicit acceptance criteria. Second, let the employee use AI while recording assumptions, sources, uncertainties and decisions rather than only the final deliverable. Third, remove the tool for a short explanation or troubleshooting exercise, and revisit the same capability later. Measure the correctness of the result, the review time and what the employee can explain or reproduce. This design is an editorial recommendation informed by the research, not an intervention validated by the German study.

This approach also shows where the economics can turn against an organization. If the work expands faster than anyone can review it, the apparent productivity gain may become a quality or safety liability. If managers treat every AI-assisted success as proof of senior-level competence, they risk assigning responsibility faster than judgment develops. If they prohibit the tools entirely, they may deny capable workers a chance to try more demanding work. The choice is not between blind delegation and blanket restriction.

Broader labor-market evidence points in both directions. PwC’s June 2026 analysis of job advertisements found that highly AI-exposed junior roles were more likely to demand traditionally senior skills such as judgment and leadership.⁴ That is a hiring-pattern observation, not proof that AI caused each change. Research from Germany’s ifo Institute, using employer-panel data, also reports that AI-adopting training firms hired roughly 14% more new apprentices on average, while training access was increasingly concentrated in more highly qualified work.⁵ Neither result guarantees an easy transition for people with fewer opportunities to learn.

The real qualification is the ability to carry responsibility

The German experiment is a useful corrective to fatalism about junior workers. In a controlled setting, AI allowed vocational trainees to tackle more advanced tasks and answer immediate comprehension questions at least as well as their unaided peers. That is genuine evidence of expanded assisted capability. It is not a license to erase the distinction between performing a task with support and being accountable for the result.

For leaders deciding what to automate and whom to develop, the practical question is sharper than “Can AI do this task?” It is: can this person use AI to take on a harder task, explain the reasoning, detect failure and eventually perform the critical parts without the tool? If organizations build that progression deliberately, AI could widen the route into skilled work. If they confuse the finished answer with acquired expertise, they may narrow it instead.

How hilighter can help

Stretch assignments only work when roles, skills and evidence of competence are visible. hilighter helps teams map skills by role, compare people with requirements through Skills & Talent Insights, recognize work and learning in Skills Evidence & Recognition, set personalized growth paths, and explore an indicative view of AI’s potential impact on jobs. A free plan covers up to three roles.

Sources

1. Brosch, H., Langer, C., Lergetporer, P., Mueller, S., Pfeifer, H. and Tschoepe, N., Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career Workers, Stanford Digital Economy Lab / IZA Discussion Paper 18999, 7 October 2026.
2. Shen, J. H. and Tamkin, A., How AI Impacts Skill Formation, arXiv preprint, 28 January 2026 (revised 1 February 2026).
3. International Labour Organization and partner institutions, Changing landscape of skills in the age of AI, 13 August 2026.
4. PwC, 2026 AI Jobs Barometer, 15 June 2026.
5. Muehlemann, S., KI in Betrieben: Mehr Ausbildung - aber Weiterbildung zunehmend fuer anspruchsvollere Taetigkeiten, ifo Schnelldienst 79(3), 2026.

Sources last reviewed: 26 October 2026. The Stanford / IZA paper is a working paper from a specific German IT apprenticeship sample; comprehension was measured immediately after tasks, not as long-term retention.