An AI-assisted task is evidence of completed work, not by itself evidence that the apprentice can repeat the skill. For a task that matters again, use a brief, low-stakes check: ask the apprentice to explain one key decision, try a changed but comparable example without generated steps, and identify a likely error. This explain, reproduce, diagnose method is a practical heuristic, not a validated test. It gives stronger task-specific evidence of learning than completion or confidence alone, while leaving AI available for useful work.
What counts as evidence that you learned the work?
A finished ticket, report, code change or customer reply answers a delivery question: was the work completed with the available tools? It does not answer a different question: what can the apprentice now do without those tools supplying the method or answer? Keep those records separate. A task can be valuable output even when it has not yet demonstrated independent mastery.
A useful distinction is between performance with assistance, retention after the help is removed, and transfer to a changed task. The last two are not identical either. Someone may remember steps for the exact exercise but struggle when one condition changes. An apprentice who can explain why a step matters, perform the core process on a fresh comparable case, and notice a plausible failure gives a supervisor better evidence of capability than a polished result alone. This three-part check is a practical synthesis, not a validated assessment instrument or a universal threshold.
A 2026 randomized experiment by Judy Hanwen Shen and Alex Tamkin studied 52 developers who knew Python but were learning an unfamiliar asynchronous programming library. After two short coding tasks, participants completed an assessment without AI covering conceptual understanding, code reading and debugging. The paper reports lower average scores for the AI-access group, while also identifying interaction patterns involving explanation and conceptual questions that were associated with better learning outcomes. Its important lesson for this question is methodological: the researchers checked what participants could do after the assisted work, rather than treating the code they produced as proof of learning. The small, bounded software study does not establish what happens in every apprenticeship or trade ([study](https://arxiv.org/abs/2601.20245)).
Sources: How AI Impacts Skill Formation
Why does an impressive AI-assisted result leave the question open?
Because the deliverable may reflect a joint workflow. A capable tool can generate a solution to a task that would otherwise be beyond a trainee's current level. That expands what the person and tool can complete together. It does not show, on its own, that the person has acquired the concepts needed to reproduce, adapt or check that solution later. Task exposure to a tool, the tool's ability to generate an answer, and the apprentice's learned skill are three different things.
The directly relevant apprenticeship research is designed to examine this distinction. Stanford Digital Economy Lab describes a randomized controlled trial in German vocational schools with final-year IT apprentices in application development and system integration. Some apprentices receive generative AI tools; others complete the same tasks without internet or AI assistance. The tasks range from standard apprenticeship work to more advanced problems, and the project says it studies task range and understanding. The opened project page describes the design but does not report findings ([project description](https://digitaleconomy.stanford.edu/project/task-expansion-with-generative-ai-the-case-of-apprenticeships/)). It therefore cannot yet tell us whether AI improved or weakened these apprentices' learning.
The coding experiment above supplies a post-task measure, but in a different population and setting. It concerns a new software library, not a full apprenticeship, a long training period, or a physical trade. The gap matters: successful performance on a harder assisted assignment can be a sensible way to contribute while learning, yet it should not be the only basis for assigning independent responsibility. Ask for a smaller demonstration with the scaffold reduced. Until direct apprentice outcomes are reported, a broad claim about AI and apprenticeship learning would outrun the available evidence.
Sources: How AI Impacts Skill Formation; Task Expansion with Generative AI: The Case of Apprenticeships
Can AI help someone learn while making the work faster?
Yes, those outcomes can coexist. An AI tool might help a newer worker handle a live task more effectively and still leave open what the worker can do alone tomorrow. Productivity measures the workflow's output; a learning check asks what knowledge or skill the worker has retained and can apply. Neither measure substitutes for the other.
A Stanford working paper by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examines the staggered introduction of a conversational AI assistant among 5,179 customer-support agents. Its abstract reports a 14% average increase in issues resolved per hour, with larger reported gains for novice and lower-skilled agents. The authors describe evidence that the tool disseminated practices from more able workers and helped newer workers move down the experience curve as suggestive, and say assistance may lead to worker learning ([study record](https://gsbpreserve.stanford.edu/view/4141/generative-ai-at-work?limit=100&offset=2515&sort=order)). That is a meaningful counterweight to the idea that assistance must block learning. But the result comes from one customer-support deployment; the abstract does not establish independent transfer on a fresh task for every agent.
This workplace evidence and the randomized coding experiment are not contradictory verdicts to average together. They study different work, tools, designs and outcomes: a deployed support workflow with suggestive evidence of learning versus a short learning experiment with a later unaided skills assessment. Together they support a narrower conclusion. Assistance can improve delivery and may support learning, but the manner of use and the outcome being measured matter. Asking questions, inspecting a suggestion and testing it may keep the learner involved; delegating the whole task may not create the same practice. The coding paper's interaction-pattern analysis is informative, but those patterns should not be presented as a proven recipe for every occupation.
For a supervisor or apprentice, the proportionate response is to keep useful assistance in the workflow while protecting some turns for diagnosis, explanation and verification. The point is not to ban tools or require unaided work on every task. It is to distinguish a helpful scaffold from evidence that a person can safely own the underlying step.
Sources: How AI Impacts Skill Formation; Generative AI at Work
How can you check without turning every task into an exam?
Choose one recurring, low-risk task and make the check brief. First, ask the apprentice to explain one consequential choice in their own words. Then offer a comparable example with one condition changed and ask them to perform the core step without generated instructions. Finally, review a plausible error together: what would make the result wrong, and how would they notice? This is an editorial practice suggestion based on the distinction between assisted output and separately measured skill; it is not a formal assessment validated by the studies above.
Consider an apprentice who used a tool to draft a small code change. Delivery review asks whether the change works and meets the request. A learning check might ask what a key part of the code does, how the behavior should change for a different input, and what test would catch a mistaken suggestion. If the apprentice can reason through the variation and detect a failure, that is stronger evidence of transferable skill for this bounded task. If they cannot, the useful response is another guided practice turn, not a career-wide judgment.
Match the check to the consequence of error. For a reversible draft, a quick explanation and test may be enough to decide what to practice next. Where a mistake could affect safety, money, private information or a customer's rights, an apprentice's unaided check does not replace required review, formal qualification or supervision. Increase independence only when task-specific performance and verification are dependable enough for the work's stakes. A single successful demonstration is a signal, not proof of permanent mastery.
The verdict is to count AI-assisted completion as real work while keeping a separate record of skill evidence. For a task likely to recur, the apprentice can ask a trainer or supervisor for one short, changed example to attempt unaided, then compare reasoning and error checks. Repeated success can support the next step in responsibility; a safety-critical miss calls for more supervised practice. The deciding question changes from ‘Did I get the answer?’ to ‘Can I explain, adapt and check this part when the answer is not supplied?’
Sources: How AI Impacts Skill Formation; Task Expansion with Generative AI: The Case of Apprenticeships
Questions readers ask
Does finishing a harder task with AI mean the apprentice learned more?
It shows that the apprentice and tool completed a harder task together. To judge the apprentice's learning, also check whether they can explain the key decision and perform a changed comparable task without answer generation.
Should apprentices avoid AI while they are learning?
The available evidence does not support a blanket rule. AI may help performance and can support learning in some workflows, but assisted output alone does not establish independent skill. Use assistance while preserving chances to reason, verify and practice without it.
What is a simple way to check whether a skill transferred?
Ask the learner to explain one key choice, try a comparable task with one changed condition, and identify a plausible error. Treat this as a low-stakes practice check, not a validated test or substitute for supervision in consequential work.
Do current studies settle whether AI helps apprentices learn?
No. A Stanford project describes a randomized trial with final-year German IT apprentices but its opened page gives the design rather than outcomes. Other research studies software developers learning a library and customer-support agents in a deployed workflow, which are useful but different settings.
Sources and notes
- How AI Impacts Skill Formation
Supports the distinction between assisted coding output and later unaided conceptual, code-reading, and debugging assessment in a randomized short-term study.
- Task Expansion with Generative AI: The Case of Apprenticeships
Supports the description of a randomized study design among final-year German IT apprentices; the page supplies no outcome findings.
- Generative AI at Work
Supports reported customer-support productivity changes and the authors' qualified, suggestive account of practice diffusion and possible worker learning.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map