The most important productivity effect of AI may not be doing the same task faster. It may be letting a person complete work that was previously outside their trained task set. A new field experiment tested that possibility with 673 final-year IT apprentices across twelve German vocational schools.[1][2]
In the working paper, apprentices were randomly assigned either ChatGPT Pro plus internet access or a no-internet control condition. The researchers scored occupation-specific tasks within and beyond the formal curriculum. On the beyond-curriculum tasks, AI access raised the share of maximum attainable points by 26–31 percentage points. This measures improvement in scored task performance, rather than a 26–31-point rise in the proportion of tasks fully completed. The authors report similar results after excluding participants who visited other websites.[1]
What the apprenticeship experiment establishes
| Measure | Value | Context |
|---|---|---|
| Participants | 673 final-year IT apprentices | Students came from twelve German vocational schools.Source 1 |
| Research design | Random assignment: ChatGPT plus internet versus no internet | The tasks were occupation specific and scored against defined outputs.Source 1Source 2 |
| Beyond-curriculum gain | 26–31 percentage points | Increase in the share of maximum attainable task points, not the percentage of apprentices completing an entire job.Source 1 |
| Immediate comprehension | No reduction detected | Four comprehension questions followed each post-treatment task; apprentices were instructed to answer without AI.Source 1 |
| AI literacy interaction | Larger effects above the median | Access alone did not eliminate differences in the ability to use the system.Source 1 |
Task expansion is different from replacement
The apprentice still had to recognise the problem, direct the assistant, judge whether the proposed steps fit the environment and submit a working result. The AI supplied capabilities that made a harder task reachable. That is closer to temporary access to a senior colleague’s procedural knowledge than to removing the human from the workflow.
The result also clarifies why AI literacy matters. A more capable model does not create equal value when one user can decompose the problem, ask for evidence, inspect commands and recover from an error while another accepts the first plausible answer. Training should therefore include how to supervise AI, not only where the button is.
Four zones for assigning work with AI
| Zone | Appropriate role for AI | Skill-development rule |
|---|---|---|
| Substitute | Routine, reversible work with a clear test | Keep periodic manual checks so the process remains understoodSource 4 |
| Complement | AI generates options while the person applies domain judgement | Require the person to explain the choice and rejected alternativesSource 4 |
| Aid | The learner attempts first, then receives hints, examples or debugging support | Preserve productive struggle before revealing a complete solutionSource 4 |
| Non-negotiable human work | High-consequence judgement, responsibility or core learning objective | AI may supply evidence but not take the accountable decisionSource 4 |
Immediate comprehension is encouraging—but not the final test
The comprehension result rebuts the simplest fear that better task performance necessarily came from blind copying. It does not settle whether the apprentices could reproduce the work weeks later, diagnose a novel failure without AI or build the tacit judgement that grows through repeated exposure. A training programme should therefore measure delayed transfer as well as completed tickets.
A separate randomised experiment with 1,174 adults found that generative AI narrowed effective productivity differences on a workplace-style business problem, while underlying education differences still shaped unassisted follow-up performance and effective tool use. That pattern is compatible with task expansion: AI can raise what people accomplish today without making their unaided human capital identical.[3]
Redesign the apprenticeship rather than defending every old exercise
If AI makes a formerly advanced task accessible, the curriculum can move upward. Apprentices can spend less time reproducing routine configuration and more time comparing architectures, investigating failures, documenting risk and defending a decision. The organisation should deliberately create those harder assignments; otherwise it may consume AI-assisted output while leaving judgement underdeveloped.
- Give the learner a task just beyond current independent capability.
- Require a short plan before AI use and a change log during the work.
- Make the learner identify one AI error, uncertainty or rejected path.
- Test the core concept immediately without the assistant.
- Retest later on a structurally similar but unfamiliar task.
- Promote the task into the normal curriculum only after the human skill objective is explicit.
This experiment gives the task-expansion idea concrete evidence. AI moved apprentices into work beyond their formal programme, with no detected reduction in the immediate comprehension measure. That finding is narrower than demonstrating equivalent comprehension or durable mastery. Repeated training and delayed unaided tests would show whether the expanded task ceiling becomes an enduring skill.
Sources & further reading
- Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career Workers Stanford Digital Economy Lab · 7 Oct 2026 · Accessed 10 Oct 2026
- Task Expansion with Generative AI: The Case of Apprenticeships Stanford Digital Economy Lab · Accessed 10 Oct 2026
- Does Generative AI Narrow Education-Based Productivity Gaps? National Bureau of Economic Research · 1 Feb 2026 · Accessed 10 Oct 2026
- SCAN: A Decision-Making Framework for Effective Task Allocation with Generative AI Stanford SCALE Initiative · 1 Jun 2026 · Accessed 10 Oct 2026