A German experiment suggests AI apprentices can complete some tasks that fall outside what they have formally been taught, without immediately losing their grasp of the work. Researchers randomly assigned access to generative AI while 673 final-year information-technology apprentices tackled occupation-related exercises. Their IZA discussion paper reports particularly striking productivity gains on tasks beyond the apprenticeship curriculum. The point is not that training has become unnecessary. It is that the boundary between what a beginner has learnt and what they can attempt may be moving.
The study was conducted across twelve vocational schools in Germany and released as an economics discussion paper. That status matters: it is research available for examination, not a guarantee that every result has been independently confirmed through journal peer review. The tasks were controlled exercises, and the findings cannot simply be transferred to a live workplace where mistakes may affect colleagues or customers. Even with those cautions, the experiment poses a practical question for employers: how should people learn when a capable assistant is available beside them?
AI Apprentices Could Attempt Work Beyond the Syllabus
Traditional training often follows a clear sequence. A person learns the foundations, practises a familiar task and gradually earns responsibility for more complicated work. The German researchers deliberately included exercises that went beyond the formal training material to examine whether access to AI allowed apprentices to perform work otherwise outside their immediate reach. In the study, it often did. The authors report productivity improvements of 26 to 31 percentage points on those beyond-curriculum tasks.
Percentage points are important here. The paper is discussing a change in measured task performance under its experimental conditions, not promising every apprentice a 26% salary increase, an equivalent saving for an employer or mastery of an entire profession. Numbers from a classroom-style exercise should not be converted into broad business claims. What they do suggest is that an inexperienced person may be able to tackle a wider range of bounded problems with well-used assistance.
Imagine a junior worker encountering a technical task with unfamiliar requirements. Without help, that person might not know which concepts to consult or how to begin. A capable assistant could offer a starting structure, explain an unfamiliar command or help interpret an error. The worker would still have to decide whether the answer fits the circumstances, check the result and understand its consequences. In a genuine workplace, a plausible but incorrect answer can create work for everyone else.
Does Using AI Stop People From Learning?
One of the most useful findings concerns immediate comprehension. The researchers report that the productivity gains did not come at the cost of reduced understanding on the comprehension measure used in the experiment. This is more informative than a simple speed result because an employer needs workers to develop capability, not merely finish an isolated task quickly. However, immediate comprehension is not the same as long-term retention, independent professional judgement or the ability to repeat the work months later without support.
The difference resembles the familiar debate about calculators in education. A tool can free attention for a harder problem, or it can allow someone to avoid learning an essential skill. The outcome depends on what the learner is asked to do and how the tool is used. With AI, that question becomes more complex because the system can provide both the answer and a fluent explanation, whether or not the explanation is correct.
An apprentice who compares possible answers, tests them and explains a choice may learn something useful from the exchange. Someone who copies an output without checking it may complete the task while gaining little understanding. The German results cannot establish which of those habits every participant will develop in the long run. They do show that extra assistance and immediate comprehension are not automatically enemies in the circumstances tested.
The authors also found that effects generally tended to be stronger among apprentices with above-median AI literacy. That suggests knowing how to use the assistant is itself a skill. It is not enough to provide access and assume the benefit is distributed equally. Workers may need practice framing a request, checking a response and recognising when to ask a person for help. In that sense, training becomes more important, even if parts of the work become easier.
The Old Experience Ladder May Need Another Rung
Employers often divide tasks according to an employee's stage of development. Beginners handle simpler activities, while more experienced colleagues take on the complex decisions. AI could change the economics of some parts of that ladder. If a junior employee can make a useful first attempt at an unfamiliar task, a supervisor may spend more time reviewing, correcting and explaining instead of preparing the entire work from scratch.
That might create opportunities for earlier exposure to interesting work. It could also create risks if managers treat tool-assisted performance as evidence that a novice is ready to operate without oversight. Passing a controlled exercise is not equivalent to understanding legal liability, security requirements or the effect of a wrong answer on an actual customer. A responsible employer should look for repeated independent competence before changing the level of responsibility attached to a role.
There is another complication. Entry-level work often provides the repetitive practice that builds expertise. If assistants remove too much of that practice, future workers might have fewer opportunities to recognise patterns and develop judgement. The experiment's immediate comprehension result is encouraging, but it does not resolve the longer-term question. A good training plan may need deliberately chosen exercises that a learner completes without automation, alongside tasks where AI support is encouraged.
LiveAIWire's reporting on junior lawyers and AI training explores a related tension: the technology can make early tasks easier while potentially changing the route by which beginners become experts. Legal work and IT apprenticeships are not the same setting, so their findings should not be merged into a single numerical prediction. The shared question is whether an organisation is using the time saved to develop people or simply to demand more output.
Access Alone Does Not Guarantee Equal Gains
A separate Gallup analysis of workplace AI adoption highlights how unevenly AI use and reported benefits are distributed among workers. Gallup's surveys are observational, not a replication of the German experiment, but they provide a useful warning against assuming that a company licence translates directly into higher performance for every employee. Job design, confidence, training and access to appropriate tools all shape what happens.
A confident user may break a problem into manageable pieces and spot when a proposed solution is unsuitable. A less familiar user may accept an impressive-looking output simply because it reads well. Neither behaviour is fixed. Clear instructions, opportunities to practise and support from experienced colleagues can change how a worker uses the technology. Employers should therefore measure not only who has an AI account, but what quality of work people produce with it and what they can explain afterwards.
This also changes how supervisors might assess performance. If one person finishes a difficult task with extensive prompting and correction, while another completes a simpler task independently, a bare productivity figure offers a poor comparison. Teams may need to assess source checking, error recognition, communication and the ability to take responsibility for the final result. An assistant can provide useful material, but the human worker remains accountable for knowing which conclusions can be trusted.
Apprenticeship Still Needs Real Work
The German study does not argue that vocational schools can abandon structured teaching. Indeed, a learner needs enough grounding to know when an automated suggestion is dangerous or incomplete. Without basic understanding, an assistant's mistakes may be hard to distinguish from a plausible solution. Training in fundamentals, workplace safety and professional standards therefore remains central even when the visible task is accelerated.
The place of AI in teaching is also a subject in LiveAIWire's reporting on teachers' time with pupils. One constructive approach is to ask learners first to describe the problem in their own words, then consult AI, compare alternatives and explain any changes to their reasoning. That can encourage understanding while allowing them to explore harder material. At other times, an instructor may reasonably remove the tool so that essential skills can be demonstrated independently. The right balance depends on the task and what the learner is supposed to acquire.
A company should also remember that generative AI is not the only source of knowledge in a workplace. Experienced colleagues know about local equipment, undocumented procedures, customer relationships and past failures. Some information may be absent from the model or inappropriate to share with it. A system that appears confident cannot replace the responsibility to ask a qualified person when the stakes or uncertainty make that necessary.
LiveAIWire previously examined whether AI makes work more rewarding. The question of job quality matters here too. A beginner who gets to attempt meaningful problems and receive good feedback may gain confidence. A beginner who is expected to accept opaque suggestions at speed may learn less and feel less capable. The same technology can sit inside either arrangement.
What the Evidence Actually Changes
The striking result is not a declaration that AI has replaced apprenticeship. It is experimental evidence that access to generative AI can extend the range of tasks some final-year IT apprentices successfully attempt. The researchers also found no immediate comprehension penalty on their tested measure. Those results justify further study of how learning pathways and supervision should adapt, without pretending that longer-term expertise or workplace performance has already been settled.
For an employer, the practical opportunity is to design assignments that stretch people while maintaining review. For a training provider, it is to teach both the underlying subject and responsible use of the assistant. For a young worker, it is a reason to learn how to challenge an AI answer rather than simply admire it. The interesting future is not one in which beginners somehow no longer need experience, but one in which well-designed assistance helps them earn it through better work.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.