What can we still do after AI has helped?

Imagine an internal learning session. A participant works through a difficult case, uses AI assistance and presents a convincing solution. One observation is still missing from the evaluation of the session: what can she explain and decide for herself when she encounters the next, slightly different task?

This question matters in clinical education and leadership development alike. People who will later take responsibility need opportunities to build their own understanding of a problem. AI can enrich those opportunities. It can also perform the very steps through which someone was meant to learn. The result depends partly on how assistance is designed and when it is introduced.

Two studies, two learning environments

Bastani and colleagues studied almost a thousand pupils at one Turkish school during mathematics practice in autumn 2023. Ordinary GPT-4 assistance improved practice performance but was associated with an average score 17% lower than the control group in the immediately subsequent unassisted assessment. A pedagogically constrained version largely avoided that disadvantage, without demonstrating a benefit over the control group in that assessment. Published in 2025, the study covers four sessions and does not test long-term professional competence. [1]

Kestin and colleagues reported a different experiment in 2025. In autumn 2023, 194 analysed students in a Harvard physics course completed two lessons using alternating instructional formats. Carefully prepared AI tutoring produced stronger immediate assessment results than the active classroom instruction studied. The experiment covers selected content and short learning periods. It does not establish a general advantage over human teachers. [2]

These studies should neither be combined into a single success rate nor compressed into a verdict on every form of learning with AI. They compare different populations, subjects and instructional formats. For practice, they open a more precise question: which mental activity does the assistance perform, and which activity does the learner actually practise?

Set the learning objective before choosing the tool

For an organisation, I would begin with a sentence: after this session, the person should be able to justify a particular decision in a particular situation. The question of what support to provide comes next.

Consider a fictional leadership-development example. A team needs to assess a proposal to introduce new software. The learning objective is to identify missing assumptions and formulate a justified follow-up question. Asking AI to find every weakness immediately may produce an excellent text. Whether participants have learnt to recognise a missing assumption themselves remains unclear.

A different sequence creates another observation. Each person first records their own decision and one uncertainty. An approved AI tool can then ask questions or offer a selected hint. Finally, participants work through a modified case. The differences between these three pieces of work become the basis of the discussion.

This sequence is my design proposal. The studies cited have not tested its effect in a German leadership programme. Its initial value lies in making the intended capability visible and the effect of assistance open to examination.

Leave a small trace of the reasoning

Keeping only the final answer erases the route towards it. For learning, a little additional documentation can help: the initial hypothesis, the decisive hint and the explanation for the final decision. A few sentences are sufficient.

Such a record should serve learning. It should not quietly be repurposed for monitoring individual performance. Organisers should explain in advance who sees the work, how it will be discussed and when it will be deleted. Fictional cases allow groups to practise without placing patient information or confidential business matters into an assistance system.

Clinical learning content brings an additional professional responsibility. The education team must define the tasks, expected reasoning and appropriate boundaries. A language model may help phrase exercise materials where a suitable environment is available. Approving the content and assessing clinical competence remain responsibilities for appropriately qualified people.

Which capabilities should remain available without assistance?

Complete independence from every aid is not a sensible general objective. Professional work uses reference materials, collaboration and technical support. The organisation needs to decide which capabilities someone should have immediately available, and where careful reference checking or timely help-seeking is the desired behaviour.

I would prepare that decision through three questions. How quickly must the person respond? What would follow from an incorrect first step? And how will the person recognise that assistance is needed? These questions provide a reasoned basis for selecting the capabilities to practise independently.

In an organisational exercise, recognising that an approval is missing might be knowledge someone needs immediately. The precise person responsible can then be found in an approved directory. The assessment would examine whether the participant notices the need for clarification and uses the appropriate route to obtain it. This makes the difference between appropriate help-seeking and unnoticed dependence concrete.

The boundary depends on the profession and the task. A general AI learning session can help prepare that distinction, but cannot certify professional competence. Those assessing such competence need suitable criteria, enough observations and the appropriate responsibility.

Learning needs a later observation

A learning event often ends administratively with its final agenda item. A later contact is more useful for examining what capability remains. After an agreed interval, the group receives a short new task. It requires the same underlying idea, even though the surface details have changed.

The interval and difficulty should suit the capability. The later task may reveal that the original session was insufficient. That calls for targeted practice or a different explanation. An unsuccessful learning attempt provides information about the educational design, provided it does not become a premature judgement of a person's worth.

Participants should also know how the later exercise connects to the original objective. They should be able to recognise what has become easier, where they still need a reference and when they would seek a colleague's help. This makes the follow-up useful to them as well as to the people responsible for the programme.

For me, the leadership decision is concrete: when introducing AI into education, also decide how independent competence will become visible. A good assisted solution is a valuable outcome. Developing professionals brings a further responsibility: create conditions in which people can increasingly understand their own next decision.