AI can make learning faster. It can also make practice easier to avoid.
That is the uncomfortable part.
The best uses are already obvious to anyone who has worked with a good model for more than an afternoon. AI can explain an unfamiliar idea. It can produce examples. It can generate practice scenarios, draft questions, summarise reading, role-play a difficult conversation, critique a first draft, or offer another way into a subject that previously felt closed.
For learning designers, this is not a small development. It gives us more ways to vary practice, support preparation, reduce blank-page anxiety, and help learners meet an idea from several angles.
But there is a quieter risk. AI can also let the learner skip the very friction through which capability would have formed.
The danger is not only cheating, though that matters. The deeper danger is hidden substitution. A person appears to complete the task, but the task has been quietly moved from them to the machine.
The answer is fluent. The learner is not necessarily more capable.
A 2026 experimental study by Judy Hanwen Shen and Alex Tamkin makes this problem more concrete. In their study of people learning a new programming library, AI assistance could impair conceptual understanding, code reading, and debugging ability when learners relied on it in ways that reduced their own cognitive engagement. The point is not that AI always damages learning. The study also identifies more engaged patterns of AI use that preserve learning. The sharper point is that productivity and skill formation are not the same thing.
That distinction should sit near the centre of modern learning design.
AI can help someone finish. It does not automatically help them learn.
In many workplaces, finishing is what gets rewarded. The document is written. The response is drafted. The analysis is summarised. The meeting notes are cleaned up. The first-pass proposal exists. A busy person, under pressure, will understandably reach for the tool that gets them over the line.
The problem is that capability often forms before the line.
It forms while choosing what matters.
It forms while noticing what the first answer misses.
It forms while comparing alternatives.
It forms while wrestling with a constraint.
It forms while receiving feedback that is a little uncomfortable but usable.
It forms while making a second attempt.
If AI removes all of that too early, the learner may become more productive and less prepared.
This is why “AI in learning” is too blunt a phrase. The better question is not whether AI is allowed. It is what job AI is doing in the learning system.
AI can be a preparatory tutor. That may be useful.
AI can be a scenario generator. That may be useful.
AI can be a practice partner. That may be useful.
AI can be a first-pass critic. That may be useful.
AI can also become a proxy performer, a judgement substitute, or a polite machine for avoiding the learner’s own attempt.
Those are different design choices.
The learning designer’s task is to create an AI boundary. Where should AI expand access, practice volume, variation, rehearsal, and reflection? Where should it be restricted because the learner needs to show their own judgement? Where must learners disclose how they used it? Where is unaided performance still necessary? Where does a human need to see the reasoning, not only the result?
This is not anti-AI nostalgia. It is pro-capability realism.
The modern learner will use AI. So will the modern worker. Pretending otherwise is not a serious design position.
But if a course is meant to build capability, it must protect the moments where capability becomes visible. The learner still needs to make an attempt. They still need criteria. They still need feedback. They still need a second try. They still need to know what they can do when the machine is absent, wrong, incomplete, or too confident.
There is a humane version of this argument.
People do not need more artificial difficulty. Work already supplies plenty of that. But they do need useful difficulty: the kind that helps them discover what they notice, what they miss, what they can now do, and what still needs support.
AI can help create that difficulty. It can also sand it away.
Good learning design knows the difference.