AI Should Not Give Answers, but Ask Questions (Part 2): From Answer Machine to Questioning Coach

AI's role: from answer machine to questioning coach

The rapid development of generative AI brings new possibilities for learning, and new concerns. The most common worry is that when AI can hand over instant, convenient answers, students may offload the cognitive work to it. One study found that frequent users of generative AI showed "significantly lower cognitive engagement": less reflection, reduced need to understand, weaker critical thinking — a "cognitive debt cycle".

The key, however, is how AI is used. Positioned mainly as an answer provider, it will not deliver learning outcomes. Designed as a questioner and a trigger for retrieval practice, it can become a tool that deepens learning. Retrieval practice normally requires the learner to volunteer for it, and that self-motivation is exactly what many students lack. Students rarely initiate free recall on their own. AI can fill that gap.

What AI-generated practice does better

AI does not merely "set questions" — the way it sets them can be far more precise and targeted than traditional exercises.

Personalised progressive difficulty. Traditional textbook exercises have fixed difficulty — either too easy or too hard. AI can adjust dynamically based on performance, keeping students in the zone where the task "slightly exceeds current ability but is achievable with effort" — the ideal state of the zone of proximal development.

Trap questions for misconceptions. The common student state is not "I know I can't do this" but "I thought I could". AI can design questions aimed at common misconceptions, testing whether a student truly understands or merely memorised a surface pattern. A student learning matrix multiplication gets a question where row-times-column and column-times-row give different answers, forcing them to face whether they truly told them apart. These are not there to trip students up, but to expose weaknesses in a safe environment — to find out what they do not understand before the exam does.

Simulated exam practice. AI can simulate real test conditions: time limits, randomised question order, no hints, mixed chapters. This front-loads the exam experience and makes retrieval practice resemble real retrieval conditions, rather than extracting from memory while it is still fresh. It creates room for spaced practice and interleaved practice, both of which are shown to significantly improve long-term retention. This kind of personalised mock exam is hard to achieve in traditional teaching.

Do not over-generate practice

AI-generated practice is convenient, but convenience is not ease. Generating a set of exercises takes seconds; a student may need hours to complete them and understand every place they got stuck. If cheaper question generation means piling on more and more exercises and tests, it becomes exploitation. When practice volume exceeds what a student can digest, learning degrades into mere compliance. The best method is worthless if the student does not want to learn. If a student lacks the motivation to sustain the effort, no method however scientific will be adopted. When learning is reduced to a sequence of quizzes, scores, and progress tracking, students easily lose their intrinsic motivation; learning becomes external pressure rather than curiosity and desire. Motivation is not an add-on — it is the engine of learning. When designing AI-assisted learning, motivation must sit at the core: balance quality against quantity.

Conclusion

Effective learning is input and output, processed and re-processed cognitively. When AI is properly designed as a questioner rather than an answer provider, it will greatly improve learning outcomes. AI's convenience should not become a justification for exploiting students — motivation is the core of learning.