Desirable Difficulties in AI Tutoring: Why Helpful Systems Should Not Make Everything Easy
AI tutoring should sometimes make learning harder on purpose: spacing, retrieval, interleaving, delayed help, and fading support can improve durable learning when difficulty stays achievable.
A tutor that makes every moment easy can make learning look better than it really is
The desirable-difficulties perspective distinguishes performance during practice from durable learning. Conditions that make practice feel fluent can produce weaker later retention, while spacing, interleaving, retrieval, and other effortful conditions can improve long-term learning.[1] This creates a design tension for AI tutors. The model is optimized to be helpful and can instantly summarize, explain, solve, translate, and simplify. If it removes every obstacle at the moment difficulty appears, learners may experience smooth sessions without building the ability to retrieve, choose, and execute on their own.
Friction is not automatically bad UX in a learning product
The correct question is whether the friction recruits a useful learning process. Waiting through a slow interface is waste; trying to recall a concept before seeing the hint can be productive.
Spacing and interleaving are examples of productive difficulty that AI can schedule automatically
The IES practice guide recommends spacing learning over time and interleaving worked examples with problem-solving exercises.[2] These methods are inconvenient for learners to manage manually because they require returning to older material at the right time and mixing topics rather than finishing one block completely. AI can handle the scheduling while leaving the cognitive work to the learner. A language tutor can reintroduce vocabulary after delay inside a new story; a coding tutor can mix debugging, syntax, and design tasks so the learner must first diagnose what kind of problem they are facing.
The system should explain why the experience feels harder
Learners may interpret slower performance as worse learning. A brief explanation can normalize effort and prevent them from choosing only the easiest practice mode.
Retrieval is desirable difficulty because the learner must reconstruct knowledge instead of recognize it
Karpicke and Roediger’s experiments demonstrated the importance of retrieval for later retention.[3] An AI tutor can exploit this by withholding the explanation for a few moments and asking for an attempt first. Even partial recall gives the tutor diagnostic evidence and gives the eventual feedback a specific target. The design should make “try before reveal” easy and low stakes. If learners fear penalty for error, they may avoid retrieval and request the answer immediately, defeating the purpose.
Difficulty needs an escape hatch
A learner who is completely stuck should be able to request a cue, example, or explanation. The goal is effortful success, not endurance for its own sake.
Feedback timing should be chosen for the learning objective rather than fixed by chat conventions
Kulik and Kulik’s meta-analysis found that the effects of immediate versus delayed feedback varied across learning contexts, with classroom studies often favoring immediate feedback while some experimental conditions produced different patterns.[4] AI products should therefore avoid assuming that instant correction is always optimal. Immediate feedback is valuable for preventing persistent misconceptions or unsafe practice, while a short delay can preserve reflection or allow a learner to inspect their own work first. The tutor can ask “What would you change before I comment?” and only then provide targeted feedback.
The interface can separate answer checking from explanation
A learner may need to know whether an answer is correct immediately while still benefiting from generating their own rationale before the model supplies a full solution.
The same support that helps a novice can become redundant for an expert
Research on the expertise-reversal effect shows that instructional methods beneficial to beginners can become unnecessary or even counterproductive as prior knowledge increases.[5] AI should therefore treat support as temporary. Worked examples, detailed prompts, step-by-step checklists, and explicit cues may be valuable early, then should fade as competence grows. A tutor that keeps explaining elementary steps forever can slow advanced learners and encourage dependence. Adaptive difficulty is partly the art of removing support at the right time.
Desirable difficulty must be calibrated because many difficulties are simply undesirable
Not every struggle promotes learning. Ambiguous instructions, inaccessible interfaces, missing prerequisites, anxiety-inducing timers, and factual errors can consume attention without building skill. The tutor needs evidence that the learner has enough knowledge to respond successfully to the challenge. If repeated attempts show no progress, the system should change the representation, revisit prerequisites, or provide more structure. Difficulty is useful only when it sits inside a recoverable learning loop.
Gamification should reward engagement with difficulty rather than avoidance of it
Traditional points systems can accidentally teach learners to farm easy items for perfect scores. A better design can recognize behaviors such as attempting before asking for help, revising after feedback, returning to a difficult concept after spacing, or completing a transfer task without cues. The system can celebrate “less help needed” or “recalled after seven days” instead of only total questions answered. That makes the progression model reflect increasing independence.
The best AI tutor manages the boundary between assistance and effort
Generative AI gives education products extraordinary power to remove friction. Learning design requires restraint. The tutor should make administrative work easier, access faster, explanations more available, and practice more personalized—while protecting the moments when the learner must actually remember, decide, explain, and perform. The result may feel less magical than an answer machine because the AI sometimes refuses to do the interesting work for the user. That restraint is a feature: the product is optimizing for capability after the session, not just satisfaction during it.
Works Cited
- 01Bjork & Bjork — Making Things Hard on Yourself, but in a Good Way bjorklab.psych.ucla.edu
- 02
- 03
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.