Interleaving, the practice of mixing different problem types or topics within a single study session, produces a cognitive benefit that blocked practice (studying one type at a time) cannot replicate: it forces discrimination. In blocked practice, the problem type is given by context—all problems in this section are quadratic equations, so apply the quadratic formula. The learner never has to determine what kind of problem she faces. In interleaved practice, each new problem could be any of several types, forcing the learner to identify the category before selecting a strategy. This additional categorization step—determining which approach applies to which problem—builds the flexible diagnostic skill that defines competent performance outside the practice environment. Studies of mathematics learning, sports training, and category acquisition consistently show that interleaved practice produces worse performance during training (more errors, slower completion) and better performance on delayed tests, especially when those tests require transfer to novel contexts. The effect is a textbook instance of the performance-learning dissociation that structures Bjork's entire framework.
The mechanism is discriminative learning. Blocked practice allows students to solve problems correctly without learning when to apply each solution type. Interleaved practice forces the learner to develop the if-then rules (if the problem looks like this, then apply this approach) that are the hallmark of adaptive expertise. A mathematics student who has practiced blocked problems can execute algorithms competently but may struggle to identify which algorithm a novel problem requires. An interleaved student develops the categorization skill that precedes algorithm selection. The difference becomes visible on tests presenting mixed problem types without category labels—the condition that most closely resembles real-world application.
The phenomenology of interleaving is consistently negative: it feels harder, produces more errors, and generates lower confidence than blocked practice. Students actively prefer blocked practice and report that it produces better learning. The preference and the self-assessment are both wrong—blocked practice produces the fluency illusion of ease during study that metacognitive monitors mistake for effective learning. Interleaving disrupts fluency, and the disruption is cognitively valuable. Each context switch forces the brain to reload the appropriate framework, retrieve the relevant rules, and actively select among competing approaches. This constant cognitive work is the source of the learning benefit and the source of the learner's dissatisfaction. The intervention that works best feels worst.
AI tools eliminate interleaving by providing type-specific solutions. When a developer asks Claude to debug a specific error, Claude identifies the error type and provides a targeted fix. The categorization step—determining whether the bug is a type error, logic error, concurrency issue, or architectural flaw—is performed by the machine. The developer receives a solution optimized for the specific problem type without having exercised the discrimination that would have built diagnostic expertise. Over hundreds of such interactions, the developer becomes expert at evaluating presented solutions and novice at the diagnostic reasoning that precedes solution selection. The expertise that never forms is the expertise of knowing which tool to reach for before you reach.
Educational applications reveal the effect's power and its resistance to adoption. Kornell and Bjork's 2008 study of category learning showed that interleaved examples produced superior discrimination ability—learners could classify novel instances more accurately—yet students overwhelmingly preferred blocked examples and judged them more effective. The preference persisted even after students experienced their own superior performance on interleaved tests. The metacognitive illusion was strong enough to override direct evidence. This resistance suggests that widespread adoption of interleaving—or any desirable difficulty—requires institutional commitment that overrides individual preference, because individuals, left to choose, will select the conditions that feel effective rather than the conditions that are effective. In AI contexts, the overriding institution must be tool design, organizational policy, or individual discipline informed by evidence.
The effect emerged from multiple research streams in the 1980s and 1990s. Studies of motor learning (variable practice in sports), category acquisition (interleaved versus blocked examples), and mathematics education (mixed versus blocked problem sets) independently converged on the finding that mixing impairs training performance while enhancing retention and transfer. Bjork and colleagues formalized interleaving as a desirable difficulty in the 1990s, connecting it to the broader framework through the mechanism of effortful processing: interleaving forces cognitive work (category discrimination, strategy selection) that blocked practice allows the learner to bypass.
Forces discrimination, not just execution. Interleaving requires learners to identify problem types before solving them, building the diagnostic skill that blocked practice (where type is given by context) does not develop.
Impairs performance, enhances learning. Mixed practice produces more errors and slower completion during training, yet produces better retention and transfer on delayed tests—a paradigmatic instance of the performance-learning inversion.
Students prefer blocked despite evidence. Even after experiencing their own superior performance on interleaved tests, learners report preferring blocked practice, revealing that metacognitive illusions resist correction by experience.
AI provides type-specific solutions. Tools that categorize problems and deliver targeted fixes eliminate the discrimination step, developing users who can evaluate solutions but not diagnose problems independently.