Practice is going badly.
The answers come more slowly than yesterday. Items are mixed instead of grouped. The learner has to retrieve something that was easy when it was still on the page. Errors appear.
Then, days later, more of the learning remains.
This reversal between how practice feels and what practice leaves behind sits near the centre of Robert A. Bjork's research programme. Bjork is Distinguished Professor Emeritus of Psychology at UCLA, where the Bjork Learning and Forgetting Lab has studied human learning, memory, forgetting and the conditions that support durable competence.
His name is strongly associated with desirable difficulties, a phrase that has travelled far beyond its technical home. The phrase is memorable enough to invite abuse.
Bjork's contribution is not the claim that harder is better.
It is the much more demanding claim that learning must be evaluated by the future performance it produces, and that some conditions which make acquisition more difficult can improve that future performance.
Learning and performance are different variables
Bjork's work repeatedly returns to the distinction between learning and performance.
Performance is what can be observed during practice. Learning is a more durable change inferred from retention and transfer after time or conditions have changed.
The distinction explains why a learner can look excellent during blocked, repeated or heavily guided practice and then struggle later. It also explains why a learner can look less fluent during spaced, interleaved or retrieval-rich practice and retain more.
Nicholas Soderstrom and Bjork's 2015 integrative review brought evidence from motor and verbal learning together around this point: performance during acquisition can be a poor guide to long-term learning, and some manipulations have opposite effects on the two.
The implication is methodological before it is pedagogical.
If the outcome is durable learning, then immediate success cannot be the only measure.
Desirable difficulty is an outcome-dependent concept
Bjork discussed “desirable difficulties” in his 1994 chapter on memory and metamemory in human training. The term refers to learning conditions that make acquisition more effortful or error-prone but improve later retention or transfer under appropriate circumstances.
Examples often include spacing, interleaving, variation and retrieval practice.
But a difficulty is not desirable because it is difficult.
It is desirable because the difficulty engages processes that produce better later learning for this learner, this material and this goal.
If the learner lacks prerequisite knowledge, the same manipulation may simply prevent learning. If the delay is too long, retrieval may fail without useful reconstruction. If interleaving makes categories impossible to discern, difficulty has exceeded its function.
The concept therefore contains its own boundary condition.
“Desirable” is not a feeling. It is an empirical claim about consequences.
Forgetting is not merely failure
Bjork's research programme also helped reframe forgetting.
Everyday intuition treats forgetting as the opposite of learning. Yet forgetting can change the conditions under which later retrieval occurs. A memory that is less immediately accessible can require more retrieval work when encountered again, and successful retrieval can strengthen later access.
Bjork and Elizabeth Bjork's “new theory of disuse” distinguished storage strength from retrieval strength. A memory can have high storage strength while being temporarily difficult to retrieve, and retrieving an item when retrieval strength is lower can produce substantial learning gains.
The theory is more specific than the slogan “forgetting is good.”
Forgetting can be costly, dangerous or final. The useful insight is that current accessibility and underlying learning history are not identical.
This helps explain why cramming can create high retrieval strength in the moment without the durable storage needed later.
Metamemory is where the learner gets fooled
Why do people often choose less effective study conditions even after experiencing more effective ones?
One answer is that effective conditions can feel worse.
Learners use cues such as fluency, speed and immediate success to judge whether learning is occurring. Massed study and repeated exposure can create those cues. Spacing and retrieval can interrupt them.
Bjork's work on metamemory highlights the resulting illusion: people may prefer the method that makes performance easy now because ease feels like evidence of lasting knowledge.
This is not simple irrationality. Fluency is often informative. The problem is cue validity. A cue that predicts present accessibility may be less useful for predicting future retention.
Metacognitive skill therefore includes learning which cues deserve trust for which prediction.
The laboratory principle became an educational provocation
The value of Bjork's work to education is partly that it destabilises a familiar equation: successful teaching equals smooth learner performance.
A classroom in which every question is answered correctly may be excellently taught. It may also be so supported that students have not yet demonstrated independent retrieval or transfer.
Conversely, a classroom with productive errors may look less efficient while building stronger later discrimination.
The key word is productive.
Instructional difficulty needs diagnosis. Are students struggling because they are retrieving after delay, comparing confusable concepts and constructing explanations? Or because instructions are ambiguous, background knowledge is missing and working memory is overloaded?
Both look difficult from the outside. Only one may deserve the label.
Bjork's programme is wider than one phrase
Reducing Bjork to desirable difficulties misses the larger architecture.
His work has addressed retrieval as a memory modifier, forgetting, memory strength, metamemory, study choices, testing, spacing and how people assess their own competence. Across these topics, a recurring theme is that human memory is adaptive and dynamic rather than a storage container whose only problem is loss.
Retrieval changes memory. Context changes accessibility. Forgetting can alter the value of later practice. Subjective impressions of learning can mislead study decisions.
This wider programme makes desirable difficulties intelligible. They are not isolated hacks. They arise from a theory of how conditions during learning change later retrieval and how people often misread those conditions.
Not every celebrated “difficulty” has equal evidence
Once a concept becomes popular, examples multiply faster than validation.
Spacing and retrieval practice have substantial evidence bases. Interleaving has strong evidence in some classification and discrimination tasks but important boundary conditions. Disfluency manipulations, such as making text harder to read, have produced much less consistent support than the early popular enthusiasm suggested.
The category “desirable difficulties” should therefore not be treated as one intervention with one effect size.
Each manipulation needs its own evidence review.
This is an important scholarly restraint: a theoretical family resemblance does not transfer the evidential strength of one member to another.
The learner is not the only level that can be miscalibrated
Institutions also confuse performance with learning.
A training programme can report high completion and quiz scores while never measuring retention. A school can optimise for examination performance and then infer durable understanding. A company can introduce an assistant that raises output quality and call the improvement workforce upskilling.
Each inference may be wrong because the metric captures supported performance rather than lasting capability.
The Institute extends Bjork's learning–performance distinction into an institutional question: what evidence would still count if the support, recency or rehearsal context were removed?
That is Institute synthesis, not a construct attributed to Bjork.
It is useful because it turns a cognitive principle into a governance test for training claims.
AI can make acquisition performance almost frictionless
Generative AI can remove many difficulties from cognitive work.
It can complete sentences, generate examples, supply explanations, retrieve facts and scaffold problem solving. In many settings that is a legitimate improvement. The goal may be augmented performance, not unaided recall.
But if the goal includes human learning, the removal of difficulty changes the practice conditions.
An AI tutor that gives an answer immediately creates a different learning event from one that waits for a prediction or retrieval attempt. A coding assistant that supplies every implementation step creates a different practice history from one that prompts diagnosis before revealing code.
The science of desirable difficulties does not tell us one correct interface.
It tells us not to assume that lower friction during practice and stronger human learning are the same outcome.
That distinction is likely to become more important as external competence becomes easier to rent moment by moment.
There is a humane reason not to romanticise struggle
Educational language about difficulty can become morally loaded.
Learners who struggle are told the struggle is good for them. Support is framed as weakness. Accessible design is treated as dilution. Structural barriers disappear behind a psychology of grit.
Bjork's concept does not justify any of that.
A desirable difficulty is defined by improved learning consequences, not by suffering, austerity or virtue. If a barrier excludes a learner without improving the target capability, it is simply a barrier.
This boundary is especially important for disability, language differences, fatigue, stress and unequal prior preparation.
Evidence-based difficulty should be functional, not punitive.
The researcher's influence is strongest when it remains falsifiable
Bjork's ideas are powerful because they produce testable reversals.
A manipulation may reduce immediate performance. Does it improve delayed retention? Does it improve transfer? For whom? Under which schedules? At what error rate? With what prior knowledge?
Those questions keep the concept scientific.
The moment desirable difficulty becomes an all-purpose explanation for why an unpleasant method must secretly be good, it stops doing useful work.
A mature application therefore records both the difficulty and the later outcome.
Retrieval practice shows why the timing of success matters
One of the clearest examples in the Bjork tradition is retrieval practice. Attempting to retrieve information can produce more durable learning than additional study, even when restudy feels easier and produces fewer errors during the practice period. But the benefit depends on what happens around the attempt. Feedback, retrieval success, spacing and material type all matter.
This example reveals why “difficulty” is an incomplete description. A retrieval attempt is not useful because it feels unpleasant. It is useful when the act of reconstructing knowledge modifies later accessibility and when errors are managed in ways that support correct learning.
The same logic applies to spacing. Waiting changes retrieval strength. Too little delay can leave practice redundant; too much delay can make the earlier learning effectively unavailable. The scientifically relevant variable is not maximum struggle but the relation between current accessibility and the learning operation created by the next encounter.
This is why Bjork's ideas resist reduction to classroom theatre. An instructor cannot infer desirable difficulty from visible student discomfort. The later evidence has to justify the adjective.
Metacognitive illusions are partly ecological
Learners do not form judgments in a vacuum. Educational environments repeatedly reward certain cues. If exams occur immediately after intensive review, fluency can be a reasonably predictive cue. If assignments remain open-book, recognition and search may be sufficient for the task. If every practice block repeats one problem type, blocked performance can look like competence because the environment keeps confirming it.
Problems arise when those cues are carried into a different future. The learner assumes that what predicted success in the practice ecology will predict success after delay or transfer.
This makes metamemory partly an environment-design problem. Better calibration may require not only teaching learners about illusions but giving them experiences in which delayed retrieval and transfer provide corrective evidence.
The Institute's synthesis is that a learning environment teaches a theory of learning implicitly through the metrics it rewards. If it rewards only smooth acquisition, it may train the learner to trust smooth acquisition. That claim needs institutional research, but it follows directly from taking cue validity seriously.
What the evidence does not show
Bjork's work does not show that harder learning is always better learning.
It does not show that errors are beneficial without feedback or opportunity for correction.
It does not show that all desirable-difficulty manipulations have equal evidence.
It does not show that immediate performance is irrelevant.
It does not show that guidance should be withheld from novices.
It does not show that a method beneficial for retention automatically improves far transfer.
It does not show that removing AI assistance will necessarily improve human capability.
And it does not justify using cognitive science to rationalise avoidable institutional hardship.
What remains alive in Bjork's programme
The durable contribution is a set of disciplined suspicions.
Be suspicious when learning looks too easy to measure only because the answer was just present.
Be suspicious when a training intervention is praised entirely on performance during training.
Be suspicious when subjective fluency is treated as proof of future access.
And be equally suspicious when difficulty is celebrated without evidence that it improves the outcome that matters.
Between those suspicions is a more serious view of learning: one that takes time, retrieval conditions, forgetting, transfer and human self-misperception seriously.
That is why Bjork's work remains relevant far beyond the laboratory.
Research references
Bjork RA. Memory and metamemory considerations in the training of human beings. In: Metcalfe J, Shimamura AP, eds. Metacognition: Knowing about Knowing. MIT Press; 1994:185–205. https://doi.org/10.7551/mitpress/4561.003.0011
Bjork RA, Bjork EL. A new theory of disuse and an old theory of stimulus fluctuation. In: Healy AF, Kosslyn SM, Shiffrin RM, eds. From Learning Processes to Cognitive Processes. Erlbaum; 1992.
Soderstrom NC, Bjork RA. Learning Versus Performance: An Integrative Review. Perspectives on Psychological Science. 2015;10(2):176–199. https://doi.org/10.1177/1745691615569000
Kornell N, Bjork RA. The promise and perils of self-regulated study. Psychonomic Bulletin & Review. 2007;14:219–224. https://doi.org/10.3758/BF03194055
UCLA Bjork Learning and Forgetting Lab. https://bjorklab.psych.ucla.edu/
UCLA Department of Psychology. Robert A. Bjork, Distinguished Professor Emeritus. https://www.psych.ucla.edu/faculty-page/bjork/