The learner answers correctly five times in a row.
The teacher relaxes. The dashboard turns green. The training session looks successful.
A week later, the learner cannot reconstruct the idea without the same cues.
Nothing mysterious has happened. The session measured performance under the conditions of practice. The later failure exposed what had, or had not, become durable learning.
The distinction between learning and performance is one of the most consequential and neglected ideas in learning science. Nicholas Soderstrom and Robert Bjork's 2015 integrative review states the problem directly: what we can observe during instruction or training is performance, and performance can be an unreliable index of the relatively long-term changes that constitute learning.
This page uses the canonical title Deep learning versus performance. “Deep learning” here is shorthand for learning that endures and can support later use. It does not refer to deep neural networks in artificial intelligence. It also should not be confused with the separate educational literature on deep versus surface approaches to learning.
The scientific construct underneath the title is the learning–performance distinction.
Performance is visible; learning is inferred
Learning is not directly observable at the moment it occurs.
Researchers infer it from changes that persist after the immediate conditions of practice have changed. Delayed retention tests, transfer tasks and re-learning measures are therefore essential when the claim concerns durable learning rather than momentary success.
Performance is different. It is what a person does now, under the current cues, feedback, assistance, fatigue, motivation and task structure.
Performance matters. A pianist must perform. A clinician must act. A child solving a problem needs some present success.
The error begins when present performance is treated as proof that the underlying capability will remain available later.
The distinction predates current education debates
The learning–performance problem has roots in early animal learning and motor-skill research. Studies showed that learning could occur without immediate changes in observable performance, and conversely that manipulations could improve practice performance without producing stronger later retention.
Over decades, similar patterns appeared across verbal and motor learning. Guidance can make acquisition easier while weakening later independent performance. Massed practice can create rapid fluency while spaced practice supports better retention. Blocked practice can make a training session look orderly while interleaving or contextual interference may support later discrimination and transfer under some conditions.
These examples differ in mechanism. Their common lesson is methodological: the curve visible during practice is not automatically the learning curve.
Soderstrom and Bjork's review matters because it integrates those traditions rather than reducing the distinction to one intervention.
Why smooth practice can be seductive
Humans are sensitive to fluency.
When an answer comes quickly, a sentence feels familiar or a sequence repeats successfully, that ease provides information about the current state of processing. The problem is that learners can overgeneralise from that information to the future.
“What can I do with these supports, right now?” quietly becomes “What will I be able to do next week?”
Those are different questions.
The same error can occur institutionally. A teacher sees correct responses after guided examples. A company sees employees complete a simulation with prompts. A software platform sees high task accuracy while hints remain available. Each may report successful learning when the system has mainly demonstrated successful supported performance.
Metacognition enters here because learners and decision-makers need to infer an invisible future state from visible present behaviour.
Conditions that hurt performance can sometimes help learning
The literature on desirable difficulties is often introduced at this point, but it needs discipline.
Spacing, interleaving, retrieval attempts and reduced guidance can make practice feel harder and lower short-term performance under some conditions. They may also improve later retention or transfer.
This does not mean difficulty is good in itself.
A task can be confusing, overloaded, inaccessible or badly taught. Performance can fall because the learner has insufficient prior knowledge, because instructions are unclear or because the environment has introduced irrelevant friction. None of those difficulties becomes desirable by being unpleasant.
A difficulty earns the label only when it engages processes that support the desired later learning outcome for the learner and task in question.
That qualification keeps the learning–performance distinction from turning into an ideology of struggle.
Conditions that improve performance can sometimes hide dependency
The converse is equally important.
Prompts, worked examples, demonstrations, answer keys and software assistance can improve performance immediately. Often that is exactly why they are useful. Novices may need guidance to build accurate models before independent practice is productive.
But assistance can create performance support without guaranteeing capability transfer.
If the learner succeeds only while the support remains visible, the training has not yet shown that the skill is independently available.
This does not make guidance bad. It means guidance should be evaluated against its intended function. Some support is scaffolding that should later fade. Some is permanent infrastructure that the learner will always use. The distinction changes what “learning” needs to mean.
A pilot trained to use a cockpit instrument does not need to perform as though the instrument does not exist. A student expected to reason when no answer generator is available may need a different degree of internalisation.
Deep learning is not one score
Durability and transfer are not interchangeable.
A learner can retain a fact but fail to apply it in a novel context. A person can transfer a principle to a similar problem but fail when surface features change. A motor skill can persist over delay while remaining context-specific.
For this page, “deep” therefore should not be operationalised as one hidden essence. It points to a family of stronger outcomes: retention after meaningful delay, flexible retrieval, conceptual integration, appropriate transfer and reduced dependence on the exact cues present during acquisition.
Which of these matters must be specified in the research or educational design.
Calling all of them deep learning without measurement would simply replace one vague success label with another.
Delayed testing changes the epistemology of education
An immediate post-test asks what is available at the end of instruction.
A delayed test asks what remains after time has done some of its work.
That delay is not a nuisance. It can be the very condition that separates accessibility created by recent exposure from more durable learning.
The appropriate delay depends on the claim. If a safety procedure must remain usable for a year, a test ten minutes after training provides only partial evidence. If a concept is needed as the basis of tomorrow's lesson, a shorter interval may still matter.
The Institute treats time-to-need as a useful synthesis variable: how long must the capability remain available before the environment will reliably refresh or support it? That phrase is not a standard psychometric construct. It is a design question built from the retention principle.
Transfer asks whether learning survived a change in context
Retention can be tested with the same kind of task used during learning.
Transfer goes further. It changes something relevant: the examples, surface features, response format, problem structure or context of use.
Transfer tests are therefore central when the educational claim concerns understanding rather than reproduction.
Yet “far transfer” is often overclaimed. Similarity is multidimensional, and evidence of transfer to one new task does not prove a general ability that will travel anywhere.
A rigorous report should describe exactly what changed between learning and test.
The learning–performance distinction becomes especially useful here because high acquisition performance can coexist with brittle knowledge that depends on familiar cues.
The distinction protects against teaching to the dashboard
Educational systems increasingly measure what can be captured quickly.
Completion rates, quiz accuracy, time on task and immediate improvement are attractive because they arrive while the programme is still running. Durable outcomes arrive later, often after budgets, cohorts and staff have moved on.
This creates a structural temptation to optimise for performance metrics.
The same pattern appears in workplace learning. Employees complete modules, pass embedded questions and receive certificates. Those data show participation and immediate task performance. They do not necessarily establish retained judgment in the conditions that matter.
An institution serious about capability needs at least some outcomes that occur after the support, novelty and recency of training have faded.
That is not a demand to delay every assessment. It is a demand to match evidence to claims.
AI assistance makes the distinction newly visible
A language model can improve a person's performance immediately.
It can draft, summarise, translate, propose arguments, detect errors and generate code. In many contexts, augmented performance is the goal. If a person and tool together produce a better result, that matters.
But augmented performance and human learning are different outcomes.
A team can become more productive while its members become less able to reconstruct the work without the system. A learner can submit better prose while practising less sentence formation. A novice can solve harder problems with an assistant while remaining unable to diagnose the assistant's mistakes.
The evidence from learning science does not license the claim that AI causes cognitive decline. That would require direct longitudinal research.
It does license a measurement warning: do not infer internal capability from tool-supported performance unless internal capability is what you actually measured.
This is one of the clearest Inner Technology implications of the learning–performance distinction.
Sometimes external support is part of the competence
There is an opposite error to avoid.
Human competence has always included tools, notes, colleagues, instruments and institutions. It would be artificial to define every capability as what a person can do naked of all external support.
The relevant question is which parts of a task need to remain internal for safe, adaptive and responsible use of the external system.
A calculator can replace mental arithmetic in many contexts without erasing the need for number sense. Navigation software can guide routes while still leaving value in spatial orientation. AI can produce candidate reasoning while increasing the importance of checking assumptions, evidence and consequences.
The Institute calls this capability partitioning: deciding which functions may be reliably externalised and which need internal competence because they are required to supervise, recover from or challenge the external system. This is Institute synthesis, not a validated learning-science construct.
Measurement design should separate acquisition, retention and transfer
A strong study or training evaluation can ask at least three questions.
First: how did performance change during acquisition?
Second: what remained after a delay?
Third: what transferred when cues or tasks changed?
These outcomes can point in different directions. Reporting only one can create a misleading success story.
Researchers should also examine whether later performance depends on feedback, whether prompts are present, and whether participants know what will be tested. In applied contexts, attrition and unequal opportunity to practise after training can further complicate delayed outcomes.
Learning is not a pure substance waiting to be measured. It is inferred through tasks embedded in conditions.
The individual can learn to ask a better question
A learner often leaves study asking, “Did that go well?”
The learning–performance distinction suggests a more useful set of questions.
Could I reconstruct this later? Could I recognise when the principle applies? Could I use it when the wording changes? Which parts currently depend on cues I may not have?
Those questions shift metacognition from a feeling of fluency toward an evidence problem.
They also make room for humility. The learner cannot know future retention with certainty. They can, however, design tests that reveal more than immediate ease.
Retrieval after delay is one such test. Explaining a concept without the source is another. Applying it to a new example is another.
The goal is not constant self-examination. It is better calibration between the claim “I learned this” and the evidence available for that claim.
What the evidence does not show
The learning–performance literature does not show that immediate performance is unimportant.
It does not show that every condition that depresses practice performance improves learning.
It does not show that guidance, scaffolding or external tools are inherently harmful.
It does not justify calling any low-fluency experience a desirable difficulty.
It does not establish one universal definition of “deep learning” across domains.
It does not show that delayed retention automatically implies broad transfer.
And it does not establish that AI-supported performance weakens independent capability. That is a contemporary empirical question requiring direct evidence, not an inference to be smuggled in from older laboratory paradigms.
Open questions
How should learning be measured when external tools will remain permanently available? Which internal capabilities are necessary for supervising AI-assisted work? How long should delays be when the target is professional competence rather than classroom recall? Which acquisition difficulties improve transfer for novices, and which merely overload them? How can institutions afford to measure outcomes after the programme has formally ended?
The deepest question may be administrative rather than cognitive.
What does an institution choose to count as success when the easiest evidence to collect is also the evidence most vulnerable to confusion with performance?
The green dashboard is not the end of the story
The learner got five answers right.
That is real data. It may be encouraging. It may show that the explanation worked in the moment.
It is not yet proof of what will remain.
Durable learning begins to reveal itself when the cues thin out, time passes and the person has to meet the idea again without the exact conditions that first made success possible.
Research references
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
Schmidt RA, Bjork RA. New conceptualizations of practice: Common principles in three paradigms suggest new concepts for training. Psychological Science. 1992;3(4):207–217. https://doi.org/10.1111/j.1467-9280.1992.tb00029.x
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
Dunlosky J, Rawson KA, Marsh EJ, Nathan MJ, Willingham DT. Improving Students’ Learning With Effective Learning Techniques. Psychological Science in the Public Interest. 2013;14(1):4–58. https://doi.org/10.1177/1529100612453266