A desirable difficulty is a condition that makes learning feel harder and slower in the moment while producing better long-term retention and transfer. The term is Robert Bjork's, and it carries the most counter-intuitive finding in the science of learning: the conditions that make studying feel effective are frequently the ones that produce the least durable learning.
Definition#
Desirable difficulty: a manipulation of learning conditions that impairs immediate performance while improving long-term retention and the ability to apply what was learned in a new context. The difficulty is desirable because the effort of overcoming it is what produces the durable learning, and difficult because it feels like failure while it is happening.
The core finding#
Robert and Elizabeth Bjork's work established a distinction that most people never make: performance is what you can do now, and learning is what you will still be able to do later. They come apart routinely. Conditions that raise performance during study, such as re-reading, massed practice, and having material presented fluently, often lower learning. Conditions that depress performance during study, such as retrieval practice, spacing, interleaving and varied conditions, often raise it.
The practical consequence is that learners systematically choose badly, because they judge their progress by how fluent the material feels. Fluency feels like mastery and frequently is not.
Why this is the mechanism underneath most of this research#
A large language model is, among other things, a very effective remover of desirable difficulty. It removes the retrieval attempt by supplying the answer. It removes the struggle to structure a problem by presenting it structured. It removes the generation effort by generating. And it does all of this while the work still gets done, which means performance stays high and only learning falls.
That is the pattern the strongest field experiment found. Bastani and colleagues gave nearly a thousand high-school mathematics students a GPT-4 tutor. Grades rose while the tool was available, by 48 per cent with a plain chat interface and 127 per cent with a guardrailed tutor. When access was withdrawn, the plain-interface group scored 17 per cent lower than students who never had access. The guardrailed version, which made students do the work, largely removed that harm.
Same model, opposite outcomes, decided entirely by whether the interface preserved the difficulty or removed it.
What is uncertain, and the important caveat#
Not all difficulty is desirable. Bjork's own framing is explicit about this: difficulty that exceeds what the learner can overcome produces failure rather than learning, and the boundary depends on prior knowledge. Confusion, poor instruction and cognitive overload are undesirable difficulties, and dressing them up as pedagogy is a real failure mode in the applied literature.
There is also a fair objection to the AI application. The Bastani result is one study, one subject, one age group, unreplicated. Whether the same effect holds for adults, for professional work, or over longer horizons is genuinely unknown, and this research lists it as such in what we actually know.
Why the learning question has no single answer#
Desirable difficulty is the reason "does AI help or harm learning" has no single answer, and the reason the question is badly posed. The variable is whether the interaction preserves the effortful step or performs it for you. The model barely matters.
That generalises well beyond education. In professional work, the repetitions that build judgement are almost always the difficult, unglamorous ones, and they are the first candidates for automation because they look like cost. Removing them raises this quarter's output and lowers the capability of everyone who would have done them, which is the argument in the missed reps restated in learning-science terms.
The design question follows: which difficulties in this work are load-bearing, and which are merely friction? They look identical on a process map and they are not the same thing at all.
Related SuperSkills research#
On learning with AI, how humans learn with AI. On assessment, assessing students when AI can do the assignment. On the professional version, the missed reps and deskilling. On the individual habit, using AI without dependency. See cognitive load.
Key sources
- Bjork, R. A. and Bjork, E. L. Desirable difficulties in theory and practice.
- Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
- Ericsson, K. A., Krampe, R. T. and Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3).
About this definition#
Desirable difficulty is Robert Bjork's term and is not a SuperSkills coinage. It is defined here because it is the mechanism underneath a large part of this research and is usually referenced without being explained. Rahim Hirji is the author of SuperSkills (Kogan Page, 2026).
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.
Cite this
Hirji, R. (2026). What is desirable difficulty? The SuperSkills Intelligence Company. Last reviewed 26 August 2026. thesuperskills.com/research/what-is-desirable-difficulty