Desirable Difficulties: Why Hard Learning Sticks

What Are Desirable Difficulties?

In learning science, a desirable difficulty is a condition that makes practice feel harder and slower while you're doing it, yet produces better retention and transfer days or weeks later. Robert Bjork coined the term in 1994. Today's standard examples are spacing, interleaving, retrieval practice, generation, and varied practice. The catch is that desirable difficulties lower your performance during practice, so they look like they're failing, which is exactly why almost nobody chooses them.

Bjork, a psychologist at UCLA, introduced the idea in a book chapter on training and developed it over the following decades with Elizabeth Ligon Bjork. Their lab kept hitting the same result: the manipulations that depressed performance during study were often the ones that improved it after. The best-known summary is a 2011 chapter by Elizabeth and Robert Bjork whose title states the principle outright, "Making things hard on yourself, but in a good way."

That chapter also holds the definition, and it comes with a condition the popular versions drop. Desirable difficulties, the Bjorks write, "are desirable because they trigger encoding and retrieval processes that support learning, comprehension, and remembering. If, however, the learner does not have the background knowledge or skills to respond to them successfully, they become undesirable difficulties." The second half is the one people skip: the same task can be desirable for one learner and useless for another.

Five conditions have decades of replication behind them:

• Spacing: Distributing study across separate sessions instead of massing it into one.
• Interleaving: Mixing different problem types or topics within a session instead of blocking them.
• Retrieval practice: Pulling information out of memory instead of re-reading it in.
• Generation: Producing an answer before you're shown it.
• Varied practice: Practicing under changing conditions instead of identical ones.

A sixth, productive failure, grew up in a separate research tradition and belongs in the same family. It gets its own section below.

What unites them is that each one forces your brain to reconstruct something rather than recognize it. Recognition is cheap and feels like knowledge. Reconstruction is expensive and builds it. The same idea also travels under the name effortful learning: the practice that takes work is the practice that lasts.

---

Examples of Desirable Difficulties: What Counts and What Doesn't

Adding friction is not the goal, and this is where the idea gets misapplied. "Many difficulties are undesirable during instruction and forever after," the Bjorks warn in that same 2011 chapter. A difficulty earns the adjective only if it passes two tests. The extra effort has to go into the material itself: retrieving it, generating it, or telling it apart from something similar. And you have to be able to get somewhere with it. Difficulty that simply blocks you isn't desirable. It's an obstacle.

Run the usual candidates through those two tests and they sort like this:

If you met this as a quiz question (which of the following is a desirable difficulty?), the middle column is the answer key and the right-hand column is the reason. Every yes makes you reconstruct something. Every no either makes nothing harder or makes the wrong thing harder.

---

The Fluency Illusion

Open a textbook chapter you've already read twice. Run your eyes down the page. The sentences feel familiar. Each paragraph clicks into place. You close the book convinced you know the material.

A week later, you can barely retrieve the main argument.

Cognitive psychologists call this the fluency illusion or the illusion of competence, and it's the single biggest reason most study time is wasted. When information processes smoothly, your brain interprets the smoothness as evidence of mastery. It isn't. Smoothness is just smoothness.

The data is brutal. Dunlosky and colleagues' 2013 review in Psychological Science in the Public Interest sorted ten common study techniques into three tiers by how useful the evidence showed them to be. Highlighting and re-reading, two of the most popular methods on Earth, landed in the lowest tier. Practice testing and distributed practice landed in the highest. The methods learners love produce weak learning. The methods learners avoid produce strong learning.

Re-reading is the clearest case. Callender and McDaniel (2009) ran four experiments on immediately re-reading a text and found little or no benefit on later tests, with only a few exceptions. Familiarity isn't memory. Recognition isn't recall.

The illusion is structural. Your brain uses ease of processing as a heuristic for "I know this." The heuristic just happens to be wrong for predicting future retrieval. Learning to distrust your sense of mastery is the first step toward studying things that stay studied.

---

What Bjork Actually Discovered

Robert A. Bjork and Elizabeth Ligon Bjork have spent decades untangling why this happens. Their 1992 chapter "A New Theory of Disuse and an Old Theory of Stimulus Fluctuation" introduced the framework that explains all of it.

The theory splits memory into two separate dimensions.

Storage strength measures how deeply a memory is wired in. It's a function of how much you've engaged with the material across different contexts and over time. Storage strength can only go up.

Retrieval strength measures how easily you can pull that memory up right now. It fluctuates wildly. It rises when you've just studied something, falls without use, and depends heavily on the cues available in the current moment.

The fluency illusion lives in the gap between these two. When you re-read a chapter, retrieval strength shoots up because the material is right in front of you. Storage strength barely moves. The moment retrieval strength fades, there's nothing underneath to support recall.

The theory's least intuitive rule falls out of that split. The higher a memory's retrieval strength at the moment you restudy or recall it, the smaller the gain in storage strength. Easy access buys a small deposit. Hard access, provided you get there, buys a large one. Difficulty during practice, paradoxically, is what creates lasting learning.

The name came two years later. In a 1994 chapter, "Memory and Metamemory Considerations in the Training of Human Beings," Bjork collected the training methods that slowed people down in practice and helped them afterward, and called them desirable difficulties. His original list had five entries, and it doesn't quite match the one that circulates today:

Generation isn't on it. By the Bjorks' 2011 chapter it has joined the list, sharing a section with testing. Reduced feedback went the other way and faded from most summaries. The section on contested cases looks at its better-studied cousin, delayed feedback.

Soderstrom and Bjork's 2015 paper, "Learning versus Performance: An Integrative Review," tightened the distinction further. Performance is what you can do during practice, today. Learning is the relatively permanent change in your ability to do it later, in a new context. Most study habits maximize performance and undermine learning.

A quarter century after the name, the Bjorks returned to the topic in a 2020 article in the Journal of Applied Research in Memory and Cognition, "Desirable Difficulties in Theory and Practice," and their tone had shifted. The mechanisms were well established. Getting anyone to act on them was not. They admitted they had been "prone to assuming, unrealistically, that simply telling learners and teachers about relevant research findings is enough," when learners arrive with naive theories about studying that have to be dislodged first. Quoting Rohrer and Hartwig in the same issue: "Too often, the classroom is where promising interventions go to die."

The takeaway sits in the center of everything that follows. If your session feels effortless, retrieval strength is doing the work and storage strength is going nowhere. If your session feels effortful in a productive way (you're searching, failing, recovering, reaching), storage strength is climbing. The discomfort is the deposit.

This is what ties together the methods Glasp readers already know. Active recall and spaced repetition, along with the Feynman technique, the protégé effect and the blurting method, aren't unrelated tricks. They're different ways of forcing your brain into the effortful work that builds storage strength. Desirable difficulties is the meta-principle. Everything else is implementation.

---

Spacing: Why Distributed Study Beats Cramming

The spacing effect is the oldest and most replicated finding in this whole literature. Hermann Ebbinghaus described it in 1885, in the same monograph that produced the forgetting curve. It still works.

The claim is simple. If you have ten total minutes to study a piece of material, you'll remember more of it on a delayed test by splitting the ten minutes across several sessions than by spending all ten in one block. The total time's identical. The distribution is what matters.

Cepeda and colleagues' 2006 synthesis in Psychological Bulletin pulled together 839 assessments from 317 experiments across 184 articles, all of them verbal recall tasks. Spacing beat massing consistently, and the gap that produced the best retention grew as the retention interval grew. A follow-up experiment by the same group in 2008 pinned down the ratio. More than 1,350 people learned a set of facts, reviewed them once after a gap of up to three and a half months, and took a final test up to a year after that. The optimal gap turned out to be a shrinking proportion of how long you want to remember something: about 20 to 40% of a one-week delay, but only about 5 to 10% of a one-year delay. Put on a calendar, those ratios place the single review roughly one and a half to three days after first study for a test a week away, and about three to five weeks after for a test a year away.

The effect holds up in real courses, at a smaller size. A 2025 meta-analysis of mathematics learning by Murray and colleagues in Educational Psychology Review found that spacing beat massing overall (g = 0.28 across 27 studies), more strongly when the material was taught in isolation (0.43) than when it was embedded in an actual course (0.24).

The mechanism connects directly to the storage/retrieval split. When you study something twice in a row, the second exposure happens while retrieval strength is still high, so the practice is essentially free and storage gains are minimal. When you study it again three days later, retrieval strength has decayed. Pulling the memory up takes work, and that work is what buys the storage.

In practice, spacing fights two enemies: cramming and unscheduled review (which never happens). The fix is calendar-level. Pick the things you actually want to retain, give each a recurring review slot, and trust the schedule over your sense of what needs more attention. That sense is, again, mostly fluency talking.

---

Interleaving: Why Mixing Beats Blocking

Interleaving means mixing different topics or problem types within a single study session, rather than studying one to mastery before starting the next. If you're learning algebra, you don't do twenty quadratic problems in a row. You do a quadratic, then a system of equations, then a function transformation, then back to a quadratic.

It feels worse. Performance during the session drops. Learners regularly rate interleaved practice as less effective even after they've measurably learned more from it. This metacognitive mismatch is one of the strongest illusions in the field.

Doug Rohrer and Kelli Taylor demonstrated it in 2007 with a study whose title says the finding: "The Shuffling of Mathematics Problems Improves Learning." College students learned to solve several kinds of problem, then practiced either blocked by type or randomly mixed. Tested a week later, the mixed group's performance was vastly superior. Rohrer, Dedrick and Stershic replicated it in a real classroom in 2015, with 126 seventh graders working through interleaved or blocked assignments over three months, and got the same result on a delayed test.

Why does it work? Two mechanisms. First, interleaving forces discrimination: you can't just apply the same procedure on autopilot, you have to figure out which procedure each problem calls for. That discrimination is the skill you actually need on a test or in real work. Second, interleaving spaces each topic by definition: the gap between problem-type-A items is filled with problem-type-B items, so each return to A involves real retrieval.

The discrimination mechanism also tells you where interleaving stops working, and this is the one place the popular write-ups get sloppy. Brunmair and Richter's 2019 meta-analysis in Psychological Bulletin, "Similarity matters," pooled 59 studies and 238 effect sizes. Interleaving won overall, moderately (Hedges' g = 0.42), but the benefit split hard by material: biggest for visual categories like paintings, smaller for math problems, and for studies built on words it reversed, with blocking winning outright. The pattern is what the theory predicts. If the categories you're mixing are confusable, juxtaposing them teaches you to tell them apart. If they have nothing in common, there's nothing to discriminate and you pay the switching cost for free.

For self-directed learners, interleaving is easier than it sounds and harder than it looks. You don't need a fancy scheduler. You just need to refuse the urge to "finish" a topic before moving on. Pick things you could actually mix up, like two authors who disagree on the same question or three problem types from one course, and alternate between them. Your brain will protest. Your brain is wrong. Our full guide to interleaving practice covers the scheduling mechanics.

---

Retrieval Practice: The Testing Effect

If you only adopt one desirable difficulty, make it this one. Retrieval practice (the testing effect) is the act of pulling information out of memory rather than pushing it back in. It's the engine behind active recall, flashcards, the blurting method, and most of what works in education.

Henry Roediger III and Jeffrey Karpicke's 2006 paper in Psychological Science, "Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention," is the canonical demonstration. Students studied prose passages, then either restudied them or took free-recall tests, with no feedback either way. In the first experiment, the final test came at five minutes, two days, or one week. At five minutes the restudiers won, 81% to 75%. A week out the order had flipped: the tested group recalled 56%, the restudy group 42%. A second experiment pushed the comparison further. Students who studied a passage once and then tested themselves three times recalled 61% a week later, against 40% for students who studied it four times, roughly 50% more. Same total time, different work.

Why retrieval does this is still argued over. That it changes the memory is not. Bjork put it in two words back in 1975, and his 1994 chapter repeats them: the act of retrieval is a "memory modifier." One prominent explanation, the episodic context account from Jeffrey Karpicke's lab, holds that a successful recall updates a memory's context to include the moment you retrieved it as well as the moment you learned it, and that the richer context lets a later search home in on the right memory. Even failure helps. Search for an answer, come up empty, then study it, and you'll remember it better than if you'd spent the same time reading. Richland, Kornell and Kao demonstrated that in 2009, in a paper titled "The Pretesting Effect: Do Unsuccessful Retrieval Attempts Enhance Learning?"

One gap: most of this evidence comes from verbal material. The Murray meta-analysis of mathematics learning from the spacing section found only seven studies that compared testing with restudy, and their pooled effect (g = 0.18) had a confidence interval that crossed zero. The authors read that as too little evidence to call, not as proof that retrieval fails in math.

Retrieval works in dozens of forms: closed-book recall, flashcards, free-form summary writing, teaching aloud, the Feynman technique. What they share is the underlying move: pull it out of memory before you look.

After you've highlighted a chapter with Glasp's web highlighter, close the source and try to reconstruct the argument from your highlights alone. Then open the page and check. The gap between what you produced and what's there is your storage-strength deficit. Closing that gap is the work.

---

Generation: Produce the Answer First

The generation effect is a close cousin of retrieval practice, distinct enough to deserve its own line. Slamecka and Graf (1978) showed across five experiments that learners who generated material themselves remembered it better than learners who simply read the same material.

Their paradigm was rule-based and almost embarrassingly simple. One group read word pairs outright, like "lamp - light." The other was handed a rule (associate), the stimulus word "lamp," and the single letter "l," and had to produce "light" themselves. The generators won by a wide margin, even though the reading group had seen the answer.

The principle scales up. Reading a worked proof builds less competence than attempting the proof and then comparing. Reading someone's book summary teaches less than writing your own. Watching someone code teaches less than typing the code yourself, hitting an error, and figuring out why.

Generation is desirable difficulty in pure form. It deliberately withholds information you'd be happy to receive, forcing you to produce it. The frustration of half-remembering, half-guessing is the mechanism. By the time you check your answer, you've already done the work that converts exposure into encoding.

Highlighting can be a generation tool or a passive one. Yellow-bombing every paragraph is passive: you're outsourcing judgment to a future re-read. Sparse, deliberate highlighting is generative: you're forcing present-you to commit to "this matters more than that," which is itself synthesis. The science of highlighting goes deeper.

---

Varied Practice: Change the Conditions

Varied practice means practicing the same skill in slightly different forms, contexts, and conditions, rather than under identical conditions every time.

Kerr and Booth's 1978 study with children throwing beanbags became the textbook example. They ran two age groups, 8-year-olds and 12-year-olds. One group practiced from a single distance. The other practiced from a mix of distances that bracketed the target but never included it. The varied group outperformed the fixed group on the test throw they'd never practiced. They'd built a more general motor representation.

Note the design detail, because it's usually dropped: the varied distances surrounded the target. Variability isn't a license for arbitrary noise. It works when the variations span the space the skill has to generalize across.

Cognitive learning shows the same pattern. Vary the wording of definitions you study. Vary the contexts in which you encounter a new term. Read about a concept in two different fields rather than two different chapters of the same textbook.

Variability supports transfer: the ability to use what you've learned outside the conditions in which you learned it. Transfer is what most learners want and what most study habits actively prevent. If you practice a skill in one form, you encode the form along with the skill, and you'll struggle when the form changes. Variability decouples the skill from any single form.

---

Productive Failure: Struggle Before the Explanation

Manu Kapur at ETH Zurich calls it productive failure: give learners a hard problem before teaching them the method, let them generate their own flawed attempts, and only then deliver the instruction.

The attempts usually fail. That's the design. What the failed attempts do is build a map of the problem space, so that when the canonical solution arrives, the learner has somewhere to put it and can see what their own approach was missing.

Sinha and Kapur's 2021 meta-analysis in Review of Educational Research pooled 53 studies and 166 experimental comparisons. Problem-solving before instruction beat instruction before problem-solving (Hedges' g = 0.36), and the designs that followed Kapur's productive-failure principles most closely did better, with effects between 0.37 and 0.58. Two boundaries showed up. For second to fifth graders, and for domain-general skills, instruction first came out ahead. Younger children may not yet have the problem-solving footing to fail productively.

The reading version is easy to run. Before you open an article on a topic you half-know, spend two minutes writing what you think the answer is. Then read. Your version is almost always wrong in an interesting way, and the gap between it and the author's is what you'll remember.

The six at a glance:

---

Two Contested Cases: Delayed Feedback and Hard-to-Read Fonts

Two candidates get called desirable difficulties all the time. One is a close cousin of an entry on Bjork's original list and turns out to be conditional. The other was never on the list and didn't survive replication.

Reduced and delayed feedback. The entry on Bjork's 1994 list is reduced feedback, and it came from motor-skill training. Richard Schmidt and his collaborators had found that giving feedback in periodic summaries, or fading it out as practice went on, slowed the acquisition of simple motor skills and improved their long-term retention. The usual explanation is that constant feedback becomes a crutch: performance leans on it, then sags when it's gone.

The version that has been tested on study material is delayed feedback. In Butler, Karpicke and Roediger (2007), students read prose passages, took a multiple-choice test, and saw the correct answers either right away or after a delay. On a later cued-recall test, delayed feedback won. The authors' explanation is almost anticlimactic: feedback that arrives later works as a spaced second presentation of the material.

Kulik and Kulik's 1988 meta-analysis of 53 studies had already mapped the boundary. Lab experiments on learning test content usually favored delay. Applied studies using real classroom quizzes and real course material usually found immediate feedback more effective. So the answer is "sometimes." Holding back the answer for a while can help when you're certain to come back for it. In ordinary coursework, get it fast.

Hard-to-read fonts. In 2011, Diemand-Yauman, Oppenheimer and Vaughan published "Fortune Favors the Bold (and the Italicized)" in Cognition. In a lab study, 28 participants studied fact sheets about fictional alien species, set either in clean 16-point Arial or in smaller, gray Comic Sans or Bodoni. The hard-font group recalled 86.5% of the facts, against 72.8%. A second study reformatted real course materials for 222 Ohio high school students and found higher scores on their class tests too. It looked like free learning: change the typeface, remember more.

It didn't hold up. Rummer, Schweppe and Schwede ran three slightly modified replications of the lab study, published in 2016, and found no effect in any of them. A 2018 meta-analysis by Xie, Zhou and Liu pooled 25 articles and 3,135 participants and found that making text harder to read had no effect on recall (d = -0.01) or transfer (d = 0.03). What it did do was make people study longer (d = 0.52) and feel less sure of what they'd learned (d = -0.43). The font added the feeling of difficulty without the benefit.

Sans Forgetica, the back-slanted, gapped typeface RMIT University released in 2018 to exploit the idea, fared the same. Two 2020 papers in Memory tested it across seven experiments. Readers rated it harder to read than Arial, and they didn't remember more. One of the two is titled "Disfluent Difficulties Are Not Desirable Difficulties."

Robert Bjork weighed in himself. In a 2016 commentary titled "Is Disfluency Desirable?", he and Carole Yue concluded that the original benefits "do not replicate, at least not in any general way," and suggested why: what matters for learning from text "happens after the perceptual encoding of what is on the paper or screen."

That's the first of the two tests from the examples section: the effort has to go into the material. A hard font makes seeing harder. Desirable difficulties make retrieving, generating and discriminating harder, and only that second kind of effort changes the memory you're trying to build. Later studies turned up narrow exceptions (one found a benefit only when readers weren't expecting a test), but nothing you could build a study habit on.

---

When Difficulties Become Undesirable

Not all difficulty helps, and the research on when it stops helping is more specific than most summaries admit.

A difficulty becomes undesirable when the learner can't engage with it productively. If you can't decode the words at all, slowing your reading further won't help. If you don't know what a derivative is, interleaving derivative and integration problems is noise. If your retrieval practice is so far past your storage strength that you produce nothing, you're not retrieving, you're flailing.

Cognitive load theory gives this a mechanism. Working memory can juggle only a handful of interacting elements at once. When material already has high element interactivity (lots of pieces that only make sense in relation to each other), adding generation or interleaving on top pushes total load past capacity, and the extra difficulty buys nothing because there's no spare capacity to process it with.

Chen, Kalyuga and Sweller demonstrated the crossover directly in a 2015 study in the Journal of Educational Psychology, using geometry instruction. On low-element-interactivity material, the generation effect showed up as expected: generating answers beat being shown them. On high-element-interactivity material it flipped, and studying worked examples beat generating. Then in a second experiment with more knowledgeable learners, the effective complexity dropped and generation won on everything. That's the expertise reversal effect. The correct study strategy is not fixed. It moves as you get better.

This is the second half of the Bjorks' definition from the top of this guide: without the background knowledge to respond to a difficulty, it turns undesirable.

The right zone is where you can produce something, even if it's incomplete. You're reaching, not falling. You should be failing on some attempts and succeeding on most. Getting everything right means the material is too easy or the gap too short. Producing almost nothing means you've crossed the boundary and need scaffolding, not grit.

So adjust in both directions. When something's genuinely beyond you, don't romanticize the struggle: read the explanation, get the foundation in place, then come back. When something feels too easy, don't let the comfort fool you: increase the spacing, mix in harder variants, or move to a more generative format.

Picture storage strength on one axis and retrieval strength on the other. High storage / high retrieval is recently practiced and well-learned. High storage / low retrieval is where retrieval practice does the most good, because the search is hard but the deposit is real. Low storage / high retrieval is the dangerous one: stuff you just re-read and feel you know but haven't built. Cramming lives here. Low storage / low retrieval is genuinely new material, where you need scaffolding first.

---

Do AI Tools Remove Desirable Difficulties?

By default, yes. An AI assistant is an extraordinarily efficient machine for deleting exactly the effort that builds storage strength, and the first good evidence on what that costs arrived in 2025.

Hamsa Bastani, Osbert Bastani and colleagues ran a randomized controlled trial, published in PNAS, with nearly a thousand high school math students at a large school in Turkey. Students got one of two GPT-4 tutors during practice sessions, or no AI at all. "GPT Base" behaved like the ordinary ChatGPT interface. The researchers prompted "GPT Tutor" with safeguards, most importantly to offer teacher-designed hints rather than hand over answers.

During assisted practice, both worked spectacularly. GPT Base students improved 48% over controls. GPT Tutor students improved 127%.

Then the researchers took the tools away and ran an exam.

Students who'd practiced with GPT Base scored 17% worse than students who never had AI at all. Not worse than they'd looked with the tool in hand: worse than the control group that never got one. The GPT Tutor group came out statistically indistinguishable from controls, with a point estimate near zero.

That's Soderstrom and Bjork's performance/learning dissociation, reproduced with a chatbot, at school scale. A tool that lifted practice scores by half left students worse off once it was gone. And notice the fix wasn't banning it. The fix was making the AI withhold the answer, which is the generation effect wearing a different hat.

Two smaller studies repeat the pattern. In a 2025 randomized trial by André Barcaui (n = 120), undergraduates studied with ChatGPT or without it, then took a surprise test 45 days later. The ChatGPT group scored 57.5% against 68.5%, and the paper says outright that the result fits "the 'desirable difficulties' principle." In a September 2026 preprint, Bergh and colleagues had 55 computer science students work through C programming tasks with ChatGPT or with ordinary web search. The ChatGPT group wrote better code (89% vs. 69%) and scored lower on recall, both right afterward (41% vs. 53%) and 48 hours later (39% vs. 52%).

Adult knowledge work shows the same shape. At CHI 2025, Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of AI-assisted tasks. Workers reported reductions in cognitive effort, and the correlation that matters ran this way: higher confidence in the AI predicted less critical thinking, while higher confidence in themselves predicted more. Trust in the tool substitutes for thought.

A widely shared MIT Media Lab study by Kosmyna and colleagues points the same direction, with EEG recordings showing the weakest neural connectivity in participants who wrote essays using an LLM, and those participants struggling to quote work they'd just produced. Treat it as suggestive rather than settled. As of October 2026 it's still an arXiv preprint, revised in December 2025, with no peer-reviewed version we could find. Only 54 people took part (18 of them in the final session), and a comment other researchers posted to arXiv questions its sample size, its EEG analysis and inconsistencies in how the results are reported.

None of this is an argument against using AI. It's an argument about where in the loop you put it.

Two papers from 2026 make that point with data. In a working paper from Middlebury College, 210 undergraduates were randomly assigned to study with or without GPT-4o for up to 35 minutes, then everyone was tested without it. The AI group scored about a quarter of a standard deviation higher, and kept the lead on a retention test a week later. One thing that differed from the Turkish classrooms was how students used it. The most common use was asking the model to explain a difficult concept: 44% of the AI group reported that, against 16% who said they had it draft responses (the chat logs showed more drafting than students reported). Asking for an explanation when you're stuck is a different act from asking for the answer.

A tutor that withholds answers only works if students use it. In a two-year randomized trial of Khan Academy's Khanmigo across 18 Tennessee middle schools, released as an NBER working paper in August 2026, the gains were small (0.06 to 0.08 standard deviations a year) and resembled what Khan Academy practice delivers without the AI. The median student messaged the tutor on only a third of the days they practiced, and in only 17% of the sessions where they'd made a mistake. A hint-only tutor helps only the students who ask it something.

The rule of thumb: let the AI grade you, not answer for you. Do the retrieval yourself, produce something, and only then bring in the model to compare your output against the source. You keep the difficulty and outsource the marking. Our piece on AI and learning works through the broader tradeoff, AI study modes compared looks at which assistants now withhold the answer by default, and do AI tutors actually work? sets the Turkish trial beside the trials that found gains.

---

Designing Your Routine for Desirable Difficulties

Knowing the principles doesn't make them automatic. You have to engineer them in, because the path of least resistance will always be the comfortable, low-storage option. Here's how to bake the difficulties into a self-directed learning system.

Highlight to generate, not to mark. Mark sentences sparingly, then write a brief note in your own words explaining why this passage matters and how it connects to what you already know. The sparse selection is discrimination. The note is generation. Highlighting without notes is the passive trap.

Use AI chat for retrieval, not for summaries. The wrong way to use an AI assistant is to ask it to summarize a chapter you haven't read. The right way is to read the chapter, close it, write your own summary, then paste that into Glasp's AI chat and ask it to grade your reconstruction against the source. You did the retrieval. The AI does the comparison.

Space with Kindle highlights, not memory. Your sense of when to review is broken: it's run by recency and emotion, not by storage curves. Schedule reviews on a calendar, pull up highlights from a book you read three weeks ago, and try to reconstruct the argument before scrolling. The mechanical schedule is what protects spacing from your fluency-driven instincts.

Rotate topics in the community feed. Don't read four articles in a row from the same person on the same topic. Cycle: cognitive science, then entrepreneurship, then writing, then back to cognitive science the next day. Topics this far apart can't be confused, so don't expect interleaving's discrimination benefit. What you get is spacing: the first topic has had a day to fade before you return to it. For real interleaving, alternate between two authors who disagree on the same question.

Watch a YouTube Summary, then test yourself. Close the page and write three things you'd want to remember in a year. Compare to the transcript. The gap is where your work is.

Predict before you read. Run the productive-failure move on anything you open cold. It changes what you notice.

Vary the format. Read the same idea in a book, a paper, a thread, and a video. Each medium encodes the idea differently, and your brain has to abstract across them. That's the point of varied practice.

The system view lives in our companion piece on building a learning OS. Desirable difficulties is the principle. The OS is the daily mechanics.

---

Frequently Asked Questions

What are the five desirable difficulties?

Spacing (distributing study over time), interleaving (mixing topics or problem types), retrieval practice (recalling instead of re-reading), generation (producing an answer before seeing it), and varied practice (changing the conditions of practice). Productive failure, attempting a problem before you're taught the method, is a closely related sixth. All six share one mechanism: they force reconstruction rather than recognition. Bjork's original 1994 list was slightly different: it had reducing feedback where today's lists put generation.

What is an example of a desirable difficulty?

Closing a book and writing a summary from memory is the cleanest example. So is reviewing on a spaced schedule instead of cramming, mixing problem types in one session, or predicting an answer before you read the explanation. The test is whether the difficulty comes from the material and how you engage with it, and whether you can still produce something.

Which of the following is a desirable difficulty?

The right answers are spacing your study sessions, interleaving related topics, testing yourself, generating an answer before you see it, and varying the conditions of practice. The usual distractors are cramming, re-reading, and highlighting everything. Hard-to-read fonts are the trap: older course materials list them as a desirable difficulty, but the effect didn't replicate. The rule behind the answer key: a desirable difficulty slows you down during practice and improves what you can recall and use later.

What is an undesirable difficulty?

Any difficulty that makes studying harder without making you retrieve, generate or discriminate, or one you can't get anywhere with. Hard-to-read fonts, noisy rooms and badly written textbooks are real difficulties, but the effort goes into seeing or concentrating, not into the idea. A problem far beyond your level fails for the second reason. The Bjorks' own line is the one to remember: "Many difficulties are undesirable during instruction and forever after."

Do hard-to-read fonts like Sans Forgetica improve memory?

No. A 2011 study reported that harder fonts improved recall, but the finding didn't replicate. A 2018 meta-analysis of 25 articles and 3,135 participants found no effect of hard-to-read text on recall or transfer, and two 2020 papers found no memory benefit from Sans Forgetica in particular. People find these fonts harder to read, study longer, and learn the same amount.

Who coined the term desirable difficulties?

Robert A. Bjork coined it in 1994. The source usually cited is his chapter "Memory and metamemory considerations in the training of human beings," in the MIT Press book Metacognition: Knowing about Knowing. The accessible summary most people read is a 2011 chapter by Elizabeth and Robert Bjork, "Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning," in Psychology and the Real World. Its core argument: conditions that slow acquisition and make performance look worse during practice often produce better long-term retention and transfer, and learners systematically misread that short-term dip as failure. The Bjorks revisited the topic in 2020 in "Desirable difficulties in theory and practice," in the Journal of Applied Research in Memory and Cognition, mostly to note how hard the ideas are to get adopted.

Is highlighting a desirable difficulty?

It depends entirely on how you do it, and most highlighting isn't. Marking whole paragraphs just defers the decision to whoever re-reads the page later. Highlighting sparsely, one or two passages per page, and writing a note in your own words for each one is both a discrimination task and a generation task. Treat your highlights as future retrieval cues rather than a substitute for reading.

Should I struggle on every problem?

No. The goal is productive struggle, not flailing. If you're producing nothing, you've gone past the desirable zone into overload. Back up, get the foundation in place, then return to the difficulty. You want to be failing sometimes and succeeding most of the time. And for genuinely complex material you're new to, the research says studying worked examples first beats generating: Chen, Kalyuga and Sweller (2015) found that on high-element-interactivity content, worked examples beat problem-solving until learners had enough expertise for the effect to flip back.

Does using ChatGPT hurt learning?

It depends entirely on where you put it in the loop. In the 2025 PNAS trial, students who practiced with an unrestricted GPT-4 tutor later scored 17% below students who never had AI, while students whose tutor gave hints instead of answers showed no such penalty. An experiment reported in 2026, in which the most common use was asking the AI to explain concepts, found the reverse: higher scores on a no-AI test a week later. Ask a model to do the retrieval for you and you lose the learning. Do the retrieval yourself and use the model to explain or check, and you keep it.

How long should you wait between study sessions?

Longer than feels comfortable, and longer each time. The gap that maximizes retention grows with how long you need to remember, while shrinking as a share of it. Going by the ratios from Cepeda's 2008 experiment (20 to 40% of a one-week delay, 5 to 10% of a one-year delay), the best gap is roughly one and a half to three days for a test a week away, and about three to five weeks for a test a year away. When a review feels easy, lengthen the next gap. When it feels hard, shorten it. A schedule beats your gut either way.

Do desirable difficulties apply to skills as well as facts?

Yes, and several of them started there. Varied practice and reduced feedback both came out of motor-skill research, and interleaving was studied in motor skills, under the name contextual interference, long before anyone shuffled math problems. Any skill with discrimination, transfer, and retrieval components benefits: coding, writing, design, language, instruments, athletic technique. The form of practice changes; the principle doesn't.

Why don't schools teach with desirable difficulties?

Because the easy way looks better in the short run. Massed practice and re-reading produce higher scores on quizzes given right after instruction, and worse scores a week later. Most schools don't measure delayed retention. Soderstrom and Bjork (2015) made exactly this point: confusing performance with learning is structural, not personal. As a self-directed learner, you don't have to wait for institutions to catch up.

---

Conclusion

The principle to walk away with: if your studying feels easy, you're probably not studying. Effort is the price of durable learning, and the brain pays out only when you've earned it through spacing, interleaving, retrieval, generation, variability, or productive failure.

This doesn't mean grinding harder, it means grinding smarter. Most learners are already putting in the time. They're just spending it on activities that maximize fluency and minimize storage. Swap one re-read for a closed-book reconstruction. Swap one cram session for four spaced reviews. Swap one block of identical problems for a mix. Each swap trades short-term comfort for long-term retention. And skip the fake kind of hard: an uglier font or a noisier room buys you nothing.

The stakes went up in 2026. Every tool on your desk now offers to do the effortful part for you, and the evidence says taking that offer costs you the learning even while it flatters your performance. The good news from the same research: a tool that withholds the answer produces no such penalty. You get to choose which one you're using.

Glasp was built around this trade: sparse highlighting, marginal notes, AI-graded reconstructions, spaced reviews of past reads, and a community feed that mixes topics by default. Each is a small piece of friction designed to convert exposure into storage.

Tomorrow, pick the easiest thing in your routine and replace it with the harder version. That swap is the whole game.