What the Illusion of Competence Actually Is
The illusion of competence is the mistaken belief that you've learned something because it feels familiar. It happens when your brain reads the ease of processing material as evidence that you've stored it, when in fact the two are separate things. Psychologists also call it the fluency illusion, and it's a failure of metacognition, which is your ability to judge your own knowledge.
Here's the distinction that matters. Recognizing a fact and producing a fact are different operations. When you reread a chapter you're doing recognition, because the words are on the page and your job is to follow them. When the exam arrives, or when someone asks you a question, or when you sit down to write, you're doing production. The page is blank and your job is to generate the content from nothing.
Recognition gets easier every time you reread. Production doesn't, at least not in the same way. So the signal you're using to decide whether to stop studying is generated by the task you're doing, not the task you're preparing for. That's the whole problem, and the rest of this article is about how it works and what to do instead.
The 2008 Study That Proved Students Can't Predict Their Own Recall
In 2008, Jeffrey Karpicke and Henry Roediger published a study in Science that made the gap impossible to argue with. College students learned 40 Swahili-English word pairs, like mashua for boat. Everyone started the same way, studying the whole list and then taking a test on the whole list.
The four groups differed only in what happened after a student got a pair right for the first time.
| Condition | After a pair was recalled once | Recall one week later |
|---|---|---|
| Study all, test all | Kept in both study and test | 80% |
| Drop from study, keep testing | Removed from study, still tested | 80% |
| Keep studying, drop from testing | Still studied, removed from tests | 36% |
| Drop from both | Removed from everything | 33% |
Look at the third row. Those students kept studying every single word pair in every study period, all the way to the end. They saw the material more than the group in row two did. A week later they recalled 36% against that group's 80%. Repeated studying after an item was already learned produced no measurable benefit at all. Repeated retrieval more than doubled retention, with an effect size of d = 4.03, which is enormous by any standard in psychology. The score distributions didn't even overlap. The drop-from-testing groups ranged from 10% to 60%, the repeated-testing groups from 63% to 95%.
Now the part that makes this an article about self-assessment rather than an article about flashcards. At the end of the learning phase, before they went home, the researchers asked all four groups to predict how many of the 40 pairs they'd recall a week later.
Every group said roughly the same thing. The mean predictions were 20.8, 20.4, 22.0, and 20.3 words out of 40. Everyone expected about 50%. An analysis of variance found no significant difference between the groups at all, with F less than 1.
So the students whose real outcome was 33% and the students whose real outcome was 80% made identical forecasts. Their learning had diverged by nearly two and a half times, and their sense of it hadn't moved a millimeter. That is the illusion of competence, measured.
Why Fluency Feels Exactly Like Knowledge
The mechanism was pinned down by Asher Koriat at the University of Haifa and Robert Bjork at UCLA, in a 2005 paper in the Journal of Experimental Psychology: Learning, Memory, and Cognition titled, straightforwardly, "Illusions of Competence in Monitoring One's Knowledge During Study."
Their argument is about a mismatch in conditions. When you judge your own learning, you do it while the material is in front of you. When you're tested, the material is gone and you have to supply it. As they put it, judgments of learning "are made in the presence of information that is absent but solicited during testing," and the failure to discount that present information "can instill a sense of competence during learning that proves unwarranted during testing." They named the effect foresight bias in that same paper.
In their first experiment, 24 undergraduates studied 60 word pairs of varying association strength and rated how likely they were to recall each one. Across the whole list, predictions were almost perfectly calibrated, landing within three points of actual recall. That aggregate number hides the failure. Broken down by item type, the unrelated pairs drew significantly inflated predictions, while the strongly associated pairs ran the other way and students underestimated what they'd recall. Seeing citizen next to tax makes the connection look obvious. Seeing citizen alone a week later does not. Their 2006 follow-up in Memory & Cognition went on to show the bias can be reduced, including by delaying the judgment until after study.
This is the same cognitive machinery behind the curse of knowledge. Koriat and Bjork point to Elizabeth Newton's 1990 experiment in which people tapped out the rhythm of a well-known song and predicted whether listeners would name it. The tappers, hearing the melody in their heads, predicted a success rate around 50%. Listeners actually got it less than 3% of the time. Once you know something, you can't easily simulate not knowing it. When the thing you're failing to simulate is your own future self, you overestimate what you'll remember.
One more thing is worth naming here. This bias doesn't feel like a bias. It feels like knowing. You're not being lazy or overconfident in some character-flaw sense, and your monitoring system isn't broken. It's reporting accurately on the wrong variable.
The Study Habits That Manufacture the Illusion of Competence
Not every study habit produces the illusion at the same rate. The ones that do share a property. They keep the answer visible while you work.
| Study habit | What it feels like | What it actually builds | Illusion risk |
|---|---|---|---|
| Rereading a chapter | Smoother and clearer each pass | Recognition of the text's surface | Very high |
| Highlighting as you read | Active, engaged, productive | A record of what looked important | High |
| Copying notes neatly | Thorough, organized, complete | A transcript, and hand fatigue | High |
| Watching a lecture or video | Effortless understanding | Familiarity with the explanation | Very high |
| Explaining aloud from memory | Awkward, halting, exposing | Retrieval paths you can reuse | Low |
| Answering questions cold | Uncomfortable, sometimes humbling | Durable, testable knowledge | Very low |
The first four are the illusion factories. The last two are what they should be replaced with, and notice that the comfortable activities are the deceptive ones. Robert Bjork's framework of desirable difficulties makes the same point from the other direction. Conditions that slow you down during practice tend to improve long-term retention, and conditions that speed you up tend to hurt it. Difficulty during study isn't a sign you're doing it wrong. Very often it's the only evidence you're doing it right.
Video deserves a special mention because it's the fastest illusion generator most of us use daily. A good explainer is engineered to be clear. You follow every step, nothing catches, and you finish feeling like you've understood. What you've actually experienced is someone else's fluency. Pause a lecture at the 20-minute mark, try to reconstruct the argument on paper, and the gap shows up immediately.
What Students Actually Do: 84% Reread, 11% Self-Test
If retrieval practice is this much better, you'd expect people to use it. They don't.
Karpicke, Butler, and Roediger surveyed 177 undergraduates at Washington University in St. Louis and published the results in Memory in 2009. The researchers asked one open question. What do you actually do when you study?
83.6% of them named rereading notes or the textbook, and 54.8% called it their single most-used strategy. Only 10.7% named practicing recall, meaning genuine self-testing, which works out to 19 students out of 177. Two people named it as their number one strategy.
A second question removed the ambiguity. Students imagined they'd just read a textbook chapter once and had to choose between going back and restudying it, trying to recall it without restudying, or doing something else. Of the 101 students who got that version, 57% chose to restudy and 21% chose something else, so 78% would not test themselves. Only 18% would. The honest caveat is that this version forced a trade, because choosing to self-test meant giving up the chance to reread. When a second group of 76 students was allowed to do both, 42% chose to test first.
These aren't struggling students, either. They come from a university whose students average above 1400 on the SAT. Strong students, given a direct choice, still lean heavily toward the strategy that produces the feeling of learning over the one that produces the learning. That's what a real illusion looks like. It isn't a knowledge gap you can close by telling people the finding once; it's a perceptual pull that keeps working after you know about it.
The 2026 Version: AI Summaries Compress the Illusion
Every technology that makes information easier to absorb also makes it easier to feel finished. AI is the strongest version of this we've had.
In October 2025, Shiri Melumad and Jin Ho Yun published a study in PNAS Nexus running seven experiments with 10,462 participants. The design was clean. People learned about a topic either from an AI-generated synthesis or from ordinary web search results, then wrote advice for someone else on that topic.
People who learned from the AI summary reported developing shallower knowledge, and the advice they produced showed it. It was shorter, with a mean of 84.58 words against 94.64 in the first experiment. It referenced fewer specific facts, with 0.464 unique entities per piece of advice against 0.718. It was also markedly less original. Advice from AI-learners clustered together at a cosine similarity of 0.159, nearly three times the 0.057 among web-search learners, meaning they were all writing close to the same thing. In a later experiment, recipients rated the AI-derived advice as less likely to be adopted.
The authors attribute this to the format rather than to accuracy. A summary hands you a finished, coherent object; search results hand you fragments you have to select among, reconcile, and assemble. The assembly is the learning, and the summary removes it. The effect held even in a supplementary version where the AI summaries carried live web links, though only 26% of those participants clicked one.
None of this means you should stop using AI to read. It means the subjective signal you get from an AI summary is even less trustworthy than the one you get from rereading, because it arrives faster and feels more complete. The practical rule is to treat the summary as an index of what to go read, not as the thing you read. We go deeper on that in reading with AI without losing comprehension.
Five Tests That Break the Illusion
You can't fix a monitoring error by trying harder to monitor. You need a test that runs under production conditions rather than recognition conditions. These five take between one and ten minutes.
1. The blank page test. Close everything. Write down everything you can recall about what you just studied, then open the source and compare. The gap between what you wrote and what's there is your real state of knowledge, and it's usually a shock the first few times. This is the core of the blurting method, which is popular with medical and law students for exactly this reason.
2. The delayed judgment. Don't rate your learning the moment you finish a page. Nelson and Dunlosky showed in 1991, in Psychological Science, that predictions made after a short delay are far better at separating what you'll actually recall from what you won't. Ratings made the moment you finish, while the answer is still fresh in mind, are only moderately good at it. Put the material down, come back, and judge from the cue alone.
3. The cue-only check. Cover the answer. Look only at the prompt, the heading, or the highlighted term, and see whether you can produce what follows. If you have to uncover it to feel confident, you didn't know it.
4. The explanation test. Explain the idea out loud to someone who doesn't know it, without notes. The moment your sentence stalls is the moment you've located a genuine hole. This is the engine behind the Feynman technique. Teaching forces production, and production is the only thing that leaves a retrieval path behind.
5. The transfer question. Don't ask whether you remember something. Ask where it would apply that the author didn't mention. Recognition can't fake an answer to that one.
Notice what all five have in common. Each removes the answer before asking for a judgment. That's the only structural property that matters.
How to Highlight Without Faking Competence
Highlighting gets a bad reputation in this literature, and some of it is earned. In the widely cited 2013 review by John Dunlosky and colleagues in Psychological Science in the Public Interest, highlighting and underlining were rated low utility, alongside rereading. But the finding is narrower than the headline suggests. What fails is highlighting as a terminal act. You mark the page, you feel the material register, you move on, and the mark becomes a substitute for the encoding rather than a trigger for it. We unpack the full evidence in the science of highlighting.
Highlighting survives the critique when it's the first step of a loop instead of the last step of a read. Three changes do most of the work.
Highlight less, and decide why. A page where everything is yellow contains no decisions. The value isn't in the color; it's in the judgment about what earns it. A useful constraint is three to five highlights per article, which forces you to rank rather than collect.
Attach a note in your own words. If you change one thing, change this. A highlight is the author's sentence. A note is yours, and writing it is a small act of production, the only kind of work that builds retrieval paths. When you use Glasp's web highlighter, every highlight takes an inline note, and that note is what you'll test yourself against later. Rephrasing beats copying every time, because copying can be done fluently and rephrasing can't.
Come back with the source closed. Almost everyone skips this step. Open your highlights a few days later and, before reading them, try to recall what the piece argued. Then read. Your Kindle highlights work the same way once they're imported and searchable. Spacing the return visit out is what converts a pile of highlights into memory, and we've mapped out the intervals in spaced repetition for readers.
Video needs the same treatment. If you learn from lectures and talks, YouTube Summary gives you the transcript and timestamps, which makes it possible to do something other than watch passively. Pause at each section boundary, write what was just argued, then check the transcript. The summary is your answer key, not your study material.
A Weekly Loop That Keeps Your Self-Assessment Honest
Tools don't remove the illusion. A schedule does, because it forces retrieval to happen on a clock rather than when you feel like it. This loop takes about 35 minutes a week.
While reading, no extra time. Highlight sparingly. Write a short note in your own words on anything you highlight. Don't stop to review.
Same day, 5 minutes. Close the article. Write three sentences on what it claimed and one sentence on what you disagree with or don't yet understand. That last line matters, because it's impossible to write fluently without actually having processed the argument.
Two days later, 10 minutes. Read only your own notes, not the highlights, and try to reconstruct the source's reasoning from them. Anything you can't reconstruct is a note that was too thin, which tells you something useful about how you were reading.
Weekly, 20 minutes. Take the week's material and write a single connected paragraph linking at least two sources. This is the transfer test at scale, and it's where most of the durable learning actually happens. If you'd rather do it conversationally, Glasp's AI chat will question you across your own saved highlights, which is closer to a viva than a review.
Occasionally, no fixed time. Read what other people highlighted on something you've already read. Glasp's community makes this easy, and it works as a calibration check. The passages other careful readers marked and you skipped are a direct measurement of what your attention missed. It's uncomfortable in a productive way. For a fuller treatment of turning reading into retained knowledge, see how to remember what you read.
The loop isn't complicated. What makes it work is that every step asks you to produce something with the source closed, which is the one condition under which your confidence means anything.
Frequently Asked Questions
What is the illusion of competence in simple terms?
It's believing you know something because it feels familiar. Rereading, highlighting, and watching a lecture all make material easier to process, and your brain reads that ease as proof of learning. The feeling comes from recognition, but tests and real work require production, which is a different and much harder operation.
Is highlighting bad for studying?
Highlighting on its own is weak, and Dunlosky's 2013 review rated it low utility. It becomes useful when it starts a process rather than ending one. Mark sparingly, write a note in your own words, and return to the highlights later with the source closed. The mark itself isn't the learning. What you do with it is.
How is the illusion of competence different from the Dunning-Kruger effect?
Dunning-Kruger is about general skill, where low performers in a domain overestimate their overall ability. The illusion of competence is narrower and happens to everyone, including experts. It's a moment-to-moment misreading of a specific study session, driven by how fluently material is processing right now. You can be a genuine expert and still badly misjudge whether you'll recall today's reading next week.
Can you get rid of the illusion of competence once you know about it?
No, and that's the frustrating part. Koriat and Bjork's work shows it's driven by the conditions under which you make the judgment, not by what you believe about learning. Knowing about the bias doesn't remove the feeling of fluency. The remedy is procedural, so delay your self-assessment and make it with the answer hidden.
Do AI summaries make the illusion of competence worse?
Yes, on the best evidence available. The 2025 PNAS Nexus study by Melumad and Yun ran seven experiments with 10,462 participants. People who learned from AI syntheses reported shallower knowledge and produced less original, less useful advice than people who worked through search results. The summary removes the assembly work, and the assembly work is where much of the learning lives.
What's the fastest way to check whether I actually know something?
Close the source and write, from memory, for two minutes. Then compare. Nothing else gives you a cleaner reading, because it's the only check that runs under the same conditions as the moment you'll need the knowledge.
The Feeling Is Not the Evidence
The uncomfortable conclusion from this research is that your sense of learning is a poor instrument, and it fails in a predictable direction. It overreports. It overreports most when you're doing the things that feel most like diligent work.
That's not an argument for abandoning the tools. Rereading has its place, highlighting has its place, and AI summaries are genuinely useful for deciding what deserves your attention. It's an argument for never letting any of them be the last step. Add a gap, close the source, and make yourself produce something. If you can, you knew it. If you can't, you found out cheaply, which is the entire point.
Start with one article today. Read it, highlight three passages with Glasp, write a note on each in your own words, and come back in two days to see what survived. The number that comes back will probably be lower than you expect, and that's the first honest measurement you've taken.
Keep going from there with active recall, which is the technique this whole literature points toward.
