The Highlighter Illusion
Why coloring a page feels like learning — and usually isn't
There was always one classmate who highlighted everything. Four markers in the pencil case, every paragraph read twice, then color-coded: yellow for definitions, green for examples, orange for “important,” pink for “very important.”
By the end of a study session, the page was more color than text — which raises an obvious question that’s easy to dismiss as a schoolyard observation but turns out to be a genuine cognitive-science puzzle: what did any of that actually accomplish?
The research answer is deflating: almost nothing, at least not on its own. But the mechanism behind that answer is far more interesting than “highlighting doesn’t work” — and it explains why students, professionals, and programmers alike routinely mistake the feeling of understanding for the fact of it.
In short: highlighting feels productive because recognizing familiar text is mistaken for actually knowing it — a bias called the fluency illusion. The fix isn’t a better color-coding system; it’s forcing yourself to retrieve, explain, or generate information without the source in front of you.
Below: the research behind the illusion, what it has to do with rubber duck debugging, and five techniques — several of them digital — that reliably outperform a highlighter.
What the evidence actually says
In 2013, cognitive psychologist John Dunlosky and colleagues Katherine Rawson, Elizabeth Marsh, Mitchell Nathan, and Daniel Willingham published one of the most influential papers in learning science: a systematic evaluation of ten popular study techniques, rated by the strength of the experimental evidence behind them. Highlighting and underlining received a “low utility” rating — the lowest of three tiers. At the top, rated “high utility,” sat two very different strategies: distributed practice (spacing study sessions over time) and practice testing, also known as retrieval practice.
The problem with highlighting isn’t that it’s actively harmful — it’s that it demands almost none of the cognitive effort that actually leaves a trace in memory. Coloring a sentence yellow doesn’t require explaining why it matters, restating it in your own words, or later trying to recall it without the text in front of you. It only records that something looked important — not what it means or how it connects to anything else you know.
It’s worth being precise about what Dunlosky’s team actually claimed, because “low utility” is not the same as “useless.” Highlighting can be a reasonable first pass for identifying what to come back to later — the failure mode is treating it as the studying itself, rather than as a way to select material for the retrieval practice that should follow. A highlighted sentence that never gets tested is no better than an un-highlighted one; the marker only pays off if something more demanding happens next.
Retrieval fluency: when recognition masquerades as knowledge
The mechanism that makes highlighting feel productive anyway has a name in the literature: the fluency illusion, sometimes called the illusion of knowing. The core finding, developed over decades by Robert Bjork and Elizabeth Bjork, is that people systematically confuse the ease with which information is processed with genuine understanding of it.
Look at an already-highlighted passage a second or third time, and something deceptive happens: your brain recognizes it — “I know this” — and that sense of recognition gets misread as evidence of learning. But recognition and retrieval are fundamentally different cognitive operations. In the textbook, the answer sits right in front of you. On the exam, or in production code at 2 a.m., it doesn’t. Bjork and Bjork capture this with their distinction between storage strength and retrieval strength: rereading or highlighting a passage boosts retrieval strength in the moment, because the material is present and familiar, but does little to build storage strength — the deeper trace that supports recall once the cue disappears. Take the highlighted page away, and the illusion of mastery collapses, because there was never much underneath it.
This shows up empirically too. Jeffrey Karpicke, Andrew Butler, and Henry Roediger found that most students default to exactly the strategies that work least well — rereading, reviewing notes, highlighting — and rarely test themselves through active recall, even though retrieval practice produces measurably better long-term retention in controlled studies. The reason students stick with the weaker methods is, again, the fluency illusion: retrieval practice feels harder and less successful in the moment, precisely because you can fail at it — and that friction is the point.
Running it through the DIKW pyramid
There’s a useful way to frame exactly what highlighting does and doesn’t do: the DIKW pyramid — Data, Information, Knowledge, Wisdom — a model introduced by systems theorist Russell Ackoff in 1989 and since widely used, and widely debated, in information science and knowledge management to describe how raw material gets transformed into something usable. It’s worth treating it as a heuristic rather than a settled scientific model — critics rightly point out that the boundaries between the four levels are fuzzy — but as a way of naming where a study technique intervenes, it’s a genuinely useful lens.
Data: the raw text on the page — words, numbers, symbols, uninterpreted.
Information: data organized enough to carry meaning — the sentence you just read and parsed.
Knowledge: information integrated into a structure you can use — connected to what you already know, applicable to a new problem, retrievable without the source in front of you.
Wisdom: knowing when and how to apply that knowledge — judgment.
Highlighting operates almost entirely at the information layer. You’ve registered that a sentence is meaningful and marked it as such — but marking is not integrating. Nothing about the act of dragging a colored marker across a line forces the climb from information to knowledge, which requires connecting the idea to something else, restating it, testing whether you can reproduce it, or using it to solve a problem you haven’t seen before. A page covered in yellow can look like a dense, well-processed artifact while representing almost no movement up the pyramid at all. That’s precisely why it feels productive — you’re visibly doing something to the information — while doing very little of the work that turns information into knowledge.
Retrieval practice, by contrast, only works if you’ve climbed at least one level: you cannot generate an answer from memory using structure and meaning you never built. The technique fails immediately, visibly, and informatively if you skipped the knowledge-building step — which is exactly the diagnostic value it has over highlighting, where failure is invisible until the moment it costs you (an exam, a demo, a production incident).
That classmate with four highlighters, in other words, was doing careful, colorful work entirely at the information layer — and mistaking the density of color for the density of knowledge.
Desirable difficulties: why what feels hard often works better
This connects to one of modern learning psychology’s most influential and counterintuitive ideas, which Bjork named in 1994: desirable difficulties. Conditions that make performance improve quickly and smoothly during learning often produce worse long-term retention. Conditions that make learning feel slower and harder in the moment — spaced rather than massed practice, interleaved rather than blocked practice, free recall rather than recognition — often produce much more durable, transferable knowledge.
The critical distinction is between performance — what you can observe during a study or practice session — and learning — a durable change in what you know or can do. In everyday experience, these two collapse into one feeling. That collapse is the trap: finishing a highlighted chapter with “okay, that’s locked in” is mistaking a short-term performance bump for real learning.
From the library to the terminal: the same illusion, different medium
What’s true of a highlighted textbook is true, almost line for line, of reading someone else’s code. Scroll through a codebase nodding along because each line parses cleanly, and you’re experiencing the same retrieval fluency: you recognize the syntax and the patterns, and that recognition gets mistaken for the ability actually to solve the problem or find the bug yourself. That’s the cognitive backdrop for why rubber duck debugging works as well as it does. Explaining code line by line to a rubber duck — or a colleague, or an empty chair — forces exactly the kind of complete, sequential articulation that highlighting never demands. As I described in that piece, the technique works because “human cognition operates efficiently by taking shortcuts, but these shortcuts often bypass the precise reasoning needed to solve complex problems,” and because “you cannot skip steps when explaining to someone else without recognizing you’re doing so.”
Put in DIKW terms, rubber duck debugging is a forcing function for the climb from information to knowledge. You can silently reread a function and feel like you understand it — pure information processing. The moment you have to say out loud, in complete sentences, what each line does and why, the illusion breaks exactly where your actual understanding runs out. That’s not a coincidence; it’s the mechanism. Self-explanation is itself one of the moderately effective techniques in Dunlosky’s review, precisely because it forces learners to make gaps explicit instead of leaving them hidden under a layer of familiarity.
What actually works — including digitally
The alternatives to the highlighter are well-studied, mostly free, and in several cases easier to execute today than they were with index cards.
Retrieval over rereading. The single most effective lever is testing yourself rather than re-consuming material — the core idea behind spaced-repetition flashcard apps like Anki or Quizlet. These tools combine two “high utility” strategies at once: retrieval practice (you must generate the answer before seeing it) and spacing (the scheduling algorithm resurfaces cards right as forgetting sets in).
Spacing instead of massing. An all-night highlighting session is the opposite of distributed practice. Even a simple recurring calendar reminder — one that prompts you to revisit a topic weeks later, from memory, before looking at your old notes — exploits the same spacing effect that underlies the research.
Interleaving. Instead of mastering one topic before moving to the next, effective study plans mix topics or problem types. It feels worse in the moment — performance dips, confidence drops — but produces measurably better transfer—another desirable difficulty.
Explaining instead of marking. Whether it’s rubber duck debugging for code, the Feynman technique for a textbook chapter, or narrating a research paper out loud as if teaching it to someone new, the common thread is externalizing your reasoning and letting every gap show. Digitally, a voice memo or a short, unedited screen recording works just as well as an actual duck — the microphone is the duck.
Generation over consumption. Trying to reconstruct a chapter’s key ideas, draw its structure from memory, or write your own quiz questions in a blank document — with no source material open — is active generation rather than passive recognition. This is the same generation effect Slamecka and Graf described back in 1978, and it’s a reliable way to force the climb from information to knowledge that highlighting skips entirely.
Weaving highlights into a network of ideas. One genuine way to rescue a highlight from the information layer is to stop treating it as a static mark and start treating it as a node you have to connect to something else. This is the logic behind tools like Heptabase, which pulls highlights in from read-it-later apps like Readwise and lets you drag them onto an infinite whiteboard next to highlights from other books, articles, and your own notes — forcing you to ask, explicitly, how this idea relates to that one, whether they agree, and what new idea sits between them. I’ve written about this workflow in more detail in a piece on Heptabase. Crucially, the highlight itself still isn’t the learning — dragging it onto a board and never touching it again is just highlighting with extra steps. The learning happens in the act of drawing a connection, writing a sentence explaining why two highlights belong near each other, or noticing that two sources contradict each other. That act is a form of elaborative interrogation and self-explanation, both rated moderately effective by Dunlosky’s review, and in DIKW terms it’s the one move that reliably turns information into knowledge: structure that didn’t exist in either source on its own, built by you, and therefore genuinely yours to retrieve later.
Arguing with an LLM instead of reading past it. A more recent addition to the toolbox, not yet covered by the classic 2013 literature but consistent with everything it found: using a language model as a sparring partner rather than a summarizer. Asking a chatbot to explain a concept back to you is barely better than rereading — it’s still someone else doing the cognitive work. Asking it to challenge you is a different exercise entirely: explain the concept yourself and have the model probe for gaps, pose counterexamples, or argue the opposite position and require you to rebut it. This is functionally a tireless, infinitely patient version of the Socratic method, and it inherits the same mechanism that makes rubber duck debugging work — except the “duck” talks back, which means it can catch a wrong assumption a silent rubber duck never could. Used this way, the model isn’t a source of answers to highlight; it’s a generator of the desirable difficulty that highlighting was always missing. The failure mode to watch for is the same one that undermines any tool: if you let the model do the explaining and nod along, you’re back to information-layer processing with better production values.
The actual point
That classmate with four highlighters was probably chasing something real: the sense that studying should mean engaging with a text, not just passing your eyes over it. She just replaced that engagement with a ritual that feels like thinking without doing the cognitive work that actually leaves a trace. The uncomfortable but well-supported conclusion from metacognition research is this: the feeling of understanding is not a reliable signal of whether you can retrieve that understanding when it counts — in an exam, in a live demo, or while debugging a production incident at 2 a.m. The only reliable way to find out is to actually try, without the safety net of the source in front of you, and to be willing to fail at it.