Highlighting Statistics: What the Research Shows

The headline numbers

If you only take one table away from this page, make it this one.

Each of these numbers has a story behind it, and a few of them are widely misquoted. The sections below give the context.

Does highlighting work? What the lab research says

The most cited verdict on highlighting comes from Dunlosky and colleagues' 2013 monograph in Psychological Science in the Public Interest, which reviewed ten popular learning techniques. Highlighting and underlining received the review's lowest rating, "low utility," alongside rereading and summarization. Practice testing and spaced practice earned the top marks.

Two details usually get lost when this finding is quoted.

First, the studies behind the rating tested highlighting as a passive, one-pass activity: mark some lines, reread them later, take a test. Under those conditions, marking text adds little over plain reading. The review did not test highlighting as a capture step inside a larger system, where the highlights feed later retrieval practice, notes, or spaced review.

Second, the review found that highlighting can even hurt performance on inference questions, and that benefits appear mainly when marking is selective. Marking whole paragraphs spares the brain the work of deciding what matters, and that decision is where the learning happens. Selectivity is the active ingredient, a point we unpack in The Science of Highlighting.

So the honest reading of the lab literature is narrow: highlighting alone does not produce durable learning, highlighting as the front door to active recall and review is a different, largely untested activity.

How people actually highlight: data from millions of highlights

Lab studies typically watch dozens of students read a few passages. Glasp operates a social highlighting platform with over 1,000,000 users and millions of highlights, and publishes peer-submitted research on that behavioral data. The full series lives on the Glasp Research page; three findings matter most here.

What people mark inside a document is mostly shared. Across readers of the same document, highlights cluster on the same passages. A model that knows only what the crowd marked predicts an individual's next highlight almost as well as a model tuned to that individual: the personal lift within a document is tiny, about +0.017 AUC (Paper I, arXiv:2606.09024).

Individuality lives in document choice. The same modeling shows a much larger gain, roughly +0.13, when predicting which documents a person will engage with at all (Paper II, arXiv:2606.10398). In plain terms: you are not special in what you underline, you are special in what you decide to read.

Readers form factions, but only within a document. On contested texts, readers split into distinct sub-groups that mark different regions, with strong statistical separation (z ≈ +6.3). Yet membership in those factions barely carries from one document to the next (Paper III, arXiv:2606.11613). Disagreement about what matters is real, but it is situational, not an identity.

Can highlights be predicted?

Two more results from the same series answer practical questions.

Cold-start prediction works, on the long tail. For a brand-new document with no readers yet, a trained model can predict which passages the crowd will eventually highlight, beating the strong "people highlight the opening" baseline by about +0.04 average precision. The advantage concentrates on the long tail, the less-popular passages where simple heuristics fail (Paper IV, arXiv:2606.11654).

Reading identity is durable. A profile of a reader's interests frozen after their first six months predicts their document choices years later with no measurable decay: the paired retention ratio is 1.003, and the frozen profile still beats a popularity baseline by more than 3 to 1 on next-document prediction (Paper V, arXiv:2606.12904). Your reading taste behaves like a trait, not a mood.

For anyone building a personal knowledge management system, that last number is quietly encouraging: the interests you capture today will still describe you in two years, so the library you build compounds instead of expiring.

Reading speed and attention: the context highlighting lives in

Highlighting statistics only make sense against the backdrop of how people read.

Reading speed. The most careful estimate comes from Brysbaert's 2019 meta-analysis of 190 studies: adults silently read non-fiction at about 238 words per minute on average, with wide individual variation. Claims of ordinary readers tripling that speed without losing comprehension have not survived testing.

Attention. Gloria Mark's screen-attention research at UC Irvine tracked knowledge workers over nearly two decades: average attention on a single screen fell from around two and a half minutes in 2004 to about 47 seconds in recent years. The famous "8-second attention span, shorter than a goldfish" statistic, by contrast, has no traceable primary source and is best treated as a myth, something we dig into in The Attention Span Crisis.

Put together: a 5,000-word article costs the average reader around 20 minutes of increasingly fragmented attention. A good highlight layer is how any of that investment survives the next tab switch, which is the practical case for tools like a web highlighter that keep the capture step at zero friction.

AI and studying in 2026

The environment around highlighting changed faster in the last three years than in the previous thirty.

The HEPI / Kortext Student Generative AI Survey 2026 puts numbers on it for UK undergraduates: 94% now use generative AI in some form for assessed work, up from 88% in 2025 and 53% in 2024. The share who paste AI-generated text directly into assessed work reached 12%, up from 8% the year before.

The lab literature has not caught up with this shift, and the honest statistical statement is that there is no long-run evidence yet on learning effects. What the behavioral data above does suggest is a division of labor: AI is strongest at compressing what a document says, while your highlights record the only thing AI cannot know, which is what mattered to you. Summaries from an AI YouTube summarizer tell you whether a source deserves attention; your highlights are the residue of having paid it.

What the statistics mean for your study system

Reading the numbers together points to a specific way of working rather than a verdict for or against the highlighter.

1. Highlight selectively, then do something with it. The "low utility" rating attaches to mark-and-reread. The evidence-backed techniques, retrieval practice and spacing, need raw material, and highlights are the cheapest way to collect it.
2. Spend your effort on selection. The data says document choice is where individuality lives. Curating what enters your reading queue matters more than perfecting your marking style.
3. Trust your library to age well. Reading identity is stable over years, so highlights captured today keep their value; export them somewhere permanent, whether that is Kindle highlights into your notes app or a connected MCP setup that lets your AI assistant search them.
4. Assume fragmented attention. With sub-minute screen attention, systems that require long uninterrupted sessions fail by design. Capture in seconds, review in minutes.

Frequently Asked Questions

Is highlighting a waste of time?

On its own, mostly yes: the Dunlosky review rated mark-and-reread "low utility." As a capture step feeding retrieval practice, spaced review, or notes, the criticism no longer applies, because the studies behind the rating were not designed to test that workflow.

What percentage of students use AI for schoolwork?

The HEPI / Kortext 2026 survey found 94% of UK undergraduates use generative AI for assessed work in some form, up from 88% in 2025 and 53% in 2024. Direct inclusion of AI-generated text in assessed work stands at 12%.

How fast does the average person read?

About 238 words per minute for silent non-fiction reading, per Brysbaert's 2019 meta-analysis of 190 studies. Fiction runs slightly faster, and large individual differences are normal.

Do people highlight the same things?

Largely, yes. Glasp's published research finds that within a document, an individual's highlights add only about +0.017 AUC of predictive signal over what the crowd already marks. Individuality shows up in which documents people choose, not which lines.

Where do these highlighting statistics come from?

The lab findings come from peer-reviewed psychology research, primarily Dunlosky et al. (2013). The behavioral findings come from Glasp's published research series on its own platform data, available with code and reproducibility bundles at glasp.co/research and on arXiv.

Conclusion

The statistics on highlighting resolve an old argument by splitting it. The lab is right that marking text and rereading it teaches little. The behavioral data is right that highlights are far from noise: they are consistent enough to predict, stable enough to define a durable reading identity, and personal precisely in what they are attached to. The tool was never the problem; the missing second step was.

If you want the numbers behind this page, every Glasp study cited here is public, with data and code, at glasp.co/research. And if you want your own reading to start leaving a trace, Glasp is where those millions of highlights came from.