Reading Was Always Social: What Millions of Shared Highlights Reveal

The Myth of the Solitary Reader

Picture "a reader." You probably see one person, alone, silent, maybe in an armchair under a lamp. That image feels timeless. It isn't.

Around 400 CE, Augustine described watching his teacher Ambrose read: "his eyes ran over the columns of writing and his heart searched out the meaning, but his voice and his tongue were at rest." The passage from the Confessions is often cited as one of the earliest descriptions of silent reading in Western literature. Scholars debate how unusual it really was, but the broader picture holds: for centuries, the normal way to encounter a text was out loud, with other people in the room.

The private, silent, solitary reader is a historically recent figure, a product of cheap printed books, mass literacy, and rooms of one's own.

This essay makes one claim, that reading is and has nearly always been a social behavior, and backs it three ways: the historical record, controlled classroom studies, and Glasp's published research on millions of shared highlights, which found the social structure of reading inside the most private-seeming gesture a reader makes.

A Short History of Reading Together

The history of reading is mostly a history of company. A compressed tour:

The chavruta, a tradition with roots in the Babylonian Talmud, treats a text as something you can only fully understand through a partner who disagrees with you productively. Victorian families gathered around the good lamp to hear the newest Dickens installment read aloud, often the evening it was purchased; Dickens wrote with that listening audience in mind.

Notice the trajectory. Reading went from communal by necessity (few texts, fewer literate people) to communal by choice (book clubs, seminars) to, briefly, solitary by default. Marginalia kept a thin social channel open even then, but for most of the 20th century, what you noticed while reading died with your copy of the book.

Then highlights went online, and the social layer became measurable for the first time.

What Millions of Highlights Actually Show

Here's where this stops being a history essay and becomes a data story.

Glasp operates a social web highlighter used by hundreds of thousands of readers who have collectively saved millions of highlights across the open web. Because many readers highlight the same articles, the data can answer a question the solitary-reader model never could: do two people reading the same text mark the same things?

In June 2026, Glasp published two research papers on this question, both publicly available on arXiv. The first, "Personal Salience: Highlighting Is Social, but Individuality Lives in Selection" (arXiv:2606.09024), used a co-readership design: it compared readers specifically on documents they had both read, removing the easy confound of people simply reading different things. It separated three layers of the highlighting decision: generic salience (what any reader tends to mark), crowd salience (what this community converges on), and personal salience (whatever remains that's unique to you).

The headline result is humbling for anyone who thinks of their highlights as a fingerprint. Highlighting turned out to be highly social: models built from crowd signals predicted which sentences a reader would highlight better than models personalized to that individual reader. At the salience layer, the own-versus-other gap on shared documents was tiny, around +0.017. Readers on the same page largely agree about which sentences deserve the marker.

The second paper, "Selection, Not Salience: The Shape and Limits of Personalization in Social Highlighting" (arXiv:2606.10398), stress-tested the practical consequences. At the sentence level, personalized auto-highlighting never beat impersonal baselines; even on generous candidate pools, re-ranking by personal taste lost to plain salience order. The papers are candid about method, too: naive evaluation setups leaked near-duplicate text in roughly 42 percent of pairs, inflating personalization scores by up to +0.15 average precision. Strip the leakage out and the individual salience signal nearly vanishes.

Read that finding the way the papers do, not as "readers are interchangeable," but as "salience is shared." When thousands of people independently mark the same sentence, that's not conformity. Nobody is watching anyone else's cursor. It's convergent judgment, the same phenomenon that makes the chavruta work: separate minds, same text, overlapping sense of where the weight sits. The marks we make while reading are less like signatures and more like votes, and votes only mean something in a population. What that convergence looks like in aggregate is the subject of the Curiosity Graph; this piece is about the act itself.

Selection Is the Self

So if everyone marks roughly the same sentences, where did your individuality go? The research found it, just not where most people assume.

Both papers converge on an asymmetry. The individual signal in salience (which sentences you mark within a text) is whisper-thin: that +0.017 gap. The individual signal in selection (which texts and passages you choose to engage with at all) is roughly eight times larger, around +0.14 at the span level. At the document level it's stronger still: a reader's history identified which documents were theirs with an own-versus-other gap of +0.169 against community negatives, and +0.119 even against topic-matched hard negatives. When the first paper decomposed that selection signal, most of it was stable thematic preference: your topics, your beats, your recurring questions.

In plain terms: you are what you read, much more than which lines you underline once you're reading it. Identity lives in the choosing. The second paper puts overall measurable personalization at roughly +0.13, topic-dominated, with no reliable gain at the salience layer.

This reframes what a highlight is for. A single highlight tells you mostly what any attentive reader would have marked. A history of highlights, spread across hundreds of chosen sources, tells you who the reader is. The individual gesture is social; the accumulated trail is personal. That's exactly the structure you'd expect if reading were a social act performed by individuals with distinct curiosities, which is what the historical record suggested all along. For what the marking gesture does cognitively, see the science of highlighting.

One caveat the papers insist on: the numbers are benchmark gaps for one platform's readers, not universal constants. The qualitative shape, social salience and personal selection, is the finding that matters.

The Classroom Evidence

Education researchers reached the same conclusion from a different direction: make reading social on purpose, and people read more and learn more.

The cleanest study is Miller, Lukoff, King, and Mazur's 2018 paper in Frontiers in Education, run in a flipped introductory physics course at Harvard. Students did their pre-class reading on Perusall, a social annotation platform where they could see and respond to classmates' questions in the margins. The comparison cohorts, demographically similar classes from prior semesters, had used a simpler annotation tool without the social features.

The results: most students actually did the reading, with about 80 percent getting through at least 95 percent of it, far above what the pre-class reading literature usually reports. And the social annotation cohorts scored 5 to 10 percent better on all but two of the ten in-class exams, with effect sizes around 0.3. Same course, same flipped design, same kind of students. The difference was whether reading happened alone or in the visible company of other readers.

Remi Kalir and Antero Garcia's book Annotation (MIT Press, 2021) supplies the frame for why. They define annotation as an everyday and social activity, one that "provides information, shares commentary, sparks conversation, expresses power, and aids learning." From medieval rubrication to modern platforms, their through-line is that a note on a text has almost always been a note to someone, even a future stranger.

Put the evidence side by side. History says we read together for millennia. The classroom data says restoring the social layer improves effort and outcomes. The highlight data says that even readers acting alone converge on what matters. Reading was always social. We just briefly lost the instruments to see it.

Why the Social Layer Matters in the AI Era

This would be a pleasant academic point if it weren't suddenly load-bearing.

AI assistants have made solitary text consumption nearly frictionless. Any article, paper, or book can be compressed into bullet points in seconds, no other human involved at any step. Notice what that removes: every trace of other readers. No marginalia, no shared marks, no sense of where ten thousand previous minds slowed down. The summary is a private tunnel through a text that used to be a commons.

And here's what the research makes precise: the social layer is exactly the part AI can't replicate. A model can approximate generic salience; that's roughly what a summary is. What it cannot produce is the fact of convergence, the record that particular humans, with reading histories and stakes and taste, independently decided a passage was worth keeping. Crowd salience is an empirical fact about people, not a property of text. You can't prompt your way to it.

That makes shared highlights a strange kind of resource: more valuable as synthetic text gets cheaper, because they're one of the few signals still anchored to human attention. The Curiosity Graph makes this argument at the aggregate level; collective intelligence makes it for groups of readers. If reading's value increasingly lies in the human layer, reading socially stops being a nostalgic preference and becomes the rational strategy.

Learning in Public Is the New Chavruta

The chavruta survives because it solves a real problem: a text read alone can only tell you what you already brought to it. A partner forces the second reading.

Most of us won't sit across a table from a study partner for two hours a day. But the modern equivalent doesn't require scheduling. When your highlights are public, every reader who encounters them gets a quiet chavruta partner, a record of where another mind paused, and you get the same from theirs. A reader in 2026 marking the same paragraph you marked in 2024 is completing a loop the Talmudic study hall would recognize.

This is the deeper case for learning in public. Public learning is usually pitched as career advice: show your work, build an audience. The reading research suggests something less transactional. Your selection trail, the documents and passages you chose, is the most individual thing about your reading life. Publishing it is publishing the useful part: a curriculum nobody else could have assembled. A Glasp profile works as exactly this, a public commonplace book where the selection is the authorship.

The Victorian family shared one text with the room. The public reader shares a reading life with anyone who finds it. Same instinct, better reach.

How to Read Socially Online Today

Here's the practice. None of it requires abandoning solitary deep reading; it wraps a social layer around it.

Highlight where others can see it. The minimum viable social reading is making your marks public by default. Glasp's web highlighter does this for the open web. A private highlight helps you. A public one helps you and every subsequent reader.

Read the layer before you leave. After your own pass through an article, check what other readers marked. You'll reliably find confirmation (the crowd marked your passage too) and the miss, the one passage someone else caught that you skimmed. The miss is the chavruta moment.

Follow readers, not just feeds. Feeds optimize for reaction. A reader you follow is a selection engine, and selection, per the research, is where the individuality is. Three or four readers whose curiosity overlaps yours will surface better material than any recommender. The community page is built for finding them.

Form a tiny annotation circle. The Miller study's effect didn't come from a crowd; classes of 70-something students saw it. Two or three colleagues reading the same material with shared annotations recreates the condition for free. Pick one shared text a week, and leave short notes, not just marks: a highlight says "this mattered," a note says why.

Let your profile become your commonplace book. Don't curate it for an audience. Sanding your trail down to look impressive removes exactly the thematic texture that makes it yours. Highlight what actually stops you.

Frequently Asked Questions

What is social annotation?

Social annotation is the practice of marking up texts (highlights, notes, questions) in a shared layer where other readers can see and respond to the marks. It ranges from classroom platforms like Perusall to open-web tools like Glasp. Kalir and Garcia's Annotation (MIT Press, 2021) frames it as the modern form of a very old behavior: notes on texts have worked as communication between readers since medieval manuscripts.

Do people highlight the same things?

Largely, yes. Glasp's published research (arXiv:2606.09024) compared readers on documents they had both read and found highlighting highly social: crowd-based models predicted a reader's highlights better than models personalized to that reader, and the individual signal in which sentences get marked was very small (about +0.017). Individual differences show up strongly in which texts people choose to read, not in which lines they mark.

Wasn't silent solitary reading always the norm?

No. Reading aloud, often in company, was the common mode for much of antiquity and the medieval period; Augustine found Ambrose's silent reading notable enough to describe around 400 CE. Communal forms persisted for centuries afterward, from monastic refectory reading to Victorian family circles. Solitary silent reading as the default is only a few centuries old.

Does social annotation actually improve learning?

The best-documented case is Miller, Lukoff, King, and Mazur (2018, Frontiers in Education): in a flipped Harvard physics course, cohorts using a social annotation platform for pre-class reading completed far more of it and scored 5 to 10 percent higher on nearly every exam than comparable prior cohorts using a non-social tool, with effect sizes around 0.3.

If my highlights match everyone else's, why highlight at all?

Two reasons. First, the act itself aids your own encoding and retrieval regardless of who else marked the passage; see the science of highlighting. Second, the research locates your individuality in selection: your accumulated trail of chosen sources and passages is several times more distinctive than any single mark, and each highlight is one entry in that trail.

Conclusion

The solitary reader was a two-century experiment, and a productive one. Deep private reading isn't going anywhere, and nothing here argues it should. But three independent sources point the same way. History says reading was communal for most of its existence. Classroom studies say adding the social layer back makes people read more and learn more. And Glasp's published research on millions of highlights says the social structure never left: readers converge on what matters within a text, and express who they are through what they choose to read.

Your highlights are votes in a shared judgment. Your reading choices are the signature. Both get more valuable when they're visible, to you, to other readers, and to a web filling up with text no human chose.

The practice costs almost nothing. Read something worth your attention today, mark the passage that stops you with Glasp's web highlighter, make it public, and see what previous readers caught that you didn't. Then browse the community and follow two readers whose trails overlap yours. The study hall, the parlor, and the margin were never about the room. They were about reading in the company of other minds, and that's available again.