How College Students Are Reading in the Age of AI
How social reading is transforming a solitary task into a collaborative learning process. Insights from 20,000+ college courses.
CEO, Hypothesis
7 min read
As we begin a new academic year, I’ve spent much of the last several months in meetings with provosts, deans, and instructional designers at more than 150 colleges and universities. Almost every one of them said some version of the same thing: it’s getting harder to get students to read. In most cases they would add that’s just the beginning.
Headlines echo that concern, warning about declining completion rates or worrying about how generative AI tools encourage passive, machine-summarized consumption in place of real reading. But after those 150-plus conversations, and after spending real time in a full academic year of usage data, I’ve come to believe the deeper issue is structural, not motivational. Traditional course design has long treated reading as an isolated, pre-class activity: something that generally happens alone, off the record, with no visibility into whether it happened at all. Some publicly cited statistics state that less than 20% of all assigned pages are actually read.1 When reading stays invisible and solitary, verifying comprehension and driving real classroom dialogue becomes much harder. When we turn static text into a collaborative, visible workspace instead, reading shifts from an isolated homework chore into the foundation of active inquiry.
Social annotation has a particular role to play here. It functions as a written system of record for the learning process. When you think about it, that is exactly what notetaking was always meant to do. Simply consuming knowledge has never been enough on its own; today’s students need to think critically about what they read, test their interpretations against their peers’, and defend their reasoning. That is what social annotation is built for.
We’re all familiar with the concept of annotating a document—highlighting and taking notes in the margins of the page. Social annotation with Hypothesis extends that capability across all types of academic content and gives students the chance to discuss the document where they are reading it.
Since 2019, Hypothesis has partnered directly with colleges and universities to integrate social annotation directly into their learning management systems. Across our platform’s history, more than 3 million students have generated over 100 million annotations, highlights, and engagement activities across thousands of course documents. That history gives us a unique, data-informed vantage point on how students actually interact with course materials, not just how we assume they do.
What 20,000+ Courses Reveal About Modern Reading Habits
Examining de-identified, aggregate data from the 2025–2026 academic year, spanning 120,000+ assignments with verified student activity, across 20,000+ active courses at our partner institutions, gives us a clear, current picture of reading behavior:
01
Sustained engagement, not just cram sessions: independent research on annotation-enabled reading backs this up directly. In a VitalSource & Hypothesis study, the median number of days students engaged with their course texts was more than three times higher in Hypothesis-enabled sections at UT Austin, and two-and-a-half times higher (10 days versus 4) at the University of Minnesota. Students in Hypothesis-enabled sections returned to course texts across substantially more days, which suggests that collaborative reading can shift engagement toward a more sustained habit.
02
Active, asynchronous dialogue: a substantial share of annotations are direct peer-to-peer replies — students using the margins of a text to ask questions, clarify difficult concepts, and build on one another’s thinking, not just highlight passages for themselves.
03
Sustained pedagogical integration: across these 20,000+ courses, instructors assign a median of 3 social annotation activities per course and an average just under 6, with many building it into weekly modules. Because Hypothesis connects natively via LTI 1.3 into Canvas, Blackboard, D2L Brightspace, and Moodle, faculty can make their existing readings annotatable without forcing students onto a new platform or adding administrative overhead.
Minnesota · Hypothesis
10 days
Minnesota · without
4 days
Top Disciplines: Data, Case Studies, and Faculty Practice
Collaborative reading is no longer confined to literature seminars. Usage spans every major academic division, and course-level data shows sustained adoption across disciplines.
Humanities & Composition
Data & evidence: English/Writing/Literature is our single largest category: 32.4% of all courses with a mapped subject area (5,800+), averaging 5.3 assignments per course. History and Philosophy/Religion, while smaller in volume, show the deepest per-course engagement: History averages 8.7 assignments per course and Philosophy/Religion 9.6 — both well above the 5.9-assignment average across all disciplines.
Case study: at Indiana University Bloomington, first-year composition courses generated more than 78,000 student annotations over a two-year span, a benchmark for what sustained, semester-over-semester adoption in a writing program can look like.
“Hypothesis allows me to suggest the value of slow reading. It encourages close reading and resists the productivity-driven learning that big tech promotes.”— Nick LoLordo, Senior Lecturer, Honors College, University of Oklahoma
Implementation recommendation: use social annotation to scaffold primary source analysis. Prompt students to identify rhetorical strategies and ask divergent questions directly in the text margins. Instructors can assign structured roles such as summarizer, questioner, evidence finder, so every student has a clear path to contribution.
STEM & Health Professions
Data & evidence: STEM & Health Professions courses (Computer Science/Tech, Nursing/Health Sciences, Biology, Chemistry/Physics, and Math/Statistics combined) make up 22.6% of mapped courses, and several of these disciplines carry above-average per-course intensity: Chemistry/Physics averages nearly 8 assignments per course, well past the light-touch role technical fields are sometimes assumed to give reading.
Case study: in a VitalSource & Hypothesis integration study, Viranga Perera, Assistant Professor of Instruction in Physics at UT Austin, integrated Hypothesis into introductory physics and saw the median number of days students engaged with the text more than triple.
“Because students came to class knowing the basics, we could spend class time talking about the meaning of what they read.”— Viranga Perera, Assistant Professor of Instruction, Department of Physics, UT Austin
Implementation recommendation: transform dense journal articles, clinical case studies, and lab manuals into shared workspaces. Prompt students to define specialized vocabulary, dissect complex data visualizations, or flag confusing methodological steps on a shared PDF before stepping into the lab or lecture hall.
Business, Law & Applied Communications
Data & evidence: Business, law and applied communications represented just 8.1% of mapped courses, despite their heavy reliance on document analysis, case interpretation, critique, and evidence-based argument. Social annotation reinforces critical skills like critique, synthesis, structured feedback. These map directly onto what these fields need for the workforce.
Case study: at the University of Minnesota Twin Cities, Professor Allison Brenneise used social annotation in Communication Studies; sections using Hypothesis showed a median of 10 days of engagement versus 4 days in sections without it — a 2.5x increase.
“Students didn’t just read the text — they took positions on it.”— Allison Brenneise, Professor, Department of Communication Studies, University of Minnesota Twin Cities
“They’re engaging with the material directly—and forming their own interpretations—before ever turning to AI.”— Diana Fordham, Instructional Designer and Lecturer in Social Sciences, Missouri Southern State University
Implementation recommendation: have students collaboratively analyze business case studies, legal statutes, or ethical codes. In media and communications courses, faculty can assign students to act as editors and fact-checkers — critiquing published articles or evaluating AI-generated drafts directly on the page.
Three Pillars of Institutional Impact
1
Driving Student Engagement Through Visible Learning
Traditional reading assignments give educators little direct insight into student preparation until work is graded or exams are taken. Social annotation turns reading into an accountable, visible process. By revealing student thinking before class begins, instructors gain real-time diagnostic data on comprehension and gain a jumping off point for targeted, high-value discussion.
2
Helping Students Think Originally and Avoid Overreliance on AI
When large language models can summarize a textbook chapter in seconds, preventing cognitive offloading requires assignments that demand genuine, passage-anchored human feedback. Social annotation anchors student thinking directly to source texts. By requiring students to analyze specific excerpts, evaluate arguments, fact-check claims against evidence, and articulate their reasoning alongside peers, institutions preserve the human-driven critical literacy that automated tools cannot replace.
3
Demonstrating Digital Collaboration to Prepare for the Workforce
Modern professional environments rely heavily on asynchronous, document-centric collaboration across distributed teams. Social annotation trains students to communicate effectively in digital environments. They are taught to offer constructive feedback, synthesize complex information, analyze documents as a group, and co-create knowledge. Building these habits into course design helps students develop the collaborative and analytical competencies the contemporary workforce actually requires.
What Do We Want the Future of Reading in Higher Education to Look Like?
These pillars aren’t abstract. Two recent, widely discussed developments make the stakes concrete.
MIT reached a similar conclusion in August. Its Ad Hoc Committee on AI Use in Teaching, Learning and Research Training called for institutions to reconsider course assessment and design as AI becomes capable of credibly completing much of the work traditionally assigned to students. The harder task is preserving the cognitive struggle of wrestling with a text: not just deciding where AI should be permitted.
Researchers at Wharton (Steven Shaw and Gideon Nave ) recently described a phenomenon they call “cognitive surrender”: a pattern in which high trust in AI tools leads people to defer to machine output over their own reasoning, even when their own reasoning would have gotten them to a better answer. It’s a useful, more precise term for something many of us in higher ed have sensed anecdotally for a while — the risk isn’t that AI answers questions, it’s that people stop checking the answers.
At Brown University this year, an economics professor drew national attention after exam scores in his class appeared to be inflated, and, by his account, the class average dropped sharply once he moved to an in-person, AI-free final. The story spread quickly as a cheating scandal. It’s worth noting that not everyone reads the underlying data the same way: a follow-up piece in Inside Higher Ed argued that the chart from Brown says less about widespread cheating than about how unprepared many courses still are to design assessments for an AI-saturated classroom. I don’t think either extreme — “AI is ruining education” or “this is overblown” — gets us anywhere useful. What both readings of the story agree on is that when students can offload the work of thinking, a large share of them will, unless a course is designed so that thinking has to happen visibly and out loud.
“I use AI tools in my own day-to-day work. The issue isn’t whether we use AI; it’s what thinking we’re willing to outsource to it.”
To be clear, that isn’t an argument that AI is bad for education. It’s an argument that we have to be deliberate about how it’s used. Students need to learn that distinction too.
h.
The question I’d put to every provost, dean, and instructional designer we work with is a simple one: what do we actually want the future of higher education to look like, and is reading, the slow, effortful, argued-over kind, still going to be part of it? Our answer is yes. But it only stays part of it if institutions design for it on purpose.
The Path Forward
Sustained, critical interaction with complex text remains the bedrock of academic success and professional readiness. To the provosts, deans, instructional designers, and faculty partnering with us: thank you for bringing student thinking into the light. We remain committed to building reliable, LMS-integrated solutions that support rigorous instruction, protect student data privacy, and deepen collaborative learning across every discipline.
About the data
This analysis uses de-identified, aggregate activity data from the 2025-2026 academic year covering more than 20,000 active courses and 120,000 assignments with verified student activity at Hypothesis partner institutions. Discipline analysis includes courses for which a subject area could be reliably mapped. The analysis describes activity within Hypothesis-enabled courses and should not be interpreted as a representative sample of all higher education courses.
Sources & case studies
- “Is This the End of Reading?” — The Chronicle of Higher Education (source for the under-20% reading statistic)
- Generative AI and Social Annotation — Hypothesis case study (LoLordo, Fordham, Rigolino)
- VitalSource & Hypothesis Integration — Hypothesis case study (Perera, Brenneise)
- Indiana University Enhances Learning with Hypothesis — case study (Hodgson)
- Report – AI and Education – MIT’s Ad Hoc Committee on
AI Use in Teaching, Learning, and Research Training - Shaw & Nave, “Thinking—Fast, Slow, and Artificial” (SSRN, cognitive surrender)
- Wharton Executive Education, “Thinking Fast, Slow, and Artificially”
- Inside Higher Ed — “Brown Professor Suspects Most of His Class Used AI to Cheat”
- Inside Higher Ed (opinion) — “That Chart From Brown Isn’t Really About Cheating”
Frequently Asked Questions
What data is this analysis based on?
De-identified, aggregate activity from the 2025–2026 academic year: 120,000+ assignments with verified student activity across 20,000+ active courses at Hypothesis partner institutions. It reflects usage within Hypothesis-enabled courses, not a representative sample of all higher ed.
Does social annotation replace reading, or just track it?
Social annotation does not replace reading. It requires students to engage with the actual source text, anchor comments to specific passages, and respond to peers. The goal is to make reading visible and accountable, not to automate it.
How does Hypothesis address concerns about students offloading work to AI?
By anchoring assignments to specific passages and requiring students to articulate reasoning and respond to peers in the margins, social annotation demands the kind of passage-level engagement that a generative summary can’t substitute for.
Which learning management systems does Hypothesis work with?
Hypothesis connects natively via LTI 1.3 to Canvas, Blackboard, D2L Brightspace, and Moodle, so faculty can make existing course readings annotatable without moving students to a new platform.
Is social annotation only useful for humanities or reading-heavy courses?
No — usage spans every major academic division. STEM & Health Professions account for 22.6% of mapped courses, and disciplines like Chemistry/Physics average nearly 8 assignments per course, showing meaningful adoption well outside traditional reading-intensive fields.