Imagine a student opening a difficult essay on a laptop. The first paragraph offers a problem, the second complicates it, and the third introduces a distinction that makes the opening seem less certain. The student stops. Rather than reread the passage, she asks an AI assistant to explain the argument in five bullet points.
Seconds later, the ideas appear in an orderly list. The names are correct. The main conclusion is recognizable. The student moves on.
Has she read the essay?
The question is more serious than a quarrel over which software students are allowed to use. Reading is not merely the transfer of propositions from a page into a brain. It is a meeting with someone else's arrangement of thought: the hesitation before a claim, the example that undermines it, the ambiguity that refuses to disappear, the sentence whose meaning changes after the next paragraph. A summary can convey what a text says. It may not preserve what the text does to a reader.
Yet the opposite story—AI as the enemy of reading—is too simple. A reader who would otherwise abandon a demanding text may use an explanation or translation to get through the door. An overwhelmed student may need a map before exploring the territory. In 2026, we are witnessing not the end of reading but an unsettled negotiation over which parts of reading remain human work.
The crucial question is not whether AI produces a good summary. It is whether the encounter leaves someone more capable of reading, questioning and judging the next text.
A summary is a map with someone else's decisions already made
An argument is not a bag of facts. Its force often depends on sequence.
Consider an essay that begins by defending a policy, introduces evidence against it, and ends by arguing that the disagreement rests on a mistaken definition. A concise AI summary might correctly report the final position. But the reader who encounters only that conclusion misses the author's intellectual journey: why the initial case seemed plausible, what evidence disturbed it, and which assumptions had to be rebuilt.
This matters because sophisticated reading is not simply identifying the destination. It is following—and sometimes resisting—the route.
Summarization necessarily selects. It decides that one passage is central and another dispensable; that an ambiguity can become a settled point; that a counterexample can be compressed into a qualification. Human readers make such decisions too, but they can revisit the original sentences and change their minds. An AI-generated summary can make those decisions invisible, presenting an interpretation in the visual form of an answer.
That does not make summaries useless. Scholars have always relied on abstracts, reviews, indexes, translations and other readers' accounts. The danger emerges when we forget their status: a summary is a representation of a text, not the text itself. Its usefulness depends on the reader's purpose, and its reliability requires comparison with the source.
An AI system can also omit something important or introduce a claim the author never made. The more fluent the prose, the easier it can be to mistake interpretive confidence for fidelity.
The question a reader should ask is not only, “Is this summary accurate?” It is, “What have I not been invited to notice?”
What a British experiment tells us about remembering what we read
There is a practical reason to care about the reading process, beyond literary sentiment. The activity we perform with a text changes what may remain available to us after the screen closes.
In a pre-registered randomized experiment reported in Computers & Education, Pia Kreijkes and colleagues studied 405 students aged 14–15 in seven English secondary schools; 344 were retained for the final analysis. Students worked with history passages using note-taking, an AI chatbot, or AI plus note-taking, then completed comprehension and retention tasks three days later.
The result unsettles an easy assumption about efficient study. On the tested measures, note-taking alone outperformed using the chatbot alone. Combining notes with the chatbot also improved literal retention and comprehension relative to chatbot-only study, although that combination did not show a significant advantage for free recall. Students often experienced AI as easier or more helpful even when note-taking produced better delayed results. Read the research.
This does not prove that handwritten notes are inherently superior to every AI tool, that all kinds of reading benefit equally, or that the effects persist for years. The experiment concerned specific history passages, a particular chatbot, a short study activity and a three-day interval. Its importance is sharper and narrower: the easiest-feeling way to process a text may not be the way that best supports later independent understanding.
Why might note-taking help? A student who creates meaningful notes has to select ideas, connect them and produce a representation of the material. That effort can be educationally consequential. Merely copying sentences into a document would not necessarily do the same work; neither would pasting a chatbot's bullet points into a notebook.
The larger issue is authorship of the mental representation. What has the reader actually constructed?
But the students themselves are inventing another kind of reading
If that experiment were the only evidence, schools might conclude that AI should be kept away from assigned texts. A different study suggests why that would be an incomplete response.
In a paper published in Studies in Higher Education in July 2026, Thomas Corbin and colleagues interviewed 24 philosophy undergraduates at three Australian universities about their use of generative-AI summarizers. The researchers found instances of students replacing reading with summaries. But they also documented four hybrid practices: using AI for orientation, clarification, translation and cross-checking. In those cases, the tool often helped students engage with material they might otherwise have struggled to approach. Read the study.
The distinction is crucial. One student asks a model to substitute for reading because the reading seems dispensable. Another uses the model to understand an unfamiliar historical term so that the original chapter becomes possible to read. A third tests their own interpretation against a competing explanation and returns to the page to adjudicate the difference.
All three used AI. They did not perform the same intellectual act.
The Australian interviews cannot tell us how frequently these practices occur among all university students, nor whether one pattern causes superior long-term learning. Twenty-four participants in a reading-intensive discipline are not the entire student population. Their accounts do, however, challenge the vocabulary of panic. “Using AI to read” is not one behavior. It is a family of behaviors with different purposes and consequences.
The most promising use of assistance is sometimes not to reduce the amount of reading, but to make more demanding reading possible.
The hidden distinction: useful friction and pointless friction
Teachers are often told that learning requires effort. This is true in an important but frequently misunderstood sense. Not all difficulty has equal educational value.
A reader trying to understand the structure of a philosophical argument may benefit from wrestling with an apparent contradiction. The uncertainty is part of the reasoning. Remove it too quickly and the student may lose the occasion to consider competing interpretations.
But a reader unfamiliar with the language of the text may be blocked by vocabulary that is incidental to the intended learning goal. Explaining an obsolete term, translating a passage or providing relevant historical background might reduce that obstacle without stealing the central reasoning.
The same distinction matters for disability and access. Text-to-speech, translation, plain-language supports and other accommodations may be essential conditions of participation. The aim should never be to impose the same difficulty on every learner in the name of purity. Instead, we must ask which capability the reading activity is meant to develop and which obstacles merely obstruct it.
Here is an editorial test worth applying: Does the tool remove a barrier to thinking, or does it remove the thinking the reader was meant to practice?
For a historian studying rival explanations of an event, an AI-generated glossary might be a bridge; a ready-made verdict about which historian is right might be a bypass. For a student learning English, translation might make a scientific paper accessible while reducing opportunities for language practice if used indiscriminately. Context, not a blanket rule, determines the difference.
Some friction deserves preserving because it is where judgment begins. Other friction deserves dismantling because it prevents judgment from beginning at all.
Reading is also learning how an author can be wrong
The most consequential loss may not involve memory. It may concern dissent.
When we read a serious work, we do not merely download statements. We assess the evidence, notice where the author changes the meaning of a term, identify what has been left unexplained, and compare the argument with what we already know. Often the most intellectually productive moment is disagreement.
A chatbot that offers a persuasive summary may create the impression that the controversy has been resolved before the reader has encountered the reasons for disagreement. If the model supplies both the claim and its “balanced” counterargument, the human reader can become an observer of a debate performed by the system rather than a participant in the reasoning.
A 2025 CHI paper by Hao-Ping Lee and colleagues surveyed 319 knowledge workers, collecting 936 examples of generative-AI use. They reported that greater confidence in AI was associated with less self-reported critical-thinking effort, while greater confidence in one's own ability was associated with more. They also found a shift in reported critical-thinking work toward verifying information, integrating responses and supervising tasks. Read the study.
This was a survey of work practices, not an experiment showing that AI damages readers' reasoning. It cannot establish that trusting AI causes lower critical-thinking ability. Yet it raises an essential question for the culture of reading: who remains responsible for deciding whether a generated interpretation deserves belief?
Verification is not a decorative last step. To verify an account of a text, one must be willing and able to go back to the text, locate the evidence and entertain the possibility that the account is wrong.
There is a difference between asking AI what a paragraph means and asking it to identify two plausible interpretations, then checking which interpretation the surrounding pages support. The first may end the conversation. The second can start one.
The real risk is a new inequality of reading
It would be too comfortable to describe AI-assisted reading solely as an individual's habit. Students read under radically unequal conditions.
Some have quiet rooms, experienced teachers, family support, accessible books and time to revisit an argument. Others combine coursework with paid work, caregiving, language barriers or overloaded schedules. A summary may be the only feasible entry point on a difficult day.
This produces a troubling possibility. Students who already possess strong disciplinary knowledge may use AI as an efficient critic and navigator. Students with weaker foundations may use it primarily as a substitute for texts they have never been taught how to interpret. In that case, the same technology could widen an existing gap in the ability to read independently.
That is a plausible risk, not an established universal trend. We should resist turning it into an unverified prediction. But it shifts the policy question. Merely prohibiting chatbots will not correct unequal access to reading instruction, and merely distributing chatbots will not ensure that students learn to interrogate what they receive.
A serious response would teach the practices that capable readers often perform silently: identifying the central claim, distinguishing evidence from interpretation, tracing a disagreement, questioning a translation, and returning to the source when a summary feels too neat.
These practices are teachable. They should not remain privileges acquired by accident.
What the next generation's reading lesson could look like
Imagine a class reading two opposing essays about whether schools should permit generative AI in assessed work. Rather than demanding that every student either use or avoid the technology, the teacher organizes the activity around a sequence of encounters with the source.
First: encounter the author's voice
Students read a selected passage without receiving an AI summary first. They underline the claim they find hardest to accept, write one question about the evidence and identify a sentence that changes their interpretation. Those struggling with language or access receive appropriate support.
The purpose is not to prove independence by withholding help. It is to give students an initial position from which they can judge later assistance.
Second: bring the machine into the argument
Students ask AI to produce a short account of the author's position, then identify what it leaves out. They might request a rival interpretation of an ambiguous paragraph, a definition of an unfamiliar term, or an objection to their own reading.
Now AI is not a surrogate author of the student's response. It is an additional interpreter whose work must be tested.
Third: return to the original sentences
Students locate the passages that support or contradict the model's explanation. They discuss why two summaries can differ while each sounds authoritative. They distinguish a direct claim from an inference.
The habit of returning is the point. Without it, verification risks becoming another button that promises certainty.
Fourth: make a judgment without the answer on screen
Each student explains the argument in their own words, states a defensible objection and identifies what evidence would change their view. Days later, they revisit the question without the summary beside them.
This is a proposed classroom design, not a validated protocol with a guaranteed effect size. It draws on the difference between passive access and active interpretation illustrated by the studies above. Its educational promise is to make reading visible as an act of reason rather than an act of consumption.
One could call this source–assistant–source reading: begin in the original, use the assistant strategically, and return to the original before concluding.
The book beside the argument
Learning Coach Editions, the publisher behind Books & Ideas, has a declared interest in the problem of learner autonomy. Its English-language book You Were Never Taught How to Learn approaches study, comprehension, memory and the role of AI from the learner's perspective. Its documented outline includes the illusion of understanding, how AI can help or hinder learning, and the development of independent study habits. The edition is listed as published in the Learning Coach catalog; readers can find it in the English books collection.
The book offers an adjacent educational perspective, not independent scientific verification of the research reviewed here. The broader question reaches beyond any single Learning Coach title: if a young person can obtain a plausible explanation of almost anything on demand, how do we help that person develop the judgment to know when an explanation is incomplete, superficial or wrong?
The answer is not to protect students from tools. It is to give them the intellectual habits with which to use tools without surrendering their agency.
What remains of the human reader
The future reader may encounter a book surrounded by explanations, instant translations, voice interfaces, summaries, annotations and conversational guides. Some of those tools will open extraordinary new doors. They may allow someone to enter an unfamiliar discipline, approach a text in another language or discover an author they would never have found alone.
But the reader must still decide what deserves attention. The reader must still feel the force of an objection, recognize an unanswered question, and choose which argument to carry into the world. No summary can have those responsibilities on a person's behalf.
Perhaps the deepest mistake is to assume that reading was always about minimizing the distance between a question and its answer. Sometimes that distance is where the reader develops a mind of their own.
We should welcome AI that helps more people enter difficult books. We should be wary of AI that persuades them they have finished reading before the argument has begun.
A generation reading with AI will still need to learn how to read against it.
Sources and editorial notes
Kreijkes, P., et al. (2026 print issue; online 2025). “Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools.” Computers & Education, 243, 105514. DOI. Randomized experiment, 405 recruited; 344 retained for analysis; outcomes measured after three days.
Corbin, T., et al. (2026). “Hybrid reading practices: how GenAI summarisers can support rather than replace reading.” Studies in Higher Education. DOI. Qualitative interviews with 24 philosophy students at three Australian universities; not a population estimate or causal trial.
Lee, H.-P., et al. (2025). “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” Proceedings of CHI 2025. DOI. Survey associations and self-reports; not causal evidence of skill loss.
