AI can make a difficult text accessible in seconds. It can also become the text the reader actually consumes. An eight-week study of undergraduates found both possibilities inside the same workflow: students asked useful comprehension and reasoning questions, yet often used the generated summary as their primary material and consulted the assigned reading only selectively.[1][2]
The question is therefore not whether an assistant can summarise a chapter. It is whether the interaction leaves the reader able to explain the author’s claim, locate the supporting passage, challenge an assumption and transfer the idea to a new case after the chat is closed.
What 15 students asked across eight weeks
| Measure | Value | Context |
|---|---|---|
| Reading sessions | 239 | The study collected 838 prompts from assigned course reading sessions.Source 1 |
| Comprehension prompts | 59.6% | Requests to explain, summarise, clarify or identify the main idea.Source 1 |
| Reasoning prompts | 29.8% | Requests to compare, analyse, infer or connect ideas.Source 1 |
| Metacognition prompts | 8.5% | Requests that reflect on understanding, strategy or learning.Source 1 |
| Decoding prompts | 2.1% | Requests focused on words, sentences or basic textual decoding.Source 1 |
| Exactly three prompts | 72% of sessions | Three was also the assignment’s required minimum, suggesting compliance rather than sustained exploration in many sessions.Source 1 |
The missing move is usually metacognition
Comprehension is a sensible first use. A concise explanation can remove a vocabulary or background barrier that would otherwise stop progress. The risk appears when the explanation becomes the endpoint. The student may understand the assistant’s restatement while never testing whether it accurately represents the source, which qualification changes the claim or where the author’s evidence is weakest.
The study found a natural movement from comprehension toward reasoning within sessions, but that progression often stopped early. Participants recognised that better prompting took effort yet rarely sustained it, with efficiency dominating their choices. This is a product-design problem as much as a study habit: the assistant makes the shortcut immediate and the deeper reading optional.[1]
A six-pass AI reading workflow
| Pass | Ask the AI to do | What the reader must still do |
|---|---|---|
| 1. Map | List the document’s sections, thesis and defined terms | Open the text and confirm the structure and central claim |
| 2. Locate | Point to passages relevant to one question | Read the contiguous section, not only the extracted sentenceSource 3 |
| 3. Explain | Clarify one difficult concept at the reader’s current level | Restate it unaided and identify what remains unclear |
| 4. Challenge | Generate objections, alternative explanations or missing evidence | Decide which criticism actually follows from the source |
| 5. Retrieve | Quiz the reader with the text closed | Answer before requesting hints or corrections |
| 6. Transfer | Create a new case requiring the same idea | Solve it without copying the source example |
Good document systems locate, then read
DeepRead addresses a related technical failure. Many retrieval systems flatten a document into isolated chunks, losing section order and surrounding argument. Its agent keeps paragraph coordinates and can first locate a relevant passage, then read a contiguous range in the original section. That “locate then read” pattern is closer to serious human reading than extracting one semantically similar fragment.[3]
SparseRead reaches a compatible engineering conclusion from the other direction. Exposing an entire artifact when only a small part is needed wastes tokens and can dilute useful evidence. Its controlled reading layer reduced token volume by as much as 92.9% and wall time by as much as 89% across the reported workloads while preserving or improving task quality. Selective reading can be efficient, but only when it remains source-anchored and has a fallback for uncertainty.[4]
Reading assistance should produce evidence of learning
A useful session ends with more than a polished summary. Keep a short record containing the question, cited passage, the reader’s own explanation, one unresolved objection and the result of a closed-book check. That transforms the interaction from outsourced prose into inspectable learning activity.
AI’s reading potential is substantial because it can supply unlimited clarification, comparison, questioning and practice. The highest-value design does not ask the model to replace the reading. It uses the model to make more of the reader’s attention active: finding the right section, noticing the hard idea, testing understanding and returning to the source with a better question.
Sources & further reading
- How College Students Use AI to Navigate Course Readings: Evidence from an Eight-Week Study Yue Fu and colleagues / arXiv · 26 Jul 2026 · Accessed 10 Oct 2026
- How college students use AI to navigate course readings: evidence from an 8-week study Human–Computer Interaction · 18 Aug 2026 · Accessed 10 Oct 2026
- DeepRead: Document Structure-Aware Reasoning to Enhance Agentic Search Zhanli Li and colleagues / arXiv · 4 Feb 2026 · Accessed 10 Oct 2026
- Read Less, Solve More: Token-Efficient Sparse Reading for AI Agents Zedong Liu and colleagues / arXiv · 23 Aug 2026 · Accessed 10 Oct 2026