The AI Reading Paradox
You can have ChatGPT summarize a 10,000-word investigative piece in about ten seconds. You'll "know" what it said. You'll be able to describe the thesis, name the main characters, maybe list three takeaways. You'll feel informed.
A week later, if someone asks you about it, you'll remember almost nothing.
Now contrast that with an article you read slowly last year. You highlighted three passages. Maybe you argued with the author in the margin. Two years later, you're still quoting lines from it. You can find the link from memory. The ideas snuck into how you think.
Same information, radically different outcomes. The paradox is that AI makes reading faster at the exact moment faster stops being the thing we need. We're not short on content. We're short on comprehension, and comprehension has a stubborn minimum time cost that no model can compress for you.
This isn't a case against AI. It's a case against one specific failure mode: using AI as a replacement for reading instead of a multiplier on reading. The difference looks small from the outside. Inside your head, it's the difference between learning and the theater of learning.
Cognitive Offloading: What Research Shows
The science on this is older than the current AI wave. In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner published "Google Effects on Memory" in Science. They ran four experiments and found a consistent pattern: when people believed information would be saved on a computer, they recalled the information itself less accurately but remembered where to find it better. Memory shifted from content to pointer.
Barr, Pennycook, Stolz, and Fugelsang extended this in 2015 with "The Brain in Your Pocket," showing that heavier smartphone use correlated with lower performance on analytic thinking tasks, especially when problems required overriding an intuitive wrong answer. Storm, Stone, and Benjamin (2017) found similar patterns with external storage and encoding: the act of saving something externally reduced how deeply it was processed internally.
Then generative AI arrived and cranked the effect up.
Lee, Sarkar, and colleagues (2025) at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers in a study titled "The Impact of Generative AI on Critical Thinking." The finding that made it into headlines: higher confidence in AI correlated with lower critical thinking effort. When users trusted the model, they stopped interrogating its output. When they distrusted it, they engaged more. The researchers called this a shift from "task execution to task oversight," but oversight only works if you still have the domain knowledge to spot errors.
Michael Gerlich's 2025 paper in Societies pushed further. Surveying 666 participants, Gerlich found a strong negative correlation between frequent AI tool use and critical thinking scores, mediated by cognitive offloading. Younger participants who relied most heavily on AI scored lowest on independent reasoning measures.
| Study | Year | Finding | What It Means for Readers |
|---|---|---|---|
| Sparrow, Liu, Wegner | 2011 | People remember less when they know info is stored elsewhere | Reading-to-save ≠ reading-to-know |
| Barr et al. | 2015 | Heavier smartphone use tied to weaker analytic thinking | External tools can erode internal skills |
| Storm et al. | 2017 | Saving to external storage reduced encoding depth | The offload happens before you notice |
| Lee, Sarkar et al. (Microsoft/CMU) | 2025 | AI confidence inversely correlated with critical thinking | Trust the model less, think more |
| Gerlich | 2025 | AI reliance negatively correlated with critical thinking, strongest in young users | The pattern compounds with age |
The mechanism is simple and biologically boring. Your brain treats AI the way it treats any other external drive. If you don't engage with the material, you don't encode it. If you don't encode it, you don't remember it. If you don't remember it, you can't think with it. Reading with AI poorly isn't a failure of the AI. It's a failure to do the part the AI can't do for you.
For a longer treatment of the thinking-trap angle, see our piece on the AI thinking trap.
When AI Helps: The Three Green Lights
Nothing above is an argument against AI. It's an argument against one specific workflow. AI genuinely amplifies reading in at least three situations, and all three share a property: they happen after you've already engaged with the text yourself.
Clarification. You're reading a paper on attention mechanisms and you get stuck on a line about softmax normalization. Asking Claude to explain that one concept in the context of what you've already read is genuinely useful. It un-sticks you. It increases your engagement with the rest of the paper instead of replacing it. You're still the one doing the reading.
Synthesis across sources you've actually read. You've read five articles on remote work over the last month. You've highlighted passages in each. Asking an AI to find tensions, contradictions, or unexpected overlaps across those highlights is a job it's good at and a job you probably wouldn't do manually. The key is that you did the reading. The AI is operating on material you've metabolized, not substituting for the metabolism.
Socratic drilling on your own notes. Take your highlights from a book and ask the AI to quiz you on them. Ask it to argue against the author. Ask it what a reasonable critic would say. You're using the model as a cheap, patient sparring partner, which is one of the things it's genuinely best at. This is close to what personal RAG over your notes enables at scale.
Notice the pattern. In all three cases, you bring the reading, the AI brings conversational range. Your understanding grows. The AI isn't pretending to be you.
When AI Hurts: The Three Red Lights
The opposite pattern is everywhere because it's easier, and easier almost always wins in the short term.
Pre-reading summary. You see a long article. Instead of reading it, you paste it into ChatGPT and ask for a summary. You read the summary. You tell yourself you'll go back to the full article later. You don't. Now you have a compressed, decontextualized, possibly hallucinated version of someone else's work, and a false sense of having encountered their thinking. Worse, you've skipped all the texture: the specific examples, the tonal signals, the passages that would have stuck.
Single-prompt digest of long documents. "Summarize this 50-page report." The output is fine. It's also comprehension theater. You'll cite the report in a meeting. Someone will ask a specific question. You'll realize you have no idea, because the summary compressed away exactly the detail that mattered.
Verbatim paraphrasing into your own thinking. You ask the AI for its take. You absorb its framing. You start repeating its phrasing as if it were yours. This isn't plagiarism in the legal sense, but it's a kind of plagiarism of the brain. You end up with views you didn't arrive at, defending them with arguments you didn't construct. When challenged, you can't defend them deeply because you never built the scaffolding.
Here's a decision matrix you can actually use.
| Situation | Should You Ask AI? | Why |
|---|---|---|
| Haven't opened the article yet | No | Read at least a few minutes first. Engagement is the asset. |
| Stuck on a specific concept mid-reading | Yes | Clarification increases engagement with the source. |
| Need the gist of a 50-page report before a meeting | Skim first, then ask | Skim gives you hooks the summary alone can't provide. |
| Comparing 5 articles you've already highlighted | Yes | Synthesis over digested material is a real amplifier. |
| Writing your own take and want a counter-argument | Yes | Socratic use sharpens your position. |
| Want to cite a specific stat or quote | Verify in the source | LLMs fabricate citations often. Never trust the first answer. |
| Need to remember this long-term | Read it yourself | AI summaries are low-adhesion by design. |
Keep the table close. Most bad AI reading habits come from skipping this check and defaulting to "summarize."
The Hallucination Problem Nobody Wants to Talk About
Beyond the cognitive cost, there's a factual one. LLMs confidently fabricate. Citations, dates, quotes, statistics, and sometimes entire papers that don't exist.
Stanford HAI's 2024 research on legal AI tools (Magesh et al., "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools") tested Westlaw AI and Lexis+ AI and found hallucination rates between 17 and 33 percent on domain-specific legal queries, even in products marketed as retrieval-augmented. On general-purpose models, rates on cited claims have been reported between 15 and 40 percent depending on domain, prompt style, and how strictly "hallucination" is defined. A 2023 study on medical question-answering by Pal, Umapathi, and Sankarasubbu ("Med-HALT") found similar magnitudes.
This matters for reading because reading is, among other things, a fact-checking activity. If your encounter with an idea is filtered through an AI that confidently made up the supporting citation, you're not reading. You're consuming lightly contaminated synopsis.
A practical implication: any claim you pull out of an AI summary and plan to use in writing, teaching, or arguing needs verification in the source. This is why the "AI-last" protocol below puts verification in step four rather than trusting the model at step one.
The AI-Last Reading Protocol
Here's the workflow that tends to produce real comprehension rather than the feeling of it. It's four steps, and the order matters more than the content of any single step.
| Step | Action | Time | Why It Matters |
|---|---|---|---|
| 1 | Read first, even just 10 minutes | 10 to 30 min | Primes context, generates hooks for later retrieval |
| 2 | Summarize yourself in 2 to 3 sentences | 2 min | Forces encoding; reveals gaps in understanding |
| 3 | Ask AI specific questions about your highlights | 5 to 15 min | Leverages engagement you've already built |
| 4 | Verify any claim you plan to use | 2 to 10 min | Catches hallucinations before they propagate |
Step 1: Read first. Even ten minutes. Highlight three or four things that strike you: surprising claims, confusing passages, lines you might use later. Glasp's web highlighter is built for this because your highlights persist and can feed step 3.
Step 2: Summarize yourself. Two sentences is fine. This is the encoding step. The friction of writing two sentences is where memory actually forms. Skip this and you skip the point.
Step 3: Ask AI specific questions. Not "summarize this." Things like: "What does the author mean by X in the third section?" "How does this compare to the Sparrow 2011 argument I highlighted last month?" "What's the strongest counter to the claim in my second highlight?" This is where Glasp's AI chat becomes useful, because the AI is working on your highlighted material, not on a generic paste-in.
Step 4: Verify. If you're going to quote a statistic, cite a study, or rely on a date, open the source and confirm. Every time. Yes, every time. For background on why this is worth the friction, see our piece on the AI reading assistant.
The protocol inverts the common pattern. Most people do AI-first reading: prompt, skim, maybe go back. AI-last reading makes the AI the last step, not the first. It takes slightly longer, and you remember enormously more.
Using Glasp's AI Stack Correctly
Glasp is built around the AI-last assumption, though it's worth being explicit about how the tools fit.
Glasp's web highlighter is step 1. You read, you highlight, and the highlights are stored, indexed, and searchable. This is the "engagement first" layer, and without it the later steps don't have anything to work on.
Glasp's AI chat is step 3. It operates over your highlights, not over arbitrary text you pasted in. That design choice is intentional. The chat is strongest when it's synthesizing, questioning, and cross-referencing material you've already metabolized. It's weakest when you ask it to replace the reading, and the product doesn't pretend otherwise.
YouTube Summary is a useful case study in getting this right. You can generate a summary of a video, but the best use isn't "watch the summary instead of the video." It's "watch the video, highlight the moments that matter, then use the summary and transcript to find specific sections again." The summary becomes a retrieval tool, not a replacement for watching.
Glasp's PDF highlighter and Kindle highlights extend the same pattern to long-form reading. The highlights carry forward into your library. Later, you can chat with your notes across every book you've read, which is the kind of synthesis that was genuinely impossible five years ago.
And because Glasp is a social reader, you can also see what other people highlighted in the same piece. That's a different kind of AI-adjacent layer, a human curator signal, and it's surprisingly useful for picking up what a thoughtful reader noticed that you missed.
Case Studies: Two Students, Two Readers
Concrete examples make the abstract concrete.
The two students. Maya and Jun are in the same graduate seminar. Maya uses Claude to summarize every assigned paper before class. She shows up with crisp talking points. Jun reads the papers, highlights what he finds interesting, writes two-sentence summaries in his own words, then asks Claude what he missed and whether his interpretation of the methods section is right. He also shows up with talking points, but he spent roughly twice as long getting there.
End of semester, they take the same comprehensive exam. Maya remembers the AI's framings, not the papers. When the exam asks her to critique a methodology she supposedly read, she struggles because she never actually encountered the specifics. Jun does better, not because he's smarter but because he built scaffolding. Two years later, Jun is still quoting those papers. Maya can't remember the titles.
The two readers. Priya and Sam both read widely for work. Priya has ChatGPT generate book summaries instead of reading books, because she's busy. She "reads" 40 books a year this way. Sam reads 12 books, highlights heavily in Glasp, and uses Glasp's AI chat to find patterns across his highlights every few months.
Priya talks about books in shallow, interchangeable ways, because the summaries she ingested were shallow and interchangeable. Sam can tell you which author said what, how two authors disagree, and where his own thinking has shifted. He's read fewer books and understood more of them. This is what remembering what you read looks like in practice.
The lesson isn't that AI is bad. It's that AI amplifies what you put in. Put in engagement, get amplified comprehension. Put in nothing, get amplified nothing.
What Five Years of AI-Native Reading Could Do to Us
Speculation, but grounded speculation.
Calculators didn't destroy math, but they did shift what people are good at. Most adults today can't do long division in their heads as fluently as their grandparents could. GPS didn't destroy navigation, but spatial memory research (Dahmani and Bohbot, 2020) shows people who rely heavily on turn-by-turn directions build weaker cognitive maps of their own cities.
Reading is more cognitively load-bearing than either arithmetic or navigation. It's how we think with other people's ideas across time. If a generation grows up outsourcing the first encounter with every text to an LLM, the likely outcome isn't illiteracy. It's something stranger: a population that "knows what books said" in a surface sense but can't think with the ideas, because they never internalized them. They can quote the summary. They can't argue with the author. For a broader view of what this trend could do to learning, see AI's impact on learning.
This isn't inevitable. Calculators coexist with math education. GPS coexists with people who still know their neighborhoods. The question is whether AI reading coexists with deep reading, or whether the cheaper option crowds out the slower one by default.
The path forward is mostly individual. It's adopting a protocol that keeps the AI useful without letting it take the parts of reading that are you. That's all this article is trying to argue.
Frequently Asked Questions
Is using ChatGPT to summarize an article bad for learning?
If you read the summary instead of the article, yes. If you read the article, write your own summary, and then use ChatGPT to check your understanding or fill gaps, no. The order matters more than the tool.
What's the best way to use AI for reading research papers?
Read the abstract and introduction yourself. Skim the methods. Read the discussion. Highlight what you don't follow. Then use AI to explain the parts you highlighted, compare the paper to others you've read, and stress-test your interpretation. Never quote a statistic from an AI summary without checking the source, because hallucination rates on citations are high.
Do AI summaries hallucinate important details?
Often. Stanford HAI's 2024 work found 17 to 33 percent hallucination rates even in retrieval-augmented legal AI tools. General-purpose chatbots range from 15 to 40 percent on cited claims depending on the domain. Treat any AI-generated specific (a date, a quote, a statistic, a citation) as unverified until you've checked the source.
Should I use AI to explain concepts I don't understand?
Yes, this is one of the strongest uses. Clarification increases engagement with the source rather than replacing it. Be specific in your prompt. "Explain softmax in the context of the attention mechanism I just highlighted" works better than "explain attention."
Is AI-assisted reading different from reading with AI?
It can be, though the phrases overlap. "Reading with AI" in this article means using AI as a companion to your own reading, in the AI-last pattern described above. "AI-assisted reading" in some marketing material really means AI-first reading where the model produces a digest and you consume the digest. The distinction is whether you're still the one doing the reading.
Can AI chat across my highlights help or hurt retention?
It helps, provided the highlights came from real engagement with the source. If you highlighted mechanically or asked AI to highlight for you, you're synthesizing across shallow material and you'll get shallow synthesis. If you highlighted things that struck you while reading, chatting across those highlights is one of the highest-leverage uses of AI for learning that exists today. Tools like Glasp's AI chat are designed around this loop.
Conclusion
AI can make you a better reader or a faster forgetter. The tool doesn't decide which. You do, through the order in which you use it.
Read first. Highlight what strikes you. Summarize yourself, even briefly. Then bring the AI in for clarification, synthesis, and questioning. Verify anything you plan to use. It's two or three extra minutes per article, and it's the difference between knowing something and performing the knowing.
If you want to try the AI-last workflow in practice, Glasp's web highlighter plus Glasp's AI chat are built for exactly this sequence. Highlight a piece today. Write yourself a two-sentence summary. Then ask the AI what you missed. A week later, check whether you still remember what it was about. That's the only test that matters.