The Misreading

On reading that appears to happen, an AI practice that helps, and what is at stake when neither is taken seriously.

"Knowledge isn't free. You have to pay attention." Often attributed, uncritically and incorrectly, to Richard Feynman.

The line is sharp, almost certainly true to the spirit of the man, and I could not locate it in any of his books, his published lectures, his interviews, his letters, or in the compilation his daughter Michelle Feynman drew from his spoken and written legacy. It appears on Goodreads, where the community adds quotes that Goodreads explicitly does not verify; on a fan-run Twitter account; on the image-quote sites that ricochet across social media in any given week. The line may, of course, have been said in a Caltech corridor and never recorded. The public evidence is closer to a meme that congealed into an attribution.

That is the problem in miniature. Everyone who has shared the quote has been reading without reading: trusting the picture, the format, the implicit promise that someone, somewhere, must have checked. The same gesture that authenticates a quote in our minds is the gesture that distinguishes a useful relationship with artificial intelligence from a destructive one.

The gap between the appearance of reading and the act

A piece of writing reaches a reader. The eyes cross the text. A reply, a share, a comment is produced. Inside the reader's head, an unmistakable sense of having read settles in. From inside, the sense is indistinguishable from the act. From outside, in the response that follows, the difference is sometimes visible to everyone except her.

This is not a complaint about social media or about kids today. The gap between the appearance of reading and the act of reading is older than the internet, and the empirical record on it is now a decade deep. In 2016, researchers at Columbia and the French National Institute for Research in Digital Science and Technology presented a study at the ACM Sigmetrics conference. Across roughly ten million clicks on news stories from major outlets, fifty-nine percent of the links shared on Twitter had never been opened. The shares preceded the reading. In a majority of cases, they preceded any reading at all.

Six years later, a team at the McCombs School of Business in Texas published a paper in the Journal of Consumer Psychology titled I share, therefore I know? The authors found that sharing an article online raises the sharer's subjective sense of knowing its contents, even when she has only glanced at the headline. Putting information online commits the sharer to a public expert identity, and her self-image adjusts to match.

The cybersecurity profession runs on text, and the volume crossing a senior practitioner's desk in a week is larger than any working attention budget. The triage that follows is rational. It is also where the impression-without-the-act takes hold. A vendor white paper, encountered through a peer's enthusiastic share and never opened, leaves an impression that feels as solid as one earned by reading; the gap shows up later, in the renewal cycle or the auditor's question. A board summary written for a careful reader, encountered by a hurried one, will be remembered as a different document. A NICE Framework update read at headline speed will be remembered as confirmation of whatever the reader already believed about the workforce shortage. A candidate's evidence package skimmed for tool names will be remembered as the resume the manager went in expecting to find. In each case, the original lost.

The defensive turn

The trouble begins when the gap gets exposed.

Suppose another reader, or the writer, points out that the article did not say what the first-pass reader has just claimed it said. She remembers reading. The impression is strong, recent, and confidently held. It was produced not by the text but by the headline and her prior beliefs, but she cannot tell that from where she sits. To admit the gap would mean admitting that her recent confident behavior (the share, the reply, the board comment) rested on something she had not earned.

That admission is hard. What gets produced in its place, almost reliably, is a strawman. The strawman is the article she thought she read, restated as if it were the article the writer actually wrote, and then attacked on grounds the writer never gave her. You're saying everyone has to retrain. The writer was not saying that. You're saying compliance work is worthless. The writer was not saying that either. You're saying nobody in my role has options. Also not what the writer said.

This is not malice. It is ego defense. Revising a confidently held belief in public, in front of an audience that has watched one assert the belief, is genuinely expensive. The cheaper move is to attack a position the writer did not hold, because attacking the actual position would require having read it.

The cost compounds. A piece of careful writing is an investment: sentences weighed, objections considered, claims sourced. If a meaningful fraction of the responses are not to the piece the writer wrote but to a piece the network constructed in the writer's name, the strawmanned version now circulates with the writer's byline attached. The next reader encounters the strawman first and assumes it is what the writer said. The actual argument gets crowded out by a poorer one bearing its author's name. This is why so much careful writing in this profession reads defensively, hedged, over-qualified. The writers who survive have learned to write against the strawman rather than for the argument. The strawman wins twice: once on first reading, once again in the next piece the writer is forced to compose around it.

Substitution and scaffolding

A response is now expected. Most public commentary about reading in 2026 ends up at the same destination: artificial intelligence is the cause of the decline, and the only honest answer is to use less of it. The empirical case is not nothing.

In April 2025, a team at Microsoft Research presented a study at the CHI conference of 319 knowledge workers and 936 first-hand examples of AI use in real work. Higher confidence in the AI tool was associated with less critical thinking. Higher confidence in one's own expertise was associated with more. The same tool, in different hands, appeared alongside opposite cognitive outcomes. Two months later, a group at the MIT Media Lab released a preprint titled Your Brain on ChatGPT. Fifty-four participants wrote essays under three conditions: with ChatGPT, with a search engine, with nothing. Their brain activity was monitored across thirty-two regions. The ChatGPT group showed the weakest neural connectivity. By the final sessions, many had stopped writing in any meaningful sense and were copying and pasting from the model's outputs.

Both studies describe something real. Neither describes the only thing that happens when a person sits with a chatbot and a hard text. Each describes one mode of use. Call it substitution. Substitution is what happens when the AI is asked to do the thing the user was supposed to do. The student who hands over the essay. The analyst who hands over the threat assessment. The board reader who reads the executive summary and remembers it as having read the white paper. The cognitive cost of the original task is eliminated, and the cognitive benefit goes with it.

There is a second mode. Call it scaffolding. Scaffolding is what happens when the AI is asked to make a difficult task more accessible without doing the task. The user does the task afterward, on her own, with the scaffolding removed. A GRC lead who asks a chatbot to write her response to a regulator is in substitution mode. A GRC lead who asks a chatbot to render the EU AI Act's risk-tier definitions into plainer language, reads the rendering, and then opens the regulation itself with the rendering held in mind is in scaffolding mode. The first has skipped the work. The second has done the work with help. From outside the user's head, the two activities look superficially similar. They are not the same activity.

The practice that follows is unglamorous. Take a difficult primary source: a dense academic paper, a regulation, a contract clause, a postmortem written in a vocabulary you do not work in daily. Read it once, cold. Note where understanding flickered. Hand those passages to a chatbot and ask for a simpler version. Read it. Then return to the original with the simpler version's scaffolding in the back of the mind. This is the move the practice depends on. Note what the original was doing that the simpler version was not. The writer foregrounded what she foregrounded for reasons; those reasons survive translation only partially. The point is to recover the original, not to replace it.

The mechanism is simple. Difficult texts spend the reader's working memory on parsing: syntax, vocabulary, structure. Little is left for synthesis. A simpler version lowers that first cost. The second pass through the original can then attend to what the writer was doing: what she foregrounded, what she withheld, what the easy version flattened. This is what good teachers have always done: paraphrase, gloss, reframe, then send the student back to the page. The chatbot is not the teacher. It is the gloss at three in the morning.

The practice has limits. The simpler version is sometimes wrong. It loses the writer's tone, sometimes the irony, occasionally introduces an error the original did not contain. Returning to the original is the check; the original is what the practice is for. A user who stopped at the simpler version would be back in substitution mode. The discipline that distinguishes scaffolding from substitution is the discipline of going back.

Why the two halves are one argument

The misreading and the misuse of AI are not separate failures. They are the same failure expressed in two media.

The misreading begins with a confident impression formed without consulting the text. The strawman defends the impression when the text turns out not to support it. The whole loop runs on the reader's inability to tell, from inside her own head, whether reading happened. Substitution-mode AI use begins the same way: a confident impression produced by the model is held as if it had been formed by contact with the original. The user does not return to the source. The impression becomes the knowledge.

The Microsoft finding points toward the distinction, but the practical test is simpler: does the reader return to the source, or stop at the impression? The reader who trusts her impression of an article more than the article itself defends the impression. The reader who treats her impression as a hypothesis tests it against the page. The scaffolding-mode AI user does exactly this, by design, every time she goes back to the original.

The chatbot does not read for you. It readies the page for you to read. The page is still the point.

Coda

The complaint that artificial intelligence is damaging reading and critical thinking is half right. Substitution produces the deficits the studies measure. The conclusion that AI use as such is the problem does not follow from the evidence; the same tool, used by a reader who trusts her own judgment more than the model's, appears alongside more critical thinking, not less.

The path out is the small discipline of insisting on the original after the easier version has done its work. Most readers will not. That is, in the end, what makes the practice valuable. The reader who can tell the difference between her impression of a text and the text, and who is willing, even eager, to test the first against the second, is doing something which in the present economy of attention is not common, and is therefore worth a great deal.

The Feynman quote at the top of this piece may or may not be his. The chatbot can help you find out. But only if you go back to the source. Knowledge isn't free. The cost is attention.

Source notes

Gabielkov, M., Ramachandran, A., Chaintreau, A., and Legout, A., Social Clicks: What and Who Gets Read on Twitter? ACM Sigmetrics / IFIP Performance 2016 conference, Antibes Juan-les-Pins, France, June 2016. HAL: hal-01281190.

Ward, A. F., Zheng, J. (F.), and Broniarczyk, S. M., "I share, therefore I know? Sharing online content — even without reading it — inflates subjective knowledge," Journal of Consumer Psychology, vol. 33, no. 3 (2023), pp. 469–488. McCombs School of Business, University of Texas at Austin.

Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., and Wilson, N., 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 the 2025 CHI Conference on Human Factors in Computing Systems (April 2025). Microsoft Research.

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., and Maes, P., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, arXiv preprint arXiv:2506.08872 (June 2025). MIT Media Lab.

Feynman, R. P.: quote "Knowledge isn't free. You have to pay attention." Widely attributed online (Goodreads community submissions, a fan-run Twitter account, image-quote sites). No primary source located in Feynman's books, published lectures, interviews, letters, or in the compilation prepared by Michelle Feynman (Princeton University Press, 2015).