Why does highlighting and re-reading feel like learning but vanish later?

Because marking produces a visible record of attention, and attention is not recall. Readers who choose their own highlights remember more of the marked facts and little more of the meaning, and underlining can leave people worse on questions that need connections across a text. The feeling is real. It measures the wrong thing.

The page looks like evidence of work. Then someone asks what the section argued, and you can produce the topic and not the argument. The marks show where you stopped, not what you built.

What did the experiments actually find?

Plainer than the folklore. In the classic test, students read about 8,000 words, studied for an hour, and returned a week later for a 54-item test, and the highlighting groups did no better than the readers who marked nothing. Within the marked text they did better, more so when they chose the marks themselves.

The experiment is Fowler and Barker's, from 1974, reported in a widely cited review of study techniques, Dunlosky and colleagues' 2013 paper. Undergraduates read Scientific American and Science articles. Three groups: read only, highlight as much as you like, or read a text highlighted by someone else. Each group studied for an hour. A week later they reviewed for ten minutes and took a 54-item multiple-choice test.

The null repeats across decades: Stordahl and Christensen in 1956, Idstein and Jenkins in 1972, Hoon in 1974, on immediate and delayed tests, with Air Force trainees as well as undergraduates.

Nor was the failure total. There was a small cost on unmarked material and a real gain on the marked parts. Ponce, Mayer and Méndez's 2022 meta-analysis pooled 36 articles and 85 effect sizes and found learner-generated highlighting improved memory at 0.36 but not comprehension, which sat at 0.20. Instructor-provided marks did better, at 0.44 for both. Marking is a good way to hold a fact and a poor way to hold an argument.

Why do highlights fail when you need to explain something?

Because a mark points at an item, and explanation lives in the links between items. In Peterson's 1992 study, underliners matched other groups on factual questions and fell behind on inference questions, the ones that ask you to connect two claims. Marking tells you where a fact sits. It does not build the argument.

Peterson had education majors read a 10,000-word history chapter; two groups underlined for 90 minutes, one only read. A week later, after 15 minutes of review, everyone took a 40-item test: 20 on facts, 20 on inferences. The groups were similar on facts, but students who had underlined and reviewed their marked text were worse on the inference questions, a result the review says needs replication.

The review's account is about attention. Underlining draws attention to individual concepts rather than connections across them, so the reader finishes with marked points and no map. Bell and Limber found weaker readers marked more irrelevant material and scored lower on the final exam, and training that pushed readers to mark as little as possible helped.

One caveat. In a five-country comparison of PIRLS 2021 digital reading data, highlighting item-relevant text predicted accuracy in only one country. Observational, partly inconsistent, and a reason to watch your own page.

How much highlighting is too much?

More than roughly a tenth of the page, at least in one test. When 127 people read a short story and could mark nothing, up to 150 words, or as much as they liked, the capped group scored highest on an open-book comprehension test the next day, with no difference in reading time. The limit forced a choice.

Joshi and Vogel's 2024 CHI experiment ran in a document reader they built, with participants reading a story, marking under one of those conditions, and returning 24 hours later for an open-book test of 20 multiple-choice questions in five minutes. The capped group averaged 14.3 against 12.1 for unlimited marking and 10.5 for none, about 11 and 19 percent better, with equal reading time. A pilot behind that cap compared 50, 150 and 250 words, and 150 won.

Treat that as one design's result: one story, one interface, one laboratory. What travels is the mechanism, since a cap turns marking back into a decision about what matters.

What should you do instead of a second pass?

Read the whole section before you mark anything, because the choosing is the processing. Cap the marks at about ten percent, write how two claims connect rather than what they say, and turn two marks into questions you answer out loud without looking. The spoken answer is the only step that tests understanding.

Four moves. Read first, then mark, because marking while you read lets the decision eat the comprehension. Budget the marks at ten percent of a long document, and remove something when you hit the cap. Write relations where you can: an arrow between paragraphs, a margin note that this figure contradicts the earlier one. Then turn two highlights into questions and answer them from memory.

Rereading keeps one job: repair. When a sentence did not land, go back a few lines. That is different from a second pass over the chapter, which mostly re-exposes you to words you already recognize.

The AI version of the mistake is worth naming. When a model hands you the highlighted version, you get the marks without the choosing, and the choosing is where the benefit was measured. Before you read its list, write down the two sentences you would have chosen.

If you want a daily setting where the check happens, Absorb's comprehension workout on the App Store gives you reading, listening and analysis rounds with three scores in ten minutes.