Most of us study by looking. We reread the chapter, run a highlighter over the important sentences, and skim our notes one more time before the meeting or the exam. It feels productive. The page gets more familiar each time, and familiarity feels a lot like knowing.

A large body of research says that feeling is misleading. One of the most reliable ways to make something stay in your head is to close the book and try to pull it back out. Cognitive psychologists call this retrieval practice. The result it produces is known as the testing effect. It is one of the simplest thinking tools there is: no software, no method to learn, only a blank page and the willingness to find out what you can't remember.

The experiment that put it back on the map

The idea is old. Educational psychologists were comparing reading with “recitation” more than a century ago. Arthur Gates published a classic study in 1917, building on a German study by Kühn from 1914. Interest came and went. In 2006, the psychologists Henry Roediger and Jeffrey Karpicke published a study that brought it back to the centre of learning research.

Students read prose passages. Some then reread the passage. Others took one or three free-recall tests, recalling as much as they could, without any feedback. Everyone took a final test five minutes, two days or one week later.

The results split in two directions. On the five-minute test, rereading won. On the delayed tests, two days and a week later, the students who had practised recall retained substantially more. And the students who had reread were more confident that they would remember the material. The authors' conclusion fits in one line: “Testing is a powerful means of improving learning, not just assessing it.”

That gap between how learning feels and how well it lasts is the heart of the story. Rereading pays off right away and then fades. Recall feels harder and less fluent, and it lasts longer.

We don't notice it working

Two years later, Karpicke and Roediger published a study in Science on learning foreign-language vocabulary. Once students had produced a word correctly, the researchers varied what happened next: the word could stay in further study but leave testing, stay in further testing but leave study, or be dropped from both. Extra studying after that point had no effect on delayed recall. Extra testing had a large positive effect. And the students' predictions of how well they would do were uncorrelated with how well they actually did. In the authors' words, the results suggest that “even university students seem unaware of this fact.”

Students' habits match. In a 2009 survey of 177 college students, Karpicke, Butler and Roediger found that most of them repeatedly reread their notes or textbook. Relatively few tested themselves while studying. The authors put this down to “illusions of competence”: the fluency of rereading is easy to mistake for mastery.

Not just for flashcards

A fair objection is that all this sounds like memorising. Vocabulary and lists of facts are one thing; understanding a scientific idea is another. In 2011 Karpicke and Janell Blunt tested this in Science as well. They compared retrieval practice with elaborative studying through concept mapping, using texts of the kind found in science education. Retrieval practice produced greater gains, including on questions that tested comprehension and required inferences. It even won when the final test asked students to draw a concept map.

That finding was not left unchallenged. A group of researchers led by Joel Mintzes published a formal comment questioning the study “on methodological and epistemological grounds,” and the authors replied. That kind of exchange is worth knowing about. It is how science is supposed to work, and it shows that a single striking study is not the last word.

What the reviews say

A single study, however elegant, is a weak foundation. The case for retrieval practice rests on reviews of a very large literature.

  • In 2013, John Dunlosky and colleagues evaluated ten popular learning techniques for Psychological Science in the Public Interest. Practice testing and distributed (spaced) practice were the only two to receive a high utility rating. Rereading and highlighting were rated low utility, and the authors suggested practice testing in place of rereading.
  • Christopher Rowland's 2014 meta-analysis in Psychological Bulletin compared testing with restudy. Its findings support the role of effortful processing: initial recall tests yielded larger benefits than recognition tests, such as multiple choice.
  • Olusola Adesope and colleagues' 2017 meta-analysis in Review of Educational Research found practice tests more beneficial for learning than restudying and all the other comparison conditions they examined.
  • Moving from the lab to real classrooms, Chunliang Yang and colleagues (2021) pooled data from 48,478 students in 222 independent studies. Quizzing raised achievement “to a medium extent” (g = 0.499). The size of the effect depended on factors such as feedback, the number of repetitions, and how closely the quiz matched the final test.
  • A systematic review by Pooja Agarwal, Ludmila Nunes and Janell Blunt (2021) coded 50 classroom experiments. The majority (57%) of the effect sizes showed medium or large benefits. The authors also flagged a blind spot: only 6% of the experiments were conducted outside so-called WEIRD (Western, educated, industrialised, rich, democratic) countries.

Where it may not work as well

An honest account also needs the boundaries. In 2015, Tamara van Gog and John Sweller edited a special issue arguing that the testing effect decreases, or may even disappear, as learning materials become more complex. Their measure of complexity was “element interactivity,” meaning many pieces of information that must be understood together. Several experiments in that issue, for example on learning to solve problems from worked examples, found no testing effect. Karpicke and William Aue replied in the same journal that the testing effect is “alive and well” with complex materials. They criticised how element interactivity was defined and pointed to studies the review had left out.

That debate has not been settled here, and we won't pretend otherwise. Two practical points come out of the reviews themselves. First, the benefit tends to show up after a delay, not straight away. Second, conditions matter: the review authors highlight feedback, the format of the practice, and how often it is repeated.

Failing to remember is part of it

One more finding makes retrieval practice easier to live with. You might worry that trying to recall something and failing would only reinforce your confusion. Nate Kornell, Matthew Hays and Robert Bjork tested exactly this in 2009. They used questions people could not possibly answer, such as fictional trivia or guessing a weakly related word, and then showed the correct answer. Unsuccessful attempts still enhanced learning compared with simply reading the question and answer together. The authors suggest that taking challenging tests, instead of avoiding errors, may be one key to effective learning.

Using it as a reader and a writer

Retrieval practice turns reading into something closer to writing. You stop receiving the text and start reconstructing it. A few ways to use it, all consistent with the studies above:

  • Close the book, then write. After a chapter or an article, write down what you remember on a blank page before you look back. Then check it against the source and fill the gaps. That check is your feedback.
  • Prefer recall to recognition. “Explain this in your own words” asks more of you than “pick the right answer,” and in Rowland's meta-analysis recall tests gave larger benefits.
  • Come back later. In these studies the payoff shows up after days, not minutes. Spreading recall sessions out over time also fits the high rating Dunlosky and colleagues gave to distributed practice.
  • Distrust fluency. If the material feels easy because you've just seen it, that is exactly the situation in which the studies above found people misjudging what they would remember.

The same caution applies to AI tools, including the ones we build. A model that summarises an article for you can become another version of rereading: smooth, familiar and quickly forgotten. A more useful request is often the reverse. Ask it for questions about the text, answer them from memory, and only then compare. Tabulect, Polyquill and Feedcairn all let you edit their prompts or write your own, so you can point them at that kind of work. The research, though, does not depend on any app. A blank sheet of paper is enough.

Reading puts ideas in front of you. Recall shows you which ones you have actually taken in.