한국어

The Pretesting Effect: Why Guessing Before You Study Works

At a glance
  • Answering questions before you study, even wrongly, makes the answers stick better than spending that time reading.
  • Two minutes of questions and eight of reading beat ten minutes of reading on the asked-about material a week later (55% vs 45%).
  • Much of the gain came from trying to answer. Students who only memorized the questions learned less than those who guessed.
  • It helps what you were asked about, and almost nothing else (g = 0.54 vs g = 0.04 in a meta-analysis).
  • Before you ask AI, write a ten-second guess. A wrong guess still helps you remember the answer.

Answer questions about a topic before you study it, even if you get nearly all of them wrong. In one experiment, students who spent 2 of 10 minutes on questions and 8 on reading remembered more of what the questions asked about a week later than students who read for all 10 minutes. The gain came from trying to produce an answer, not just from seeing the question.

What is the pretesting effect?

The pretesting effect is the finding that trying to answer questions before you learn the material improves your memory for the answers, even when your first attempts are wrong. Researchers also call it the prequestion effect, and the questions prequestions.

It sounds backwards. Tests are usually something you take after studying, to see what you know. Pan and Carpenter (2023), in a review of this research, put it plainly: testing people on what they do not know looks like a fruitless exercise. Yet across text passages, videos, lectures and trivia, the attempt to answer prepares you to learn the answer when it comes.

The usual advice for previewing a chapter is to skim the headings and read ahead. That is not harmful, but in the Richland experiments below, reading for longer was the comparison condition. The same minutes spent partly on questions did better.

Does it work if you get the answers wrong?

Yes. Richland, Kornell and Kao (2009) gave college students an essay about a patient who lost color vision after a brain injury. One group had 2 minutes to answer five questions about the essay, then 8 minutes to read it. The other group read for the full 10 minutes. To rule out a simple attention effect, a key word from each tested sentence was printed in bold for both groups, so everyone could see what mattered.

The pretested group answered only 10% of the pretest questions correctly. The researchers then removed even those, and compared final scores only on the questions people had gotten wrong.

What students were asked about first stuck better a week later

  • Highlighted content
  • Other content
0%20%40%60%80%100%Highlighted content · Read for 10 min · 45%45%Other content · Read for 10 min · 39%39%Read for 10 minHighlighted content · Questions 2 min + read 8 min · 55%55%Other content · Questions 2 min + read 8 min · 42%42%Questions 2 min+ read 8 minAsked-about content:55% vs 45%
Percent correct on a test one week later. For the question-first group, items answered correctly on the pretest were removed. "Highlighted" means printed in bold. 158 undergraduates (Experiment 4).Source: Richland, Kornell, & Kao (2009), Experiment 4 results
See the numbers
Highlighted contentOther content
Read for 10 min45%39%
Questions 2 min + read 8 min55%42%

Content that was not asked about came out about the same in both groups (42% and 39%, not a significant difference). Answering questions did not make people learn less of the rest, but it did not make them learn more of it either. In Experiment 1, tested right after reading, the gap was larger: 75% for pretested items against 56% for the reading-only group.

TestQuestions firstReading only
Exp. 1, right after reading75%56%
Exp. 4, one week later55%45%

The questions were almost all answered wrong, and the content they asked about still lasted longer.

Kornell, Hays and Bjork (2009) found the same direction with materials where correct answers were impossible or very unlikely, such as invented trivia questions and weakly related word pairs. Guessing first and then seeing the answer led to better learning than seeing the question and answer together.

Is it enough to just read the questions first?

No. Memorizing the questions helped, but answering them helped more. Experiment 5 of Richland and colleagues split this apart. One group answered the questions. A second group was told they would quiz someone else and should memorize the questions without answering. A third read the essay for longer.

Answering the questions beat memorizing them

0%20%40%60%80%100%Percent correct · Read longer · 63%63%Read longerPercent correct · Memorize the questions · 78%78%Memorizethe questionsPercent correct · Answer the questions · 90%90%Answerthe questionsNot enough alone
Percent correct on the highlighted content in a test right after reading. For the answer group, items answered correctly on the pretest were removed. 76 undergraduates (Experiment 5).Source: Richland, Kornell, & Kao (2009), Experiment 5 results
See the numbers
Percent correct
Read longer63%
Memorize the questions78%
Answer the questions90%

Memorizing the questions lifted scores from 63% to 78%, even though the key sentences were italicized for every group. The authors suggested some of that gain may have come from people quietly trying to answer despite being told not to. Trying to answer lifted them to 90%. The authors concluded that part of the benefit comes from the search itself, the work of reaching for an answer you do not have.

How big is the effect, and when does it shrink?

Moderate for what you were asked, and close to zero for everything else. St. Hilaire, Chan and Ahn (2024) pooled the prequestion studies in a preregistered meta-analysis. The effect on prequestioned material was g = 0.54 across 97 comparisons. The effect on material that was not asked about was g = 0.04 across 91 comparisons, effectively none.

Guessing gave the biggest effect, and unasked content gained almost nothing

00.250.50.751Effect size · Asked about (all) · 0.540.54Asked about(all)Effect size · Guessed · 0.650.65GuessedEffect size · Did not guess · 0.220.22Did notguessEffect size · Not asked about · 0.040.04Not askedaboutGuessing was thestrongest moderator
Effect size (Hedges' g) against studying without prequestions. 0.2 is often read as small and 0.5 as medium. "Guessed" and "did not guess" split the effect on asked-about content.Source: St. Hilaire, Chan, & Ahn (2024), results
See the numbers
Effect size
Asked about (all)0.54
Guessed0.65
Did not guess0.22
Not asked about0.04

Of all the moderators they tested, the strongest was whether people were asked to guess: g = 0.65 when they guessed, g = 0.22 when they only saw the questions. Factual questions (0.58) showed a marginally larger effect than conceptual ones (0.28), and adults (0.62) larger than school-age children (0.22), though the authors cautioned against reading too much into those two differences.

When you see the answer also matters, depending on the material. In Pan and Carpenter's (2023) summary, in word-pair studies the benefit appeared only when the correct answer came right after the guess. With text passages the answer came minutes later in the reading, and with trivia the benefit held even when the answer came a full day later. The same review lists real educational settings where pretesting helped, from elementary school students learning about space exploration to medical students learning surgical procedures.

Why don't more students study this way?

Because being wrong feels bad, and the benefit is hard to feel. In a survey summarized by Pan and Carpenter (2023), 78% of respondents rated reading as more effective than guessing first. In another study they describe, pretested word pairs were remembered best, yet participants still ranked pretesting as the least effective method they had used.

This illusion gets stronger with AI. When something is unclear you ask straight away, and a clear answer feels like understanding. The few seconds of producing your own answer first are what drop out. How hint-only AI tutors differ from answer-giving ones is covered in Is ChatGPT Study Mode Good?.

A version you can try today

Ten minutes per lecture or reading is enough.

  1. Turn the chapter's headings, bold terms or end-of-chapter questions into three to five questions. Headings become questions easily ("The light reactions" becomes "What do the light reactions produce?").
  2. Give yourself 30 seconds per question and write an answer. If you don't know, write your best guess. The point is not to leave it blank.
  3. Then read or attend the lecture, and correct each answer as you find it.
  4. A week later, cover the answers and try the same questions again.

The same works with AI. Before you hit send on a question, write one line of "I think it's probably...". It takes ten seconds, and in the studies above even wrong guesses helped the correct answer stick. Why questions help memory in general is in Does Asking Questions Help You Learn?.

In Aloom

Aloom starts from your questions. You ask AI, keep only the answers worth keeping, and each one joins what you already know in a knowledge graph. Then each day you recall the number you chose, 1, 3 or 5, as questions before you see the answers. Reaching for your own answer before the real one opens is the same work this note is about. The more you recall a node, the darker its ink. See how Aloom works

FAQ

What is the difference between pretesting and retrieval practice?

Retrieval practice is recalling something you have already studied. Pretesting is trying to answer before you have studied it. Both involve reaching for an answer, and both help memory. Pretesting prepares you to learn new material, while retrieval practice keeps what you have learned.

Won't guessing wrong make me remember the wrong answer?

In these experiments, guessing wrong and then seeing the correct answer led to better memory for the correct answer, not worse (Richland et al. 2009, Kornell et al. 2009). The key is to always check the answer after you guess.

What kind of questions work best for pretesting?

Fill-in-the-blank, short-answer and multiple-choice questions have all worked (Pan & Carpenter 2023). In the meta-analysis, factual questions showed a slightly larger effect than conceptual ones, though the difference was not quite significant. Since unasked content gains almost nothing, write questions about what you most need to know.

References