All articles
6 min readlearning / training

0.55 SD from one prompt: when self-explanation works

After a demo in a training, I used to ask the room whether they had questions. Nods, no hands. The prompt that changes the next 10 minutes is narrower: everyone writes, in two sentences, why the step I showed works, before anyone speaks. Two meta-analyses put that habit near half a standard deviation. A preregistered trial published in April found no effect of it on factual material.

Two meta-analyses, around half a standard deviation.

Kiran Bisra, Qing Liu, John Nesbit, Farimah Salimi and Philip Winne pooled 69 effect sizes from 64 research reports in which the only difference between conditions was a prompt to explain the material to yourself. The random-effects mean was g = 0.55, holding across 20 coded moderators including task type, subject area, education level and treatment duration 1.

Li-Ping Tan and colleagues ran the same exercise in 2025 for digital learning environments: 204 effect sizes from 56 studies, overall g = 0.46. Split by outcome, retention came in at g = 0.31 and transfer at g = 0.33. Two of their moderators matter to anyone building a session. Effects were larger when the learner controlled the pace rather than the system, and larger for conceptual and mixed knowledge than for procedural knowledge 2.

The original demonstration dates to 1994. Michelene Chi, Nicholas de Leeuw, Mei-Hung Chiu and Christian LaVancher asked 14 eighth-graders to explain each line of a passage on the human circulatory system aloud, with no training in how to do it, while 10 others read the same passage twice. The prompted group gained more from pretest to posttest. The students who produced the most explanations all finished with a correct mental model of the circulatory system, which many of the re-readers and the sparse explainers never reached 3.

Plain facts got nothing.

Benjamin Harders and Mirjam Ebersbach preregistered a test of the same manipulation on factual material. 208 participants learned fictional facts in the lab, either prompted to self-explain or re-reading, then sat an immediate test and another one 2 weeks later. The retention interval had a large effect on scores. The learning strategy had none, at either test, with no interaction between the two 4.

The authors then looked inside the explanations people had produced. Counting only the correct ones, the number of inferences predicted test performance. The number of paraphrases did not 4. A learner who restates my sentence in new words gains nothing. The gain shows up when they derive something the text never stated.

That lines up with Tan's moderator on conceptual versus procedural content 2. A price list, a keyboard shortcut or a client name holds no mechanism, so there is nothing to infer from it, and you pay the minutes for nothing.

The minutes come out of the exercises.

Kelley McEldoon, Kelly Durkin and Bethany Rittle-Johnson put 69 children in grades 2 to 4 through mathematical equivalence problems like 3 + 4 + 8 = _ + 8, in three conditions: self-explanation prompts, additional practice problems, or a control. Against the control, the prompts raised both conceptual and procedural knowledge. Against extra practice, the benefit was more modest and appeared on some subscales only. The authors closed by asking for more attention to how much self-explanation gains over alternative uses of the same time 5.

Bethany Rittle-Johnson, Abbey Loehr and Kelly Durkin pooled the mathematics literature in 2017 and found small to moderate gains on immediate procedural knowledge, conceptual knowledge and procedural transfer, with much thinner evidence for classroom effects or long-term retention. The immediate effect was larger when the design scaffolded a high-quality explanation, and it did not vary with whether time on task was held constant across conditions 6.

A 2026 AIED paper shows the same trade in a modern setup. Eason Chen and colleagues gave 92 learners a fixed 60 minutes on limits and derivatives in three conditions: no self-explanation, menu-based prompts, or open-ended prompts with feedback from a language model. Post-test scores did not separate the three. On transfer items whose correct answer was "not enough information", the open-ended group wrote better explanations, around 11.9 percentage points above control (p = .030), while their multiple-choice accuracy on those same items held flat (p = .183). They reached that having worked far fewer practice problems in the hour 7. The explanation quality was scored by a model and the practice volume differed between groups, so I read this one as a cost estimate rather than a result.

The protocol.

1. Prompt where a mechanism exists. Conceptual and mixed knowledge carried larger effects than procedural knowledge in 204 digital effect sizes 2, and factual material carried none against re-reading 4. Before writing a prompt, check that the content has a because in it. Configuration steps, shortcut keys and vocabulary go to retrieval practice instead.

2. Ask for an inference, not a paraphrase. Inferences predicted scores in Harders and Ebersbach's data; paraphrases predicted nothing 4. My test for a draft prompt: if a learner can answer it by restating my slide, rewrite it. "Why does this work on a 200-page contract and fail on a scanned invoice?" survives the test. "Explain what the step does" does not.

3. Scaffold the shape of the answer. The 2017 meta-analysis found larger immediate effects when the design supported a high-quality explanation 6. I give a two-part frame: "This step does X because Y. Remove it and Z happens." Two sentences, written, 90 seconds.

4. Take the time out of the exercise block, deliberately. Prompts beat a control and barely beat extra practice 5, and the learners who explained in Chen's study got through many fewer problems in the same hour 7. One prompt per 3 exercises, with the fourth exercise cut. A prompt stacked on top of a full practice set runs the session over and reads as homework.

5. Prompt the wrong answer too. Among the design principles from the 2017 meta-analysis: have learners explain why correct information is correct, and why a common misconception is wrong 6. In my AI trainings that second one carries most of the value, since the failure modes are where people reach for the tool wrongly.

6. Let them pause. Learner-controlled pacing beat system-controlled pacing in the 2025 pooling 2. In practice that means stopping the recording or the demo while people write, rather than talking over the 90 seconds.

7. Check transfer at a delay. Pooled transfer sat at g = 0.33 on immediate measures, with the long-term and classroom evidence thin 26. Three weeks after the session, send 2 scenarios the training never covered and read the written answers.

The slide after my next demo is a blank box with one line above it: "Remove step 3. What breaks?" Ninety seconds, written, before anyone talks. Then three exercises instead of four.

Sources.

  1. Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., & Winne, P. H. (2018). Inducing Self-Explanation: a Meta-Analysis. Educational Psychology Review, 30(3), 703-725. doi.org/10.1007/s10648-018-9434-x
  2. Tan, L.-P., Gong, S.-Y., Wang, Y.-J., Guo, X.-R., Xu, X.-Z., & Wang, Y.-Q. (2025). Enhancing Academic Performance Through Self-Explanation in Digital Learning Environments (DLEs): A Three-Level Meta-Analysis. Educational Psychology Review, 37, 20. doi.org/10.1007/s10648-025-10001-x
  3. Chi, M. T. H., de Leeuw, N., Chiu, M.-H., & LaVancher, C. (1994). Eliciting self-explanations improves understanding. Cognitive Science, 18(3), 439-477. doi.org/10.1207/s15516709cog1803_3
  4. Harders, B., & Ebersbach, M. (2026). No Causal Self-Explanation Effect for Factual Knowledge. Applied Cognitive Psychology, 40(3), e70174. doi.org/10.1002/acp.70174
  5. McEldoon, K. L., Durkin, K. L., & Rittle-Johnson, B. (2013). Is self-explanation worth the time? A comparison to additional practice. British Journal of Educational Psychology, 83(4), 615-632. doi.org/10.1111/j.2044-8279.2012.02083.x
  6. Rittle-Johnson, B., Loehr, A. M., & Durkin, K. (2017). Promoting self-explanation to improve mathematics learning: A meta-analysis and instructional design principles. ZDM Mathematics Education, 49(4), 599-611. doi.org/10.1007/s11858-017-0834-z
  7. Chen, E., Tang, X., Zhao, Y., Chen, M., Elmir, M., McLaughlin, E., Yuan, M., Wang, Y., Agarwal, S., Cochrane, J., Lin, J., Wu, T., & Koedinger, K. (2026). Practice Less, Explain More: LLM-Supported Self-Explanation Improves Explanation Quality on Transfer Problems in Calculus. AIED 2026. arxiv.org/abs/2604.00142

Keep reading