Explanation and understanding

An explanation tries to leave the listener with an accurate mental model of how something works or why something happens. Whether it succeeds is hard to judge from the inside, for the explainer and the listener alike. Listeners overrate how well they understand how things work, explanations can make them feel they understand more than they do, and a polished explanation that contains misconceptions can feel as convincing as a correct one. The main questions are what makes an explanation produce understanding rather than the sense of it, how the explainer should adapt to what the listener already knows, and whether explanations written by AI language models help or mislead.

History

A history of research on explanation (philosophy of explanation, instructional-explanation research, explainable AI) is still to be written.

Core concepts

The illusion of explanatory depth

People think they understand complex phenomena better than they do. Rozenblit and Keil found that people rate their understanding of complex phenomena as higher than it is, and that trying to explain them lowers their rating (Rozenblit & Keil, 2002). The illusion is much stronger for explanatory knowledge than for knowledge of procedures or narratives, and (with some exceptions) than for facts. Warning people beforehand that they would be tested reduced the drop in rating but did not remove the illusion (Rozenblit & Keil, 2002).

The same gap appears for explanations received from others. After a physics explainer video, learners who had low prior knowledge were more prone to the illusion, and demanding follow-up tasks reduced it more than the video alone; the authors advise against using videos in isolation (Hörnlein & Kulgemeyer, 2025). In another experiment with low-prior-knowledge learners, a video that explained force through misconceptions left learners as sure they understood as a correct video did, but with more misconceptions and fewer correct ideas (Kulgemeyer & Wittwer, 2023). Confidence, in other words, is a poor guide to whether the explanation worked.

Explaining can also change what people believe. In one study, explaining how policies would work lowered people’s rated understanding and made their positions less extreme, and the more their rated understanding dropped, the more uncertain they were about their position (Association for Psychological Science, 2013a).

What listeners find convincing

Judged quality and learning can come apart. Novices rated good explanations as more satisfying than bad ones, but adding irrelevant neuroscience information inflated their ratings of the bad explanations (Weisberg et al., 2008). A 2024 replication found the effect modest and people better at telling good explanations from bad ones than the original suggested (Väth, 2024).

Judgments of explanations are also sensitive to mechanism information: in five experiments (1,730 participants), explanatory judgments responded more to mechanism information, and less to covariation, than judgments of causal strength did (Vasil & Lombrozo, 2022). That measures how people judge explanations, not what they learn from them. In one study, detailed causal-mechanism information only slightly reduced skepticism toward interventions on root causes and did not change a preference for fixing symptoms, though this was a decision outcome rather than a test of understanding (Müller, 2026).

Leaving out tangents

Interesting but irrelevant material lowers learning. A meta-analysis of 58 studies with more than 7,500 learners found that such “seductive details” reduced learning scores, and its lead author stresses that this is not an argument against engaging explanations (Zaske, 2020). Mayer’s coherence principle says learners learn more deeply when extraneous material is excluded; his 2005 review reported support in 10 of 11 tests, with a median effect size of 1.32 (Mayer, 2005). Seductive details are defined as interesting material tangential to the learning objective (The Learning Scientists, 2019).

Refutation of misconceptions

A refutation text states a common misconception, refutes it and then explains the correct idea. A meta-analysis of 44 comparisons found a moderate advantage over texts without refutation (g = 0.41, n = 3,869; publication-bias adjustment lowers the estimate); the effect did not vary significantly with age or prior knowledge, and few comparisons favoured the non-refutation text (Schroeder & Kucera, 2022). In an experiment on eighth graders, refutation texts corrected misconceptions held with high confidence (hypercorrection) while low-confidence misconceptions did not benefit (Van Loon et al., 2015). Misconceptions held with high confidence are often not corrected during ordinary study (Van Loon et al., 2015).

Repeating a misconception in order to correct it is not known to backfire: reviewers of misinformation corrections conclude that backfire effects are not robust and that avoiding repetition of the misconception is unnecessary, which is indirect support for refutation texts (Swire-Thompson et al., 2020).

Analogy

In four experiments, science novices answered basic questions best after literal explanations but inferential questions best after analogical ones (Donnelly & McDaniel, 1993). For learners with prior knowledge, analogy helped when the presentation discouraged contact with their prior knowledge and hurt when it encouraged it (Donnelly & McDaniel, 2000). Because an analogy carries attributes that the target does not share, it can leave oversimplified or distorted concepts when the mapping breaks down or the analogy is unfamiliar (Toma & Rahman, 2021); a review makes the same point (Richland & Zhao, 2026). Single, highly concrete analogies carry more risk of misunderstanding, perfect one-to-one mappings are rare, and aligned multiple comparisons are described as particularly effective (Richland & Zhao, 2026). Teachers’ and researchers’ accounts add that analogies help only with supporting cues such as imagery and gestures, and that stating the shared and unshared attributes is crucial (University of California, 2007; Toma & Rahman, 2021).

Adapting to the listener

Experts often fail to consider a layperson’s limited knowledge. In an 80-dyad experiment, giving experts individuating information about the layperson led to specific adjustments that made the advice more effective and efficient; the information did not make the experts simplify in general (Nückles et al., 2005). In experiments on predicting what others know, people’s own knowledge contaminated their estimates, and making them rely less on their own knowledge did not make the estimates more accurate; the authors explain the bias by a lack of diagnostic cues about others’ knowledge (Tullis & Feder, 2022). Both findings point toward getting information about the listener, not trying harder to imagine them. The common ground article covers this problem in conversation.

The expertise reversal effect holds that instruction that helps low-prior-knowledge learners loses its effect or reverses for high-prior-knowledge learners, so instruction should be adjusted as knowledge grows (Wikipedia contributors, 2026d). Information essential to novices may be redundant for more knowledgeable learners (Cobb, 2021), and a blog reports a 2025 meta-analysis of 60 studies in which low-prior-knowledge learners gained from high-structure instruction (d = 0.505) and high-prior-knowledge learners from low-structure instruction (d = -0.428), with a caution against turning this into step-by-step rules (Watson, 2025b).

A related finding concerns text coherence. Two later papers describe an interaction in which low-knowledge readers learn more from high-coherence text (explicit linking words and argument overlap) and high-knowledge readers more from low-coherence text (Renkl et al., n.d.). A replication found it in linear text but not when the same material was read as hypertext (Salmerón et al., 2005).

Self-explanation

Prompting students to explain a text to themselves while reading, compared with rereading, improved understanding in a small experiment (Chi et al., 1994). In a meta-analysis, self-explanation prompts beat instructor-scripted explanations in the same task (g = .35) but beat no extra explanation by more (g = .67), so much of the benefit is available from instructor explanations (Bisra et al., 2018). When an instructional explanation was available on demand next to self-explanation prompts, learners did worse than with prompts alone, because the available explanation reduced the incentive to self-explain; another study found instructional explanations as effective as self-explanations, with better results when learners compared their own explanation with the given one (Bisra et al., 2018).

Questions, controversies, and new developments

Do simpler explanations work better?

Lombrozo’s experiments found that people prefer simpler explanations (those with fewer causes) and treat simplicity as raising an explanation’s prior probability, so a simpler explanation is favoured even when less likely (Lombrozo, 2007). Later work says the literature is divided: the presence of causal mechanisms reduces adherence to simplicity, and complex multi-cause explanations can give a greater sense of understanding (Zemla et al., 2023). In a study of explainable agents, satisfaction depended on the consistency of relationships among the causes, and the authors say the question is not simple versus complex (Blom et al., 2024). These findings concern preferences and satisfaction, not learning.

Do refutation texts reliably help?

The meta-analytic advantage is moderate, but the evidence is mostly short one-time passages, and only two studies tested beyond a month (Schroeder & Kucera, 2022). One later experiment found that refutation texts changed reading processes (bridging inferences and elaborations) but not post-reading learning, and the authors call earlier findings mixed (Hunsu et al., 2023).

Are conversation and interaction better than a monologue?

The ICAP framework hypothesises that learning rises from passive to active to constructive to interactive engagement (Chi & Wylie, 2014). In a Wizard-of-Oz study with a human expert, free-form conversation after static explanations of an image classifier improved self-reported comprehension, which was self-rated and not tested (Zhang et al., 2023). Overhearing a tutee ask deep questions beat a monologue only on the amount and relevance of written recall; no other difference was significant (Driscoll et al., 2003).

Can understanding be checked cheaply?

Teach-back (asking the listener to explain it back) is common in patient education, but in a systematic review all 26 included studies combined it with other interventions, so its effect cannot be isolated; only 5 were randomized trials, and the evidence that it improves knowledge retention is limited and inconsistent (Yen & Leasure, 2019). Self-explanation prompting has stronger support, described above.

Do AI-generated explanations help or mislead?

The evidence is thin and partly indirect. In a pre-registered experiment, explanations from a large language model increased reliance on both correct and incorrect answers; reliance on incorrect answers fell when sources were given or the explanations contained inconsistencies (Kim et al., 2025b). Across seven experiments (n = 10,462), learning from language-model summaries produced shallower knowledge than web search with the same facts (Melumad & Yun, 2025). In a preprint, human raters were deceived by incorrect GPT-4 Turbo science answers (Bajpai et al., 2024), and analogies generated by a model helped high school students’ understanding but needed teacher guidance to prevent over-reliance (Shao et al., 2025). In a small three-group study (34 per group), the group that consulted ChatGPT overestimated its explanatory ability most, more than a group that read identical texts without the chat; coder-rated explanation quality was lower in the ChatGPT group (M 2.47) than in the read-the-same-texts group (M 2.75), against 1.8 for a no-material control (Aslanov et al., 2026). This is a single conference paper.

Asking what the listener knows

A 2026 preprint proposes that a conversational agent can find out the user’s expertise by asking targeted questions, and so tailor responses for comprehension (Cao & Huang, 2026). This is a proposal; it is not shown to raise understanding.

Broader connections

Explanations are one form of the instruction studied in transfer of learning, where the question is whether understanding carries over to new uses, and in AI tutoring, where chat-based explanation is compared with structured lessons. The sentence-level craft of making an explanation readable is the subject of writing for comprehension. The belief that a speaker or an AI system shares one’s knowledge, and the failure of that belief, are treated in common ground. How people draw conclusions from the causes an explanation gives is part of reasoning.

Further reading

• Rozenblit and Keil Rozenblit and Keil (2002): the founding studies of the illusion of explanatory depth.
• Schroeder and Kucera Schroeder and Kucera (2022): the meta-analysis of refutation texts, with moderators and publication-bias adjustment.
• Bisra and colleagues Bisra et al. (2018): a meta-analysis of self-explanation prompts, including comparisons with instructor explanations.
• Kulgemeyer and Wittwer Kulgemeyer and Wittwer (2023): an open-access experiment on how misconception-laden explainer videos produce confidence without understanding.
• Melumad and Yun Melumad and Yun (2025): seven experiments comparing learning from language-model summaries with web search.