Writing for comprehension

Writing for comprehension is the attempt to build rules for text that readers understand and can act on: use plain words, short sentences, active voice, headings, lists, exact numbers, and adapt to the reader. Many of these rules are intuitive and many are repeated in style guides, but the controlled studies behind them are patchy. Some rules hold in experiments (adding explicit links between ideas helps readers who know little about a topic), some do not hold up (active voice, bolding and dyslexia fonts), and many work for one kind of reader and backfire for another. Three disputes run through the field: whether readability formulas measure comprehension at all, whether simplifying text helps the readers it is meant for, and whether a rule that helps novices harms experts.

History

How plain-language movements, readability formulas and text-structure research developed has not been researched. To do: add this section once that history is sourced.

Core concepts

Measuring readability

A readability formula turns surface features of a text, such as word and sentence length or word frequency, into a difficulty score. Two widely used tools, Flesch-Kincaid and Lexile, were checked against students’ cloze scores (the share of deleted words a reader can restore) on 330 passages: Flesch-Kincaid correlated at r = -.76 and Lexile at r = -.70, which the authors call only moderately high (Cunningham et al., 2018). Within any single grade band below university level, both predicted measured text difficulty poorly, and the authors judge that the tools may lack adequate validity for their current uses (Cunningham et al., 2018).

Tests against comprehension are weaker. In a study of 54 adults reading 12 Wikipedia texts, correlations of formula scores with multiple-choice comprehension were weak; the best model (New Dale-Chall plus the reader’s ability) explained about 10 percent of the variance in comprehension and Flesch-Kincaid about 6 percent, and the authors warn that the small text sample limits generalisation (Crossley et al., 2023). In a study of 487 participants reading health texts, formulas were largely unrelated to measured comprehension, while longer texts raised perceived difficulty and verb-heavy content eased comprehension, features the formulas do not model (Ahmed et al., 2025b). Many formulas have only ever been tested on predicting reading scores rather than comprehension, and models using features related to processing and comprehension performed better (Crossley et al., 2023). Newer formulas with linguistic features predict human readability judgments better than Flesch-Kincaid (Crossley et al., 2023).

Language models are now used as raters. Against human readability judgments of plain-language summaries, most metrics, including Flesch-Kincaid, correlated poorly, and the best language model reached r = 0.56 (Cachola et al., 2025). Zero-shot GPT-4 Turbo and GPT-4o mini estimates correlated with human judgments at r = 0.76 and r = 0.74 (Trott & Rivière, 2024). These figures concern human readability ratings, not tested comprehension.

Fluency is not comprehension

Readers are poor judges of their own understanding. A meta-analysis of 94 studies reports pooled r = .24 (95% CI .21 to .26) for how well readers’ relative judgments of their comprehension track tested comprehension (Prinz et al., 2020a), and another of 115 studies finds a weak correlation (Yang et al., 2022). Interventions that make readers use cues about the situation described (a delayed summary, keywords, a diagram, rereading) raised accuracy by g = 0.46 across 28 effect sizes and 2,236 participants; self-explaining did not help (Prinz et al., 2020b). Delayed summary writing and keyword generation also improved how well judgments track tests in the larger meta-analysis (Yang et al., 2022).

The same gap shows in text rewriting. In a large crowdsourced evaluation, language-model-written plain-language summaries were indistinguishable from human-written ones in subjective ratings, yet human-written ones produced significantly better comprehension (Guo et al., 2025). Among the subjective ratings, faithfulness had the largest coefficient for predicting comprehension and simplicity the smallest (0.048 to 0.037, N = 459) (Guo et al., 2025). Easy text can also make readers over-trust: in 43 undergraduates, easy texts made non-experts agree more with claims consistent with their prior beliefs and trust their own judgment, an easiness effect that did not appear for belief-inconsistent claims (Scharrer et al., 2021). A similar pattern holds for jargon: in nine experiments, jargon made scientific explanations more satisfying to laypeople while reducing comprehensibility (Cruz & Lombrozo, 2025).

Cohesion and coherence

Making the links between ideas explicit (cohesion) matters more than shortening sentences or words, in the adolescent experiment that compares them. In adolescents, only reducing cohesion made multiple-choice comprehension worse; harder syntax, vocabulary or decoding did not (Spencer et al., 2018). Syntax manipulation had the smallest effect on oral reading rate (3.43 words) and a 2.21 point decrease in free recall, smaller than cohesion (3.30 points) (Spencer et al., 2018).

The benefit depends on what the reader already knows. Readers with little domain knowledge benefited from a coherent text, and high-knowledge readers from a minimally coherent one, though the benefit for high-knowledge readers depends on whether the test measures text-based or situational understanding (McNamara et al., 1996). This is the reverse cohesion effect (Bohn-Gettler & Kendeou, 2014). It is not always found: in 2020 high school data (n = 511) cohesion manipulations had no overall effect on comprehension while prior knowledge had robust effects (McCarthy et al., 2020). Fourth graders with more knowledge handled low-cohesion narrative texts better, and the authors conclude that adding cohesion cues alone is not likely to be enough for young readers (McNamara et al., 2017). Reading skill itself can partly offset missing knowledge: in 1,651 US high school students it helped offset deficits in science knowledge, and the effect was larger for higher-knowledge students (O'Reilly & McNamara, 2007).

Simplifying vocabulary and structure

Simplification means shorter sentences and common words. Results are mixed.

In a randomised study of 4,563 participants, language-model-simplified text raised multiple-choice comprehension by 3.9 points absolute (14.6 for PubMed texts), and the result held with or without access to the text while answering (Guidroz et al., 2025). For 37 adult second-language (L2) English users, simplification improved comprehension but increased processing time during initial reading (Rets & Rogaten, 2021). In English-as-a-foreign-language studies reported at second hand, simplified and elaborated texts both beat the unmodified text with no difference between them (483 Japanese students), and another found no significant difference between simplified and elaborated input (180 Korean students) (Le, 2011). Another study found simplified text aided comprehension at textbase and situation-model levels, against the view that elaboration helps (Ogiso, 2019).

Other studies find no gain. L2 learners who read simplified versions of four authentic texts did not recall more than those who read the authentic versions, and a high percentage of the changes were lexical (Young, 1999). Simplified passages did not raise L2 learners’ intake of target grammar, and the effect differed by experience level (Leow, 1993). In a thesis with ESL readers rating how well they understood sentences, replacing a low-frequency verb with a common word, or keeping it with a parenthetical definition, did not differ significantly from unmodified sentences (Brewer, 2008).

For technical terms, how the explanation is attached matters. A glossary raised perceived understanding but not content knowledge, whereas replacing the terms raised both (Kerwer et al., 2021). When no synonym exists, adding a parenthetical explanation still significantly simplified medical text, and placement of the term versus the explanation mattered (Gu et al., 2017). Hover definitions did not reduce the sense of difficulty or disengagement that jargon caused in 650 adults (Grabmeier, 2020).

Voice, nominalisation and abstraction

A nominalisation turns a verb into a noun (“the decision was made” for “they decided”). In a thesis study of full scientific reports, active versus passive voice made no significant difference to reading speed or reading comprehension (Rhodes, 1997). In 40 Italian adults, event nominalisations gave longer fixations and more regressions (eye movements back to earlier text) than verbal structures, but comprehension accuracy did not differ, because it was at ceiling (Venagli et al., 2025).

Whether experts write more abstractly is also unsettled. In a corpus of 10 expert and 26 student legal memos, passive and nominalisation use did not clearly separate more skilled from less skilled novices, and experts’ use varied widely (Hartig & Lu, 2014). Experienced applied-linguistics authors used significantly more nominalisation than MA and PhD students, the opposite of the pattern in which novices write more abstractly (nominalisation only, Iranian writers) (Sarani & Talati, 2015). Taddicken and Thoms state that experts underestimate the complexity of their language, and in a different finding report that more complex expert statements got higher overall debate ratings (Taddicken & Thoms, 2025).

Length and density

Little controlled work varies length and measures a reader outcome. In an experiment with 24 participants judging language-model explanations, length affected accuracy only when the explanation was wrong, and medium-length explanations went with higher accuracy than shorter or longer ones (Friedman et al., 2026). Length conditions were tertiles of word count (42 to 150 words observed) rather than fixed cutoffs, length was not manipulated independently of reading time, and one model and English-speaking participants were tested (Friedman et al., 2026). Longer explanations raised user confidence without improving accuracy, and participants’ ability to tell correct from incorrect answers was near chance with long explanations (Steyvers et al., 2025). Verbosity as a property of language-model output is treated in Verbosity in language models.

A study of 120 readers found that higher propositional density (more ideas per words) impaired surface-form memory and enhanced textbase memory, with no meaningful effect on situation-model memory (Harrag et al., 2025). In a 334-person randomised trial, a medium-length redesigned consent form outscored the long form on comprehension, and a 3-page short form (the FDA summary template) did not differ from the long form (Emanuel et al., 2026). Short privacy notices informed users, and the authors conclude that effective notice length may be bounded, since removing expected practices sometimes lowered awareness of those practices (Gluck et al., 2016).

Headings, lists and emphasis

Headings aided free recall of unfamiliar topics but not familiar ones, and affected summary inclusion more when a topic was discussed only briefly (Lorch & Lorch, 1996). A meta-analysis of signaling (cues that point out the organisation of material) finds an overall positive effect of d = .38 on learning outcomes, moderated by study quality and design features (Alpizar et al., 2020). In multimedia material, signaling gives a small-to-medium gain in transfer and comprehension, strongest for low-knowledge learners (Richter et al., 2019).

Bulleted lists have a mixed record. In three of five studies of functional texts, bulleted lists improved recall of the listed items, the benefit being weaker when the listed components were heterogeneous, and in two studies they lowered recall of content in the surrounding text (Jansen, 2014). In a citing experiment with 247 HR specialists, paragraphs beat bulleted lists in selection-decision efficiency in the difficult scenario; it measured decisions, not recall of surrounding text (Spinu & Corbeanu, 2023).

Highlighting central text information did not change its recall and reduced rereading of peripheral text, a process cost for non-emphasised text (Yeari et al., 2017). A meta-analysis of 36 articles (85 effect sizes) found that instructor-provided highlighting improved memory and comprehension (average 0.44 each), while learner-generated highlighting improved memory (0.36) but not comprehension (0.20) (Ponce et al., 2022).

Layering and progressive disclosure

In a 749-person study, participants were faster with standardised privacy-policy formats than with natural-language policies, but accuracy suffered for layered policies (McDonald et al., 2009). Collapsible or collapsible-with-preview layouts did not give a group-level comprehension advantage over continuous scrolling text, and the collapsible layout produced the most fragmented reading with frequent skipping; the study did not measure whether collapsed sections were opened or which details were missed (Xiao et al., 2026a).

Redundancy and seductive details

Redundancy (the same information in two forms, such as narration plus identical on-screen text) can help novice learners but impedes learning for learners with high prior knowledge, and the harm is explained by extraneous cognitive load (Trypke et al., 2023). The load idea ties to limits discussed in Working memory.

Seductive details are interesting but irrelevant additions. A meta-analysis of 58 studies with more than 7,500 students found that they lowered learning performance, and the effect was worse when the detail sat next to relevant diagrams (Zaske, 2020). A 2026 multi-level meta-analysis (50 studies, 177 effect sizes) finds a smaller negative effect, g = -0.16, strongest on comprehension; sample size, language and instructional context moderate it, and extraneous load is the main mediator (Wu et al., 2026c). Per a secondary summary of an earlier meta-analysis, seductive details were most harmful for static details, paper material, allowed backtracking and learners with no prior knowledge, with no difference by format, timing, size, placement, frequency or learning objectives (Paul, 2021). Another meta-analysis found prior knowledge can compensate the negative effect conditionally; low-prior-knowledge readers of the base passage outperformed on recall and problem solving (Wang & Adesope, 2016). The effect is driven mainly by diversion, and reducing diverting potential is advised (Kienitz et al., 2023).

Numbers, probability words and hedging

A meta-analysis found that people perceive verbal probability terms (common, uncommon, rare) as higher than numeric equivalents, overestimating risk (Büchter et al., 2014). A systematic review of 33 studies found that numerical readings of verbal probability terms vary widely, and most patients preferred numbers alone or with verbal labels; it could not draw conclusions on numeracy because included studies reported little demographic data (Andreadis et al., 2021). In one experiment variability in how people read probability words fell substantially when a numeric range was added (Wintle et al., 2019). Less numerate people’s risk perceptions were more affected by likelihood format, though all numeracy levels could use stated likelihood without narrative evidence; with narrative evidence, decision makers used a percentage likelihood better than frequency or verbal formats (Dieckmann et al., 2009).

Hedges in machine-written answers are a recent focus. First-person uncertainty phrasing (“I’m not sure, but…”) significantly lowered agreement with the AI against control, while general-perspective phrasing lowered it by less and not significantly; first-person uncertainty raised accuracy from 63.9% to 72.8%, general-perspective to 67.9%, not significant (Kim et al., 2024). Expressing uncertainty on a specific question lowered accuracy when the AI was correct, and the AI in that study had low accuracy and was poorly calibrated (Kim et al., 2024). A 2026 study of three uncertainty displays found that token-level uncertainty raised agreement and relation-level uncertainty reduced external verification, so uncertainty cues do not uniformly lower reliance (Villavicencio et al., 2026). Across five languages, people relied more on strongly worded certainty than on weak or moderate expressions; in Japanese they relied more on all epistemic markers (78.91% for strengtheners against 66.34% in English) (Rathi et al., 2025). A 156-person study found medium verbalised uncertainty gave higher trust, satisfaction and task performance than high or low; it varied the level, not first-person versus impersonal phrasing (Xu et al., 2025c).

Questions, controversies, and new developments

Does simplification help the people it is meant for?

The studies in Core concepts disagree. Language-model simplification raised comprehension in a large randomised study (Guidroz et al., 2025), but a randomised trial of a bundled human-written plain-language version of vaccine recommendations found no significant difference in youths’ understanding (50.9% vs 45.7%, P=.11), though readers rated it more accessible and satisfying (Stallwood et al., 2023). A GPT-4 revision of clinical guidelines gave only a small readability gain, and pharmacists found omissions in 20%, additions in 5% and meaning changes in 12% of 153 subsections; reader comprehension was not measured (Jones et al., 2026). The vaccine and guideline studies bundle several changes at once, so the effect of each rule is not isolated.

Effects of simplifying text differed by reader group (with or without intellectual disability) and by whether simplification was manual or automatic (Säuberli et al., 2024). In a randomised trial with adults with intellectual disability, simplifying the text did not by itself improve understanding; only simplified text with mediation (a helper) beat complex text with mediation (Buell et al., 2020).

Who writes the summary and for whom

Readers highly familiar with a topic tended to skip details such as study limitations in overly plain summaries, while simpler summaries gave the best reading experience for readers with little familiarity (August et al., 2024). Whether this reflects a loss of necessary information or reasonable skimming is not settled by these studies.

Writing for a reader who knows less

Writers overestimate what readers know. Studying facts made estimators overestimate what novices know and made their estimates less able to tell easy from hard items (Tullis & Feder, 2022). Hinds found that experts predicted novice task times worse than less expert people and resisted debiasing; the task was time prediction, not writing (Hinds, 1999). For the wider theory see Common ground.

Fixes are weak. Explicit instructions to attend to an uninformed addressee’s perspective did not reduce overestimation of shared knowledge (Damen et al., 2018), and directing attention to another person’s perspective before and during perspective-taking did not reduce egocentric projection (Damen et al., 2020). Experimentally reducing reliance on one’s own knowledge did not improve the accuracy of estimates (Tullis & Feder, 2022), and in Tullis and Feder’s second experiment feedback worsened calibration while improving resolution (Tullis & Feder, 2022). The authors attribute the error to missing diagnostic cues about others (Tullis & Feder, 2022). Narrative feedback reduced egocentric bias in perspective judgments; accuracy feedback did not (Damen et al., 2021). GPT-4 explanations matched their intended educational audience only half the time, versus 79% for lay human-curated ones (Joshi et al., 2025).

Comparing examples

Gick and Holyoak found that two prior analogs gave 80% total solutions on the transfer problem against 53% for one analog (45% vs 21% before a hint), and that summaries, verbal principles or diagrams given with a single analog did not foster transfer while the same devices helped when paired with two; the authors caution that failure to learn from one example may reflect missing guidance, since examples were presented incidentally (Gick & Holyoak, 1983). A meta-analysis of 57 experiments found that comparing cases usually beat other case-study formats or control conditions (d = .50, 95% CI [.44, .56]); learning was greater when learners found similarities and when principles came after the comparison (Alfieri et al., 2013). In worked-example study, students gained most from high-variability examples (Paas & van Merriënboer, 1994). These results come from problem-solving tasks; for explanatory text the evidence is thinner. Teaching the abstract concept of semiotics found no benefit from three different examples over one example repeated three times, though one particular example was more memorable and gave better learning (Funkhouser & Nicoladis, 2023), and a dissertation on expository passages found examples did not boost learning and memory compared with succinct statements; it did not compare one example with several (Cassady, n.d.).

Figurative language

A meta-analysis of 41 studies found that autistic participants comprehended figurative language less well than typically developing peers, but the difference was small and nonsignificant when groups were matched on language ability, so core language skill carries the effect (Kalandadze et al., 2018). L2 readers showed no consistent literal or figurative bias in reading phrasal verbs, and the authors advise training learners to recognise figurative meanings (Yu et al., 2025). The Open Book crossover trial (243 autistic adolescents and adults) found simplified texts, which replaced idioms and lexicalised metaphors with simple words among many other changes, gave more correct answers than originals (11.2 vs 10.0, d = 0.3) (Cerga-Pashoja et al., 2019). The paper did not report outcomes by type of adaptation, so the figurative-language step is not isolated from sentence splitting, vocabulary and image changes (Cerga-Pashoja et al., 2019).

Typography and accessibility aids

Bionic Reading (bolding the first letters of each word) gave no reading-time advantage over normal text across 100 paragraphs (Snell, 2024). In a 53-reader eye-tracking study, selective bolding did not change fixations, fixation durations or reading speed; the paper reports no comprehension scores and tested no dyslexic readers (Beelders, 2025). A news report of a 90-reader registered study says 43 readers named unbolded text the most readable, 31 start-bolded and 16 middle-bolded (Parker et al., 2026). OpenDyslexic gave no gain in reading rate or accuracy for students with dyslexia, individually or as a group (alternating-treatment design, 12 students enrolled), and the authors found no documentation that a specialised font is evidence-based practice (Wery & Diliberto, 2016). The Dyslexie font neither helped nor hindered reading in children with or without dyslexia (Kuster et al., 2018).

Content-level aids fare better, in small samples. In 62 readers (31 with dyslexia), bolded keywords raised comprehension for readers with dyslexia (Wilcoxon W = 178.5, p = 0.022) with no significant effect for the control group (p = 0.155); there was no effect on reading time or fixation duration for either group (Rello et al., 2014). The result rests on two 158-word news texts with hand-annotated keywords, and the authors explain the gain by a possible memory aid (Rello et al., 2014). Colour-isolated words in sentences slowed reading at the word but did not improve memory-based retrieval, so visual isolation alone does not guarantee a memory gain (Hofmeister & Vasishth, 2014). Text-to-speech and read-aloud tools had a small-to-moderate positive effect on reading comprehension for students with reading disabilities (22 studies) (Wood et al., 2018).

Explanations and reliance on AI answers

Explanations raised reliance on both correct and incorrect LLM responses in a pre-registered experiment (N = 308); providing sources, or inconsistencies within explanations, reduced reliance on incorrect responses (Kim et al., 2025b). Bansal and colleagues found explanations raised acceptance of AI recommendations whether or not they were correct (Bansal et al., 2021); Vasconcelos and colleagues show scenarios where explanations reduce overreliance, those where they lower the cost of verifying the AI (Vasconcelos et al., 2023). In a 125-participant study, only the dual (for and against) explanation reduced false trust below the solution-only baseline while accuracy rose (60.4% against 53.8%, 25 per condition) (Palod et al., 2026). See Explanation and understanding.

Language variety and who is served

Responses of ChatGPT to non-standard English varieties were rated worse for stereotyping and condescension than responses to standard varieties (Fleisig et al., 2024). Undesirable behaviours (lower accuracy, truthfulness, refusals) occur more for users with lower English proficiency, lower education, or from outside the US (Poole-Dayan et al., 2024). For a fuller discussion of adaptation see Inclusive design.

Broader connections

Rules about comprehension sit on three theories: cognitive load, which explains why redundancy and seductive details cost readers (Working memory); the grounding of messages in what the reader already knows (Common ground); and parsing and prediction in sentence processing (Language). Adapting to readers with different needs overlaps with Inclusive design and Neurodivergence and communication.

Further reading

• (Cunningham et al., 2018): tests Flesch-Kincaid and Lexile against measured difficulty; the clearest statement of what the formulas cannot do.
• (Crossley et al., 2023): tests readability formulas against adult comprehension of Wikipedia texts.
• (McNamara et al., 1996): the reverse cohesion study, showing that coherence helps novices and not experts.
• (Spencer et al., 2018): separates the effects of cohesion, syntax and vocabulary in adolescent readers.
• (Guo et al., 2025): compares human and language-model plain-language summaries on ratings and comprehension.
• (Alfieri et al., 2013): meta-analysis on comparing cases.