Reading on a Screen Versus on Paper: What the Evidence Actually Shows

You have probably heard that you understand less when you read on a screen. It is one of those claims that feels obviously true, and it has a real research literature behind it - several meta-analyses, thousands of participants, twenty-five years of studies.

The literature does find a paper advantage. It is small. And almost everything interesting about it is in the qualifications: which kind of text, whether you had to scroll, whether you were under time pressure, how old you are, and whether the outcome measured was your comprehension or your confidence about your comprehension.

This article is about where that evidence is solid, where it is contested, and where the popular version of the claim runs past what anyone has shown. It is about reading medium; the separate question of what you write notes on is covered in laptops in the classroom.

The headline number, and why it is smaller than you think

Three meta-analyses published within a year of each other reached almost the same figure.

Delgado, Vargas, Ackerman and Salmeron (2018) synthesised studies published between 2000 and 2017 comparing comprehension of the same texts on paper and on screen. Across 38 between-participant studies they found a paper advantage of Hedges' g = -0.21, 95% CI [-0.28, -0.14], and across 16 within-participant studies dc = -0.21, 95% CI [-0.37, -0.06] (dc is the corrected within-participant effect size) - the same figure twice, from two different designs.

Clinton (2019) restricted her review to randomised experiments disseminated between 2008 and 2018 and found g = -0.25 across 33 studies with 2,799 participants.

Kong, Seo and Zhai (2018) analysed 17 studies and also reported a comprehension advantage for paper, which the other two meta-analyses cite at -0.21.

Three independent teams, three overlapping literatures, roughly the same answer. That convergence is the strongest thing this field has.

But an effect of about a fifth of a standard deviation is small. It is not the difference between understanding a text and not understanding it. In the most-cited experiment behind these numbers, Ackerman and Goldsmith (2011) found undergraduates scoring 63.2% on screen against 72.3% on paper under self-paced study - a real gap, and one that vanished entirely in the same paper's other experiment.

There is also a caveat about that headline figure that the Delgado paper flags itself. Its "171,055 participants" sounds enormous, but the authors note that 165,778 of them came from four large-scale studies, and that because "their large samples yielded a small confidence interval for their effect sizes, their influence on the overall effect could skew the results." The number is real; the impression of breadth it creates is not quite.

The single most replicated finding: it depends on the kind of text

If you take one thing from this article, take this. The paper advantage shows up for informational and expository text. It does not show up for narrative text.

• Delgado et al. found expository texts at g = -0.27 and mixed informational-and-narrative at -0.30, while narrative-only texts came out at g = 0.01 - no effect at all.
• Clinton found the same split: expository g = -0.32, narrative g = -0.04.
• Schwabe, Lind, Kosch and Boomgaarden (2022) ran a meta-analysis dedicated to narrative texts - 32 studies, 2,239 participants - and concluded that "the analyses did not reveal a significant impact of the reading medium (screen vs. paper) on the reading comprehension of a narrative text."

Three separate teams, using partly different study pools, agree. Reading a novel on a Kindle is not the case the evidence is about. Reading a dense chapter of a textbook is.

The finding that reframes everything: it may be a scrolling effect

Clinton-Lisell and Litzinger (2026) did something the earlier meta-analyses could not: a network meta-analysis comparing four device types and paper against each other, across 56 studies and 79 effect sizes. Their device ranking was paper, then tablets, then e-readers, then computers, then smartphones.

Then they split by whether the reader had to scroll. The result:

"When scrolling was necessary to read the text, paper was notably better for comprehension compared to digital devices (Hedges' gs 0.35-0.48). However, when scrolling was not necessary to read the text, there were no reliable differences and small effect sizes (Hedges' gs = 0.03-0.12 between paper and any digital devices."

Note the sign convention flips here - positive g means paper is better. When scrolling was required, the paper advantage was larger than any headline figure in this literature. When it was not, it was close to zero.

If that holds, the phenomenon is not really about screens. It is about whether the text stays put while you read it. A paginated e-reader is, on this evidence, nearly as good as paper. A long article you thumb through on a phone is not.

One caution. It is tempting to attribute the scrolling finding to Delgado et al. as well, because their paper reports a significant paper advantage when scrolling was required (g = -0.25) and only a marginal one when it was not (g = -0.13, p = .06). But those are within-subgroup tests, and the paper explicitly states that the between-group test failed: "No moderating effects were found for educational level, text length, type of digital device, need for scrolling, open testing, or type of comprehension." Delgado et al. did not establish a scrolling moderator. Clinton-Lisell and Litzinger tested it directly, and they did.

Time pressure: contested

Delgado et al. found that the paper advantage was larger when reading was time-constrained (g = -0.26) than when readers set their own pace (g = -0.09), with the moderation test significant at Q_B = 4.12, p = .04.

That fits the mechanism suggested by Ackerman and Goldsmith (2011), whose two experiments dissociated cleanly: under fixed study time, comprehension did not differ between media at all; under self-paced study, screens were worse. Their reading of it was that the problem is not raw cognitive capacity but how well people regulate their own study time on screens.

But a later meta-analysis by an overlapping team - Salmeron, Altamura, Delgado, Karagiorgi and Vargas (2024), focusing on handheld devices - found time frame not significant (p = .853). So does text genre, in their analysis (p = .203). Their overall effect was also about half the size: g = -0.113 between-participants, dc = -0.103 within.

We are reporting a moderator that one of its own discoverers failed to reproduce six years later on a different device class. That is worth saying plainly rather than picking the version that reads better.

Has the gap grown or shrunk? Nobody agrees

This is the part of the literature that is genuinely unsettled, and popular coverage tends to pick whichever answer suits the argument.

• Delgado et al. (2018): growing. A meta-regression found the paper advantage increased by .01 points per year since 2000 (b = -.01, Q_R = 4.95, p = .03), explaining 64% of the effect-size variance.
• Kong et al. (2018): shrinking. Their abstract reports that comparisons before and after 2013 "indicated that the magnitude of the difference in reading comprehension between paper and screen followed a diminishing trajectory" - though they note this was not statistically significant.
• Salmeron et al. (2024): no change. Publication year, p = .872.
• Schwabe et al. (2022): no change. "There does not seem to be a difference over time."

Four analyses, three answers. Anyone who tells you the digital-reading penalty is definitely getting worse - or definitely disappearing as people grow up with screens - is choosing a study rather than reading the field.

Children are a different question, and the answer is different

The adult finding does not transfer. Furenes, Kucirkova and Bus (2021) meta-analysed 39 studies of 1,812 children aged 1 to 8, and the overall comprehension difference between digital and paper books was g = -0.07, 95% CI [-0.17, 0.04] - a confidence interval comfortably including zero.

The detail matters more than the headline:

Comparison	k	g	95% CI
Digital books with no enhancements	10	-0.22	[-0.36, -0.08]
Digital books with enhancements	16	-0.03	[-0.18, 0.11]
Story-related enhancements only	8	+0.17	[0.01, 0.32]
Vocabulary learning, overall	18	+0.20	[0.08, 0.32]
Vocabulary, embedded dictionary alone	4	+0.49	[0.23, 0.74]

A plain digitisation of a picture book is worse than the book. A digital book with enhancements that serve the story is better. And for vocabulary learning, digital wins outright.

The same paper found the paper advantage concentrated in school settings (g = -0.28) rather than at home (g = +0.08) or in the lab (g = +0.06), and among children from lower-SES households rather than middle or high (p = .010). Whatever is happening here, "screens are worse for children's reading" is not a fair summary of it.

The most useful finding is not about comprehension at all

Across this literature the effect on comprehension is small and heavily conditional. The effect on knowing how well you have understood is more consistent, and more useful.

Clinton's meta-analysis found that readers were better calibrated on paper - g = 0.20 across 11 studies and 698 participants - meaning their predictions of their own test performance were more accurate. Her conclusion: "Readers may be more efficient and aware of their performance when reading from paper compared to screens."

Ackerman and Goldsmith found the same thing in the original experiments. Both groups were over-confident, but screen readers were more so: a calibration bias of 10.4 on screen against 2.3 on paper in their self-paced experiment (F(1, 72) = 6.78, p = .01).

And Clinton found no reliable difference in reading time (g = 0.08, across 14 studies) - so this is not a story about people simply rushing on screens. They took about as long, understood slightly less, and were more confident about it than they should have been.

That is the practically actionable finding in this whole literature, and it is the one that gets left out.

What this evidence does not support

• It does not support "screens are bad for reading." The effect is small, absent for narrative text, near zero when scrolling is not required, and near zero overall for young children.
• It does not support a claim about e-readers. Salmeron et al. found handheld devices produce roughly half the effect of the computer-based comparisons that dominate the older meta-analyses, and Clinton-Lisell and Litzinger ranked tablets second only to paper.
• It does not support any claim about brain changes, attention capacity, or long-term damage. Nothing in this literature measures anything of the kind. These are studies of people reading a text and then answering questions about it.
• It does not settle whether the gap is growing. See above.

What it does support

Read expository material - the textbook chapter, the dense report, the thing you will be tested on - somewhere it does not scroll, and preferably on paper. Give yourself time rather than working to a deadline, because the moderator evidence, though contested, all points the same direction where it points at all. And be a little more sceptical of your own sense that you have understood something you read on a screen, because that is the finding this literature is most consistent about.

For narrative reading, the evidence gives you no reason to change anything at all.

If the problem is less the screen than what else is on it, that is a different mechanism - see why studying next to your phone wrecks your memory and the cognitive offloading guide.

A note on what is not settled

Two of the meta-analyses here - Clinton (2019) and Kong et al. (2018) - publish point estimates without confidence intervals in their abstracts, and we could not obtain the full texts to verify their intervals. The -0.21 attributed to Kong et al. comes from Delgado et al. and Salmeron et al. citing it, not from Kong's own abstract, which reports no numeric effect size. Clinton-Lisell and Litzinger (2026) is cited here from its abstract; its per-device estimates and intervals sit behind a paywall we did not cross. Where we could not verify a number, we have not printed one.