TL;DR. The evidence on reading splits into two questions that are usually confused. The first is acute: does this particular text cost you comprehension because it is on a screen? A meta-analysis of 54 studies and 171,055 participants puts that penalty at Hedges’ g = -0.21, and it gets worse under time pressure. The second is chronic: what does a reading habit build over years? Here the gap is not small. Print reading habits correlate with comprehension at roughly r = .36 to .41. Digital leisure reading habits, measured across 469,564 participants, correlate at r = .055 overall, and the effect is negative for primary and middle school readers. The chronic question is the one worth acting on, and almost nobody is asking it. Meanwhile verified EU data shows the 16 to 24 age group is the only one whose online reading has fallen since 2019. Deep reading is becoming rare at exactly the moment machines made shallow reading free.
The argument everyone is having is the wrong one
Search for the state of reading and you land in a decade-old fight about paper versus screens. That fight has a real answer. It is also, on the evidence, the smaller half of the problem.
The paper-versus-screen answer comes from a 2018 meta-analysis by Delgado, Vargas, Ackerman and Salmeron, published in Educational Research Review. It pooled studies from 2000 to 2017 comparing comprehension of the same texts on paper and on digital devices: 38 between-participants studies and 16 within-participants studies, 171,055 participants in total. Both designs produced the same result, an advantage for paper of Hedges’ g = -0.21.
Three moderators reached significance, and they matter more than the headline number:
- Time frame. The paper advantage grew when reading was time-constrained rather than self-paced.
- Text genre. The advantage held for informational texts and for mixed sets, but not for studies using only narrative texts.
- Publication year. The paper advantage increased over the years covered.
A follow-up experiment by Delgado and Salmeron, published in Learning and Instruction, tested the time-frame moderator directly rather than inferring it. One hundred and forty undergraduates were assigned to one of four conditions crossing medium (print versus screen) with time frame (free versus pressured). The result is unusually clean. In-print readers mind-wandered less under time pressure than under free time, showing they adapted to the deadline. On-screen readers did not adapt, and on-screen readers in the pressured condition comprehended less than the other three groups. Under free reading time, mind-wandering and comprehension were similar regardless of medium.
Two things follow. Remove the clock and, in this experiment, the medium penalty went away. And the authors found no differences in readers’ metacognitive calibration between conditions, meaning the readers who comprehended less did not know they had. The cost is invisible from the inside, which is why it does not correct itself.
Read those moderators together and the finding stops being about the device. A penalty that appears under time pressure, on informational text, and that grows as digital reading environments mature, is a penalty attached to how screens are read, not to the glass. The full text of that meta-analysis is behind a publisher paywall and the sub-group effect sizes for each moderator level were not accessible for this piece, so the moderator directions are reported here as the authors state them, without magnitudes.
Now hold that number, g = -0.21, and look at the second question.
The chronic question: what does the habit build?
In 2011 Mol and Bus published a meta-analysis in Psychological Bulletin covering 99 studies and 7,669 participants, tracking leisure time reading from infancy to early adulthood. Their central finding is that the association between print exposure and reading ability does not stay flat. It compounds.
The figures they give for variance explained in oral language skills climb steadily with schooling:
Table 1. Print exposure: share of variance explained in oral language skills, by educational stage
| Educational stage | Variance explained by print exposure |
|---|---|
| Preschool and kindergarten | 12% |
| Primary school | 13% |
| Middle school | 19% |
| High school | 30% |
| College and university | 34% |
Source: Mol and Bus, Psychological Bulletin, 2011 (99 studies, N = 7,669). Figures as reported in the published abstract. Outcome domain for these specific percentages is oral language skills; the authors report moderate to strong correlations for reading comprehension and for technical reading and spelling as well.
From preschool to university the share of variance explained rises by 22 percentage points, a 2.83x increase. The authors describe this as an upward spiral: better readers read more, and reading more makes them better readers. Poor readers, they note, also benefit from independent leisure reading.
That is the compounding asset. The question that matters in 2026 is whether the digital version of the same habit compounds the same way.
It does not
In 2023, Altamura, Vargas and Salmeron published a meta-analysis in Review of Educational Research asking precisely that: does leisure digital reading relate to text comprehension the way print reading does? They screened studies published between 2000 and 2022, retained 26 studies yielding 40 independent effect sizes, and covered 469,564 participants. That is a sample nearly 61 times the size of the print meta-analysis.
The overall answer: r = .055, 95% CI [.003, .107], p = .03, computed across 39 effect sizes after one influential outlier was removed. Statistically significant, and close to nothing.
The authors are explicit that this contrasts with the print literature, and cite the print figures as r = .36 for Grades 1 to 12 and r = .41 for higher education readers. Those two numbers come from Mol and Bus and are quoted here as reported in the 2023 paper, because the primary abstract gives variance-explained percentages rather than these correlations.
The stage breakdown is where the finding gets uncomfortable:
Table 2. Leisure digital reading habits and text comprehension, by educational stage
| Educational stage | Effect sizes (k) | r | Variance explained (derived) |
|---|---|---|---|
| Primary and elementary | 2 | -0.095 | 0.90% |
| Middle school | 14 | -0.025 | 0.06% |
| High school | 14 | +0.085 | 0.72% |
| Undergraduate | 6 | +0.070 | 0.49% |
| Overall | 39 | +0.055 | 0.30% |
Source: Altamura, Vargas and Salmeron, Review of Educational Research (26 studies, 40 effect sizes, 469,564 participants). The r values are the published meta-regression estimates. The variance-explained column is derived by squaring r and is a CEOtudent calculation, not a figure from the paper. Educational stage was a significant moderator (p = .047). Two pairwise comparisons reached significance: middle versus high school (p = .021) and primary versus university (p = .030).
For the two youngest groups the relationship is negative. More leisure digital reading, slightly worse comprehension. For the two oldest it turns positive, and stays trivially small.
Two honesty notes that most coverage of this study skips. First, heterogeneity is extreme: I-squared = 98.9, Q(38) = 2591.93, p < .001, with a prediction interval running from -.18 to .29. That interval spans zero comfortably, which means the pooled .055 describes an average across populations that behave very differently, not a reliable expectation for any one of them. Second, the authors tested for publication bias and found none. Egger’s test of the intercept was not significant (b = -.039, 95% CI [-1.39, .61], p = .427), and published and unpublished studies did not differ (p = .525). The small effect is not an artifact of a drawer full of hidden results. It is what the field actually found.
The original comparison: two ladders running in opposite directions
Put the two literatures side by side. Nobody does this, because they sit in different sub-fields and answer different questions. Together they say something neither says alone.
Table 3. The payoff gap: what a reading habit correlates with, print versus digital
| Comparison | Print reading habit | Digital leisure reading | Ratio (r) | Ratio (variance explained) |
|---|---|---|---|---|
| Grades 1 to 12 against overall digital | r = .36 | r = .055 | 6.55x | 42.8x |
| Higher education against undergraduate | r = .41 | r = .070 | 5.86x | 34.3x |
| Higher education against high school (best digital cell) | r = .41 | r = .085 | 4.82x | 23.3x |
CEOtudent editorial framework. The print figures are Mol and Bus (2011) as reported in Altamura et al.; the digital figures are Altamura et al.’s own estimates. Both sides are correlations between a self-reported reading habit and a comprehension outcome, which is what makes the ratio meaningful. The ratio columns are CEOtudent calculations. No ratio is given for the primary and middle school cells because the digital correlations there are negative and near zero, which makes a ratio arithmetically unstable and substantively meaningless.
The narrowest honest reading of Table 3: even taking the most flattering digital cell in the dataset and the print figure from the same source, a print habit tracks comprehension roughly five times as strongly as a digital one, and explains something like twenty to forty times as much variance.
Now set that against the screen penalty from the first section. The acute cost of reading a given text on a screen is g = -0.21. The chronic gap between reading diets is, on the same evidence base, several times larger in every direction you slice it. Both literatures are correlational on the habit side and experimental on the medium side, so this is a comparison of magnitudes, not a causal accounting. But the direction of the conclusion survives the caveat:
The device you read on is a second-order problem. The diet you read is the first-order one. Optimising your e-reader settings while your reading consists of feeds, threads and generated summaries is tuning the wrong variable.
That is the whole argument for deep reading as a competitive advantage, and it is why “just print it out” is bad advice offered in good faith.
Who has actually stopped reading
The habit data above comes from research samples. The population data tells you where the scarcity is forming.
Eurostat tracks whether individuals read online news sites, newspapers or news magazines. This is not deep reading, it is the lowest bar for reading connected prose online at all. Here is the EU27 picture:
Table 4. Individuals reading online news sites, newspapers or news magazines, EU27, percentage
| Group | 2015 | 2019 | 2023 | 2024 | 2025 | Change 2019 to 2025 |
|---|---|---|---|---|---|---|
| All individuals | 52.72 | 61.80 | 64.26 | 65.05 | 66.41 | +4.61 pp |
| Low or no formal education | 30.57 | 40.21 | 43.88 | 44.08 | 45.32 | +5.11 pp |
| Medium formal education | 54.15 | 62.58 | 63.65 | 64.02 | 64.93 | +2.35 pp |
| High formal education | 75.99 | 81.84 | 81.63 | 81.54 | 82.46 | +0.62 pp |
| Aged 16 to 24 | 62.27 | 66.64 | 62.41 | 62.69 | 61.39 | -5.25 pp |
| Aged 25 to 54 | 60.46 | 69.87 | 71.33 | 71.49 | 72.86 | +2.99 pp |
| Aged 55 to 74 | 34.85 | 46.19 | 53.94 | 55.99 | 58.43 | +12.24 pp |
Source: Eurostat, individuals internet activities (isoc_ci_ac_i), indicator I_IUNW1, EU27 from 2020, percentage of individuals. Change column is a CEOtudent calculation from the published values.
One row moves against every other. The 16 to 24 group is the only cohort whose online news reading has declined since 2019, down 5.25 percentage points to 61.39%, which is 0.92 times its 2019 level. Over the same period the 55 to 74 group rose 12.24 points, a 1.27x increase, and now sits only about three points below the youngest group after starting nearly 28 points behind in 2015.
The education gradient is still wide, 37.14 points between the highest and lowest education groups in 2025, but it has narrowed from 45.42 points in 2015. Note carefully how it narrowed: the low-education group gained 14.75 points since 2015 while the high-education group gained 6.47 and has been essentially flat since 2019, moving 0.62 points in six years. The gap closed from below, not because the top kept climbing.
This is a prevalence measure, not a depth measure. Eurostat is not asking whether anyone finished the article. But when the shallowest available indicator of online reading is falling in the cohort about to enter the labour market, the deeper measure is not plausibly doing better.
Why this becomes an advantage now, specifically
Three things had to be true at once, and now are.
Summarisation became free. The marginal cost of getting the gist of any document has collapsed to near zero. That does not make gist worthless. It makes gist non-scarce, and non-scarce things do not differentiate anyone. What a summary systematically removes is the part where an argument’s structure is visible: which claim is load-bearing, where the author hedges, what evidence is absent. Those absences are where judgment gets made. If you have read our piece on why AI hallucinates and how model errors work, the mechanism is familiar: a fluent output can be structurally hollow, and detecting that requires having read enough well-built arguments to feel the difference.
Time pressure is the aggravating factor. Delgado and colleagues found the screen penalty grew under time constraint. That is the same variable our piece on slow thinking in a fast world identifies as the one that quietly degrades decisions. Reading under a deadline on a device optimised for interruption is the worst-case configuration, and it is the default configuration of most professional reading.
The habit compounds, so the gap widens. Mol and Bus found print exposure explaining 12% of variance at preschool and 34% at university. The advantage is not a fixed bonus you either have or lack. It accumulates, in the same way and for the same reason that skills decay when they are not maintained. Someone who reads deeply for five years is not 5% ahead of someone who did not. They are on a different curve.
The practice: a protocol you can actually run
The evidence above constrains what a sensible practice looks like more than most reading advice admits. It pairs naturally with what the ranked evidence on study techniques already established about effortful versus passive learning. It says: protect the diet before the device, remove the time constraint, and prefer informational text where the penalty was found rather than narrative text where it was not. It does not say anything about how many books per year you should read, so this protocol does not either.
Table 5. Deep reading protocol, mapped to the evidence that motivates each step
| Step | What to do | Evidence that motivates it | Status |
|---|---|---|---|
| 1. Fix the diet first | Set a floor of connected long-form text per week before changing anything about your devices | Habit gap (r = .36 to .41 versus .055) exceeds the medium gap (g = -0.21) | Evidence-backed direction |
| 2. Remove the clock | Read self-paced, not against a deadline; if you cannot, defer the reading rather than rush it | Time frame was a significant moderator; in a 140-participant experiment the medium penalty appeared only under time pressure | Evidence-backed direction |
| 3. Weight toward informational text | Prioritise argumentative and expository material over narrative for skill-building purposes | Genre moderator: the effect held for informational and mixed sets, not narrative-only | Evidence-backed direction |
| 4. Read one primary source per week in full | Go to the study, filing or report rather than coverage of it | Not measured by any source cited here | CEOtudent convention |
| 5. Write the argument back in your own words before summarising | Reconstruct the structure, then compress | Not measured by any source cited here | CEOtudent convention |
| 6. Use the machine after, not instead | Read first, then ask a model what you missed | Not measured by any source cited here | CEOtudent convention |
CEOtudent editorial framework. Steps 1 to 3 restate directions the cited meta-analyses actually found. Steps 4 to 6 are deliberately labelled conventions: no source in this article measured them, and they are offered as reasonable structure rather than as evidence-backed technique. Treating them otherwise would be exactly the sort of unearned confidence this article is arguing against.
The honest version of the CEO and student split here is uncomfortable. The CEO instinct is to delegate reading, because delegation is what leverage looks like everywhere else. The student instinct is to do the reading, because that is where the compounding happens. On this specific asset, the evidence sides hard with the student. You can delegate the retrieval. You cannot delegate the comprehension and keep the advantage that comes from it.
What would change this conclusion
Three things, and they are worth watching for rather than assuming away.
The digital reading measured in the Altamura meta-analysis is mostly the digital reading of 2000 to 2022: feeds, social platforms, news sites. If a genuinely different reading environment emerges, one built for sustained attention rather than against it, the .055 figure describes a technology that no longer exists. Nothing in the data rules that out.
The heterogeneity is enormous. I-squared of 98.9 and a prediction interval of -.18 to .29 mean some populations almost certainly do get a real benefit from digital reading. The pooled average is not a verdict on you.
And the print figures rest on a 2011 meta-analysis of studies conducted mostly before smartphones. It is possible that print reading habits in 2026 select for a different kind of person than they did in 2005, and that part of the r = .41 is that selection rather than the reading. No source cited here can separate those.
None of that changes what you should do this week. The direction is stable across every cut of the data: read more connected prose, read it without a clock, and read the thing rather than the summary of the thing.
FAQ
Is reading on a screen actually bad for me?
Less than the popular argument suggests. The pooled penalty across 54 studies and 171,055 participants is Hedges’ g = -0.21, a small effect by conventional standards. It gets worse under time pressure and it did not appear for narrative-only texts. Compared with the gap between reading diets, it is the second-order concern.
Does that mean I should stop reading online?
No. It means the format of what you read matters more than the surface it is displayed on. A long argued piece read attentively on a phone is a different activity from thirty minutes of scrolling, a distinction our analysis of the attention economy’s newer weapons draws out, even though both are screen reading, and no meta-analysis cited here treats them as the same thing.
Would I notice if screen reading were costing me comprehension?
On the available evidence, no. In the four-condition experiment described above, readers’ predictions of their own test performance did not differ across conditions even though on-screen readers under time pressure actually comprehended less. Self-assessment did not track the loss.
Why is the digital reading effect negative for younger readers?
The meta-analysis reports the pattern, not the mechanism: r = -0.095 for primary and elementary, -0.025 for middle school, both based on small numbers of effect sizes at the primary end (k = 2). The authors identify educational stage as a significant moderator but do not establish causation, and the negative cells should be read as weak signals, not settled facts.
How much reading is enough?
No source in this article answers that, and any number you see quoted with confidence is almost certainly invented. The evidence supports the direction of more connected reading and the conditions under which it works. It does not support a target.
Can I use AI summaries at all?
Yes, after you have read the source, as a check on what you missed. Read our note on trust calibration before treating any generated answer as evidence. Used as a replacement they remove exactly the structural information, the hedges, the load-bearing claims, the missing evidence, that reading was supposed to give you. That is the trade, stated plainly.
Is the 16 to 24 decline in the EU data real or a measurement artifact?
It is a real movement in the published series: 66.64% in 2019 to 61.39% in 2025 for EU27, while every other age and education group rose. The indicator measures whether people read online news at all, not how deeply, so it is a floor measure. It cannot tell you whether the same cohort is reading long-form elsewhere.
Sources
Delgado, Vargas, Ackerman and Salmeron. Don’t throw away your printed books: A meta-analysis on the effects of reading media on reading comprehension. Educational Research Review, 2018.
Mol and Bus. To read or not to read: a meta-analysis of print exposure from infancy to early adulthood. Psychological Bulletin, 2011.
Altamura, Vargas and Salmeron. Do New Forms of Reading Pay Off? A Meta-Analysis on the Relationship Between Leisure Digital Reading Habits and Text Comprehension. Review of Educational Research.
Delgado and Salmeron. The inattentive on-screen reading: Reading medium affects attention and reading comprehension under time pressure. Learning and Instruction.
Eurostat. Individuals, internet activities (isoc_ci_ac_i), indicator I_IUNW1, EU27 from 2020.
Note on secondary citation: the print-habit correlations of r = .36 for Grades 1 to 12 and r = .41 for higher education readers are attributed to Mol and Bus but are quoted here as they appear in Altamura, Vargas and Salmeron, because the Mol and Bus abstract reports variance-explained percentages rather than these coefficients.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
