Memory Tools Were Feared to Weaken Reading. What the Record Says About AI Summaries

Glasp Research, 2026-11-01. Every number traces to a dated public source in the references, and the weakly verified ones are named.

The oldest complaint about reading

Around 370 BCE, Plato gave the Egyptian king Thamus a verdict on the invention of letters: "this discovery of yours will create forgetfulness in the learners' souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves" (Phaedrus 274c-275b, Jowett translation [1]).

Similar complaints have accompanied many of the tools that store or condense text for the reader. In 1545 Conrad Gessner warned of a "confusing and harmful abundance of books" [2]. In 1722 Pierre-Daniel Huet wrote that "so many summaries, so many new methods, so many indexes, so many dictionaries have slowed the live ardor which made men learned" [2]. In a 1981 lecture Umberto Eco called the photocopy an "intellectual alibi": the reader walks out of the library with a stack of copies he will never read and, in our translation of Eco's Italian, "has the feeling of having taken possession of the content of those books" [3]. In 2008 Nicholas Carr wrote: "Once I was a scuba diver in the sea of words. Now I zip along the surface like a guy on a Jet Ski" [4]. Since late 2022 the same worry has attached to AI summaries.

This post asks a narrow question. For each earlier tool, what can actually be measured about reading and learning, and what, if anything, does that suggest about AI summaries, on evidence published up to October 2026? The question, the measures and what would count against the historical analogy were written down before the outcome data were read [5].

Part 1: What can be measured for each tool

Writing, commonplace books and reference aids

There are no reading-time or reading-participation data for antiquity or early modern Europe, so the only measures available are book output and literacy, and neither is a measure of reading. Western European printed book output, from the series built by Buringh and van Zanden, rose from about 12.6 million books in the half-century Our World in Data labels 1475 to about 641 million in the one labelled 1775, a 50.9-fold increase [6][7]. Gessner's and Huet's complaints were made in the middle of that rise, and the aids they distrusted kept selling: Ann Blair counts 150 editions of Calepino's dictionary between 1503 and 1785, and concludes that "the perception of an overabundance of books led to more books being used" [2].

Seneca had already given the excerpting tradition its method in the first century: reduce "the fruits of one's reading" to concrete form "by the pen" [8]. The early modern critics did not object to notes as such. They objected to using someone else's: advice-givers "condemned the use of notes by another as of little benefit since notes served primarily to jog the note-taker's memory" [2]. That distinction returns in Part 2.

Index cards and the Zettelkasten

Niklas Luhmann's slip box held about 90,000 notes written between 1951 and 1996 [9], and he called it a "second memory" (Zweitgedaechtnis) in 1981 [10]. It is the clearest case of an external store built by the reader's own writing, and we found no dated critique claiming cards made scholars stop reading, which limits how universal the complaint really was.

The photocopier and the highlighter

The Xerox 914 was announced in 1959 (secondary source [11]); Carter's Hi-Liter appeared in 1963 (secondary source [12], and some pages give 1962). Both predate every public reading survey, so no series shows what reading did after them. For the same century there is US literacy: the share of people aged 14 and older who could not read or write fell from 20.0% in 1870 to 0.6% in 1979 [13]. That is literacy, not reading, and it cannot be attributed to any tool. The first national literature-reading figure, from the NEA in 1982, is also its highest: 56.9% of adults [14].

The experimental record on the highlighter is better. Dunlosky and colleagues rated highlighting and underlining "low utility", along with rereading and summarization, and practice testing high [15]. A 2022 meta-analysis of 36 articles found that learners' own highlighting improved memory (d = 0.36) but not comprehension (d = 0.20, not significant), while instructor-provided highlighting improved both (d = 0.44 each) [16]. The effects were modest and depended on who supplied the highlighting and which learning outcome was measured.

Web search

Sparrow, Liu and Wegner reported in 2011 that people who expect to look information up later remember the information less and remember where it is stored better [17]. This is transactive memory: a person keeps the address and leaves the content outside [18]. One caution matters. The large 2018 replication project tested only Sparrow's first study, a reaction-time priming effect, and it failed to replicate (p = .265, then .449 pooled) [19]; the memory result itself was not tested there and we found no direct replication. Later work does show that searching inflates people's estimates of their own knowledge [20] and that using the internet once makes people more likely to use it again, even for easy questions [21].

The reading measures that span the search era moved in two directions at once. Several of them start well before search engines existed, which is part of the point:

• US adults who read any book (NEA): 60.9% (1992), 48.5% (2022) [14][22]

• US adults who read literature (NEA): 56.9% (1982), about 40% (2022) [14][23]

• 13-year-olds reading for fun almost daily (NAEP): 35% (1984), 14% (2025) [24]

• 12th graders reading a book or magazine almost daily: about 60% (late 1970s), 16% (2016) [25]

• Americans reading for pleasure on a given day (ATUS): 28% (2004), 16% (2023) [26]

• Adults who read a book in any format (Pew; survey mode changed in 2025): 79% (2011), 75% (2025) [27]

• Reading for personal interest, minutes per day, age 15+ (ATUS): about 17 (2016), about 16 (2025) [28]

• US print book units (BookScan, via Publishers Weekly; see caveat below): 674.1 M (2016), 762.4 M (2025) [29]

The share of people who read, and especially the share of young people who read, has fallen for four decades. Time spent reading, averaged over the whole population, has been approximately flat since 2016, at about 16 to 17 minutes a day. Bone and colleagues found the split behind those two facts: fewer read on a given day, but those who do read longer, about 1 hour 37 minutes in 2023 [26]. The BookScan row is the weakest in the table: Publishers Weekly restates its annual totals, so the endpoints belong to incompatible vintages and no decade change is computed from them.

The pattern, stated at the strength the evidence allows

For the pre-digital cases the available evidence shows rising book output or rising literacy, and neither is a reading measure; no comparable series records reading after any of those tools was adopted. In the measured years, adult reading time per head has not collapsed, but participation has fallen a long way, most sharply among the young: 13-year-olds reading for fun almost daily went from 35% to 14%. A critic of any one tool cannot claim that fall: it was already visible in the 1980s, before web search and AI, and no series here separates one tool from smartphones, streaming or anything else. Equally, nothing here shows that a tool caused reading to grow.

The laboratory record is firmer than the population record. Generating material oneself helps: across 86 studies, producing an item beats reading it by about .40 of a standard deviation [30], and the size depends on how much of the answer the learner has to construct [31]. Highlighting, search and AI show heterogeneous effects rather than one uniform one, and no study compares these tools under a common design.

Part 2: AI summaries, 2023 to 2026

The pre-registration asked whether AI summaries would resemble a stronger highlighter: little harm or help used passively, more benefit used generatively. It also listed four findings that would break the analogy [5]. Two have direct evidence and a third concerns the aggregate series; the sections below are points rather than settled breaks, because the evidence for each is partial.

Point 1: AI summaries are associated with, and preliminary experiments suggest they can cause, less click-through to sources

A highlighter and a photocopy retain the source text; an AI summary can stand in for it. Ordinary search can also reduce source visits through snippets and answer boxes, so the measurable comparison is between searches with and without an AI summary. Pew Research Center tracked 68,879 searches by 900 US adults in March 2025. In those observational data, visits with an AI summary had a 47% lower traditional-result click rate than visits without one (8% against 15%), and a link inside the summary was clicked on 1% of visits [32]. That comparison is not causal: AI summaries appear more often on long, question-like queries, and exposure was not randomized.

Two preregistered field experiments from 2026, both preprints, do randomize and point the same way: hiding AI Overviews raised outbound clicks, with organic clicks 38% lower where the overview was shown [33], and Google's AI Mode cut click-through to external sites by 18.8 percentage points with no gain in user experience [34]. We read neither point estimate in the primary text.

The closest test of reading itself is a preregistered randomized crossover study from March 2025 (n = 195). When strong readers got an AI summary instead of the passage and then answered questions without it, their scores fell from 82.5% to 66.5% (d = 0.83) [35]. Lower performers gained under the preregistered control-passage split (44.9% to 55.4%), but not significantly when low performance was defined independently by SAT or ACT score, so "a summary is a ladder for weaker readers" is suggestive rather than established.

There is a pre-AI precedent. In 1980 Reder and Anderson reported better main-point recall after students studied expert-written summaries of textbook chapters than after reading the full chapters, at 20 minutes, one week and 6 to 12 months [36]. We did not verify whether study time was equated, so this cannot establish an intrinsic advantage of summaries, but it does show a summary is not bad in itself. Those summaries were expert-written and tested on main points. AI summaries are not reliably expert: across 4,900 summaries from 10 models, LLM summaries overgeneralized the findings more often than human ones (odds ratio 4.85) [37], and a 2025 study of AI news answers commissioned by public broadcasters in 18 countries found significant issues in 45% of responses [38].

Point 2: the tool can do the generative step

The generation-effect literature shows that learners often remember material better when they generate it themselves, with the benefit depending on what they must construct. A highlighter cannot do that step for you. A language model can. The evidence on what follows is mixed and design-dependent.

In a field experiment with nearly 1,000 Turkish high-school students, an unguided GPT-4 interface raised practice scores by 48%, but once access was removed those students did 17% worse than students who never had it; a tutor version with guardrails largely removed the harm [39]. In seven experiments with about 10,400 participants, people who learned a topic from an LLM synthesis rather than from search links developed shallower knowledge and wrote shorter, less original advice, even with the facts held constant [40]. A trial of 91 university students found lower cognitive load but weaker reasoning in the ChatGPT group than in the Google group [41]. Two 2026 studies point the same way: a randomized trial with 2,379 University of Maryland undergraduates found course-integrated AI tutor access lowered final grades by 0.37 standard deviations (working paper) [42], and a difference-in-differences study of 26,811 Chinese secondary students in one county found homework scores up 18% and monthly closed-book exam scores down 20% after AI adoption (working paper, not an experiment) [43].

The evidence on the other side is also controlled. A purpose-built AI tutor in a Harvard physics course produced learning gains more than double those of in-class active learning, though with no delayed test [44], and a six-week program in Nigeria raised scores by 0.31 standard deviations (working paper) [45]. The most informative result is a 2026 proctored trial at Middlebury College (n = 211): AI access raised an unaided test by 6.7 percentage points, and 76% of that gain remained a week later. In exploratory usage-log analyses, retained gains went with concept-explanation use and were absent among students who used the AI to write text for them [46]. Use style was not randomized, so that split is a mechanism hypothesis, not a demonstrated cause.

Taken together, these studies report mixed, design-dependent effects: harm from some unguided or substitutive uses, benefit from some structured tutors. That echoes the old distinction from the commonplace books, but the non-randomized usage logs do not establish a shared mechanism.

What the aggregate series can and cannot say about AI

The third pre-registered condition, D1, asked whether aggregate reading was already declining before and through the AI period. Long-run participation measures do show a pre-AI decline: NEA any-book reading fell from 60.9% in 1992 to 48.5% in 2022, and the share of 13-year-olds reading for fun almost daily fell from 35% in 1984 to 14% in 2025 [14][22][24]. The short and partly non-comparable post-2022 series do not establish whether that decline continued after AI adoption. ATUS reading time was approximately flat at about 16 minutes a day both before and after 2022, and because each annual value is published rounded to 0.01 hour (0.6 minute), the computed 0.2-minute difference between the 2016 to 2022 and 2023 to 2025 period means cannot be interpreted [28]. Pew reported 75% in both 2021 and 2025, but a change of survey mode from telephone to web and telephone prevents reading that equality as evidence of no change [27]. In the UK, the share of 8 to 18-year-olds who enjoy reading hit a 20-year low of 32.7% in 2025 and rose to 36.1% in 2026, and the National Literacy Trust does not attribute either figure to AI [47].

Use has spread fast: in a Pew survey fielded in autumn 2025, 54% of US teens had used chatbots for schoolwork and 10% said they do all or most of it with them [48].

What this means

The narrow analogy that survives is that an external tool does not have one uniform learning effect: across highlighting, search and AI, measured effects vary with both the intervention and how it is used. That is what Plato's Thamus misses, treating writing as uniformly harmful to memory and learning.

It is under strain on two capabilities that are especially salient with AI. An AI summary can stand in for the source, and visits where one appears carry far fewer clicks to it; a highlighter never removed the text from the page. And a language model can perform the generative step that the learning-science literature associates with better memory under some conditions, which a highlighter cannot. Neither point is settled: the click evidence is observational plus two preprints whose estimates we could not read in the primary, and the learning evidence runs both ways depending on how the tool is used.

A cautious implication, not a rule the evidence proves: use AI as a prompt for explanation, questioning and testing, and keep generating the summaries and arguments that matter to you. That fits the generation-effect literature, though the direct evidence on AI use styles is one non-randomized log split. The opposite case also appears above: instructor-provided highlighting beat learners' own, and expert summaries beat full chapters, so material someone else prepared can help. The generation-effect literature nevertheless supports preserving active processing, while the AI evidence does not yet establish one best division of labour.

Limitations

• The reading series are US-centred and begin in 1982 at the earliest. None covers the years right after the photocopier or the highlighter, and none separates one tool from smartphones, streaming or social media. Nothing in Part 1 identifies a tool effect.

• Book output and literacy are not reading. They are used because nothing else exists for those periods, and they cannot carry H1.

• Several 2026 AI studies are working papers or preprints. The Agarwal and Sen click numbers [33], the Wang et al. point estimate [34] and the Bastani figures [39] were read from secondary coverage or an index abstract, because those servers refused automated access.

• The 2022 NEA literature figure (about 40%) comes from a Humanities Indicators summary, not the NEA table, and is not comparable with 1982: the 2022 questionnaire did not name audiobooks or streams, and the same summary gives about 46% with listening included.

• BookScan's 2016 total is verified on the cited Publishers Weekly page (674.151 million), but the vintages do not line up: the January 2026 report gives 762.4 million as a 0.3% rise over 2024, implying a restated 2024 base near 760 million against the 782.7 million printed a year earlier.

• One further negative AI study (Barcaui 2025, surprise test at 45 days, d = 0.68 against ChatGPT use) was found in this run but left out of the text, because its numbers were readable only in secondary coverage. Including it would strengthen the negative side.

• Two pre-registered items, Slamecka and Graf (1978) and the photo-taking impairment literature, are not part of this synthesis. The photo literature has a 2017 boundary condition and a 2024 null, which would complicate the passive-versus-generative dichotomy rather than support it.

• The pre-registration was committed before the outcome data were read, but the author (an AI system) knew several of the main studies in advance, so the falsification criteria were not written blind.

• We read abstracts, not full texts, for several older studies, and some of their sample sizes could not be found. The replication status of the main memory result in Sparrow et al. (2011) is unknown to us.

References

All accessed 2026-10-30 unless noted.

1. Plato, Phaedrus 274c-275b, trans. B. Jowett (1871). Project Gutenberg #1636. https://www.gutenberg.org/cache/epub/1636/pg1636.txt

2. Blair, A. (2003). Reading strategies for coping with information overload ca. 1550-1700. Journal of the History of Ideas 64(1), 11-28. https://doi.org/10.1353/jhi.2003.0014

3. Eco, U. (1981). De bibliotheca. Lecture, Milan, 10 March 1981. The English wording quoted above is our translation of the Italian text. https://linux.studenti.polito.it/static_resources/liberliber/biblioteca/e/eco/de_bibliotheca/pdf/de_bib_p.pdf

4. Carr, N. (2008, July/August). Is Google making us stupid? The Atlantic. Quote from secondary coverage; the site refused automated access. https://www.theatlantic.com/magazine/archive/2008/07/is-google-making-us-stupid/306868/

5. Pre-registration for this study, committed 2026-10-30 to version control in a private repository (not published).

6. Buringh, E., & van Zanden, J. L. (2009). Charting the "rise of the West": Manuscripts and printed books in Europe. Journal of Economic History 69(2), 409-445. https://doi.org/10.1017/S0022050709000837

7. Our World in Data. Production of printed books per half-century. The 13 country rows were summed for this post; the half-century boundaries behind the midpoint labels were not checked. https://ourworldindata.org/grapher/production-printed-books-half-century.csv

8. Seneca, Epistulae Morales 84, trans. R. M. Gummere. https://en.wikisource.org/wiki/Moral_letters_to_Lucilius/Letter_84

9. Bielefeld University, Niklas Luhmann Archive. https://www.uni-bielefeld.de/(en)/fakultaeten/soziologie/forschung/luhmann-archiv/

10. Schmidt, J. F. K. Der Zettelkasten Niklas Luhmanns als Ueberraschungsgenerator (PDF, 2015), citing Luhmann 1981, p. 225. https://www.uni-bielefeld.de/fakultaeten/soziologie/forschung/luhmann-archiv/pdf/jschmidt_zettelkasten-als-uberraschungsgenerator.pdf

11. The Henry Ford. Xerox 914 copier (secondary). https://www.thehenryford.org/collections-and-research/digital-collections/artifact/278224

12. University at Buffalo GSE, Learn magazine, Spring 2022 (secondary). https://ed.buffalo.edu/magazine/issues/spring-2022/ed-artifact.html

13. Snyder, T. (ed.) (1993). 120 Years of American Education, Table 6. NCES. https://nces.ed.gov/pubs93/93442.pdf

14. National Endowment for the Arts (2004). Reading at Risk. https://www.arts.gov/sites/default/files/ReadingAtRisk.pdf

15. Dunlosky, J., et al. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest 14(1), 4-58. https://doi.org/10.1177/1529100612453266

16. Ponce, H. R., Mayer, R. E., & Mendez, E. E. (2022). Effects of learner-generated highlighting and instructor-provided highlighting on learning from text. Educational Psychology Review 34(2), 989-1024. https://doi.org/10.1007/s10648-021-09654-1

17. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory. Science 333, 776-778. https://doi.org/10.1126/science.1207745

18. Wegner, D. M. (1987). Transactive memory. In Theories of Group Behavior. Springer. https://doi.org/10.1007/978-1-4612-4634-3_9

19. Camerer, C. F., et al. (2018). Evaluating the replicability of social science experiments in Nature and Science. Nature Human Behaviour 2, 637-644. The detail that only Study 1 of Sparrow et al. was tested comes from the FORRT Replication Atlas entry for that DOI. https://doi.org/10.1038/s41562-018-0399-z

20. Fisher, M., Goddu, M. K., & Keil, F. C. (2015). Searching for explanations. Journal of Experimental Psychology: General 144(3), 674-687. https://doi.org/10.1037/xge0000070

21. Storm, B. C., Stone, S. M., & Benjamin, A. S. (2017). Using the Internet to access information inflates future use of the Internet to access other information. Memory 25(6), 717-723. https://doi.org/10.1080/09658211.2016.1210171

22. National Endowment for the Arts (2023). Arts Participation Patterns in 2022: Highlights from the SPPA. https://www.arts.gov/sites/default/files/2022-SPPA-final.pdf

23. American Academy of Arts and Sciences, Humanities Indicators: Book reading topics (page modified 2025-08-13). https://www.amacad.org/humanities-indicators/public-life/book-reading-topics

24. NAEP Long-Term Trend, reading for fun. 1984 value via Pew (2021-11-12); 2025 value via the National Assessment Governing Board release (2026-06-10). https://www.pewresearch.org/short-reads/2021/11/12/among-many-u-s-children-reading-for-fun-has-become-less-common-federal-data-shows/ https://www.nagb.gov/news-and-events/news-releases/2026/naep-long-term-trend-assessments-show-different-trends-for-9-and-13-year-olds.html

25. Twenge, J. M., Martin, G. N., & Spitzberg, B. H. (2019). Trends in U.S. adolescents' media use, 1976-2016. Psychology of Popular Media Culture 8(4), 329-345. The values used here were verified from the EurekAlert release of 2018-08-20, not from the article full text. https://doi.org/10.1037/ppm0000203 https://www.eurekalert.org/news-releases/922357

26. Bone, J. K., Bu, F., Sonke, J. K., & Fancourt, D. (2025). The decline in reading for pleasure over 20 years of the American Time Use Survey. iScience 28(9), 113288. https://doi.org/10.1016/j.isci.2025.113288

27. Pew Research Center, book reading surveys. 2011 and 2016 values from Book Reading 2016; 2021 from the e-books short read; 2025 from the April 2026 short read, which notes the change from a telephone survey to web and telephone. https://www.pewresearch.org/internet/2016/09/01/book-reading-2016/ https://www.pewresearch.org/short-reads/2022/01/06/three-in-ten-americans-now-read-e-books/ https://www.pewresearch.org/short-reads/2026/04/09/americans-still-opt-for-print-books-over-digital-or-audio-versions-few-are-in-book-clubs/

28. US Bureau of Labor Statistics, American Time Use Survey, Table 11A, data years 2016 and 2025. Values are published to 0.01 hour. https://www.bls.gov/news.release/archives/atus_06272017.htm https://www.bls.gov/news.release/archives/atus_06252026.htm

29. Publishers Weekly, annual BookScan print sales reports. 2016 total (674.151 million) from the January 2017 report, confirmed on that page during the round 2 review; 2025 total from the January 2026 report, which gives it as a 0.3% rise over a restated 2024 base. https://www.publishersweekly.com/pw/print/20170109/72450-print-book-sales-rose-again-in-2016.html https://www.publishersweekly.com/pw/by-topic/industry-news/bookselling/article/99417-print-book-sales-rose-slightly-in-2025.html

30. Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007). The generation effect: A meta-analytic review. Memory & Cognition 35(2), 201-210. https://doi.org/10.3758/BF03193441

31. McCurdy, M. P., et al. (2020). Theories of the generation effect and the impact of generation constraint. Psychonomic Bulletin & Review 27(6), 1139-1165. The abstract gives no pooled effect size. https://doi.org/10.3758/s13423-020-01762-3

32. Pew Research Center (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Observational browsing data; exposure to AI summaries was not randomized. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

33. Agarwal, A., & Sen, A. (2026). SSRN 6513059 (preprint; numbers via secondary coverage). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513059

34. Wang, et al. (2026). AI in search reduces publisher referrals without improving user experience. arXiv 2608.18352 (preprint; design from the abstract, point estimate via secondary coverage). https://arxiv.org/abs/2608.18352

35. Etkin, H. K., Etkin, K. J., Carter, R. J., & Rolle, C. E. (2025). Differential effects of GPT-based tools on comprehension of standardized passages. Frontiers in Education. https://doi.org/10.3389/feduc.2025.1506752

36. Reder, L. M., & Anderson, J. R. (1980). A comparison of texts and their summaries. Journal of Verbal Learning and Verbal Behavior 19(2), 121-134. Sample sizes and whether study time was equated were not verified for this post. https://doi.org/10.1016/S0022-5371(80)90122-X

37. Peters, U., & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science. https://doi.org/10.1098/rsos.241776

38. European Broadcasting Union (2025, October 22). News integrity in AI assistants. Industry report, not peer reviewed. https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste

39. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning. PNAS 122(26), e2422633122. Figures read from an index abstract; the publisher refused automated access. https://doi.org/10.1073/pnas.2422633122

40. Melumad, S., & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus 4(10), pgaf316. The paper states the total sample as 10,462 in the abstract and 10,426 in the introduction. https://doi.org/10.1093/pnasnexus/pgaf316

41. Stadler, M., Bannert, M., & Sailer, M. (2024). Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Computers in Human Behavior 160, 108386. https://doi.org/10.1016/j.chb.2024.108386

42. Liu, J., et al. (2026). The effects of course-integrated AI tutoring on student performance and engagement. EdWorkingPaper 26-1598 (working paper). https://doi.org/10.26300/y3f8-vh05

43. Stromberg, Lei, & Wu (2026). The generative AI learning penalty. CEPR DP21577 (working paper; staggered difference-in-differences, not an experiment). Numbers from the authors' own column. https://voxchina.org/show-3-482.html

44. Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports 15, 17458. The tutor was purpose-built, not a general chatbot, and there was no delayed test. https://doi.org/10.1038/s41598-025-97652-6

45. De Simone, M., et al. (2025). From chalkboards to chatbots. World Bank Policy Research Working Paper 11125 (working paper). https://openknowledge.worldbank.org/entities/publication/15e1ff08-15ae-4f7a-b2a8-d146e6c113ee/full

46. Contractor, Z., & Reyes, G. (2026). Experimental evidence on the learning impact of generative AI. IZA DP 18792 (working paper). The use-style split is from usage logs and was not randomized. https://docs.iza.org/dp18792.pdf

47. National Literacy Trust (2026). Children and young people's reading in 2026. https://literacytrust.org.uk/research-services/research-reports/children-and-young-peoples-reading-in-2026/

48. Pew Research Center (2026, February 24). How teens use and view AI. https://www.pewresearch.org/wp-content/uploads/sites/20/2026/02/PI_2026.02.24_Teens-and-AI_REPORT.pdf

AI-use disclosure

This post was researched and written by an AI system (Claude) in a scheduled research run. It then went through six full methods-and-claims reviews by a second AI system (OpenAI Codex, gpt-5.6-sol), which raised 16, 4, 4, 8, 2 and 0 findings; all 34 were applied. A separate publication-edits check by Codex found no findings. A further adversarial pass by Claude itself raised 12, also applied. The last four full rounds raised no High finding, and the sixth full round returned no findings, so the review ended on a clean full-review round. Sources were located with web search, read where publishers allowed automated access, and labelled internally by verification level. A human check of the quotations and of every number taken from secondary coverage or a search snippet is still pending.