What the Summary Leaves Out

In September, the OECD released the 2025 PISA results: reading and mathematics performance had declined sharply across its member countries, the lowest averages the test has recorded since it began. One in five fifteen-year-olds now scores as a low performer in reading, mathematics and science combined, up from one in six just three years earlier. The reflexive explanation, offered by education officials and reported nearly everywhere the results ran, pointed at artificial intelligence. Nearly half of OECD students now use AI chatbots weekly or more for schoolwork; separate research from Common Sense Media puts the figure at seventy percent of American teens, with sixty-three percent using AI specifically to get answers rather than to work toward them. OECD education director Andreas Schleicher's framing was blunt: learning requires "cognitive struggle," and struggle is precisely what a chatbot is built to remove.

This is a real explanation, and probably a partial one. It is also, on its own, an unsatisfying one, because it treats "AI use" as a single variable with a single effect — bad for reading, full stop. The actual research on AI-assisted reading does not support that story. It supports a stranger, more specific one, and the specificity is where the useful part of this begins.

A 2025 randomized crossover study, published in Frontiers in Education, put 195 college-aged readers through standardized comprehension passages under five conditions: no assistance, an AI-generated summary, an AI-generated outline, a question-and-answer chatbot, and a Socratic-style discussion chatbot. The results split cleanly along a line nobody had specifically predicted: AI tools significantly improved comprehension for weaker readers and significantly worsened it for stronger ones. The weakest readers gained the most from the Socratic chatbot. The strongest readers lost the most from plain summaries — a correlation of -0.785 between a reader's baseline skill and how much the AI tool helped them, meaning the better a reader you already were, the more a shortcut cost you.

That is a genuinely strange finding, and it stops being strange once you look at what a summary actually is. Reading comprehension research going back decades — the psychologists Walter Kintsch and Teun van Dijk did the foundational work in the 1970s and 80s — describes skilled reading as a constant, largely unconscious act of compression: a reader builds a "macrostructure," a gist, out of a text's local details, discarding exact wording almost as fast as it's encountered while retaining meaning. This is not a flaw in human reading. It is how reading works, and it's why you can describe a novel's plot a year later without recalling a single sentence from it. A skilled reader is already running their own internal summarizer, tuned over years to their own judgment about what matters and what doesn't.

An AI summary hands a reader someone else's finished compression instead of asking them to build their own. For a struggling reader, whose internal summarizer is underdeveloped or overwhelmed, that's scaffolding — a temporary structure that makes an otherwise inaccessible text tractable. For a skilled reader, it's a replacement for the exact cognitive work that made them skilled, substituted in at the moment it would have been exercised. You don't get better at compression by being handed someone else's compression. This is the same supervision-paradox shape that shows up everywhere else in AI-assisted cognition: the tool is most helpful to the people least equipped to evaluate whether it did the job well, and least helpful — actively costly — to the people who could have evaluated it, because it pre-empts the evaluation itself.

There's a second, harder-to-name effect sitting underneath the comprehension data, and it concerns not what gets learned but what gets attended to. A 2025 meta-analysis in Frontiers in Psychology, synthesizing 124 experiments on digital reading, found a moderate-to-strong negative effect of interruption and multitasking on comprehension (Hedges' g = -0.64) — confirmation of something reading researchers have argued for years: comprehension of anything nontrivial requires sustained, uninterrupted attention, and the digital reading environment is structurally hostile to supplying it. A chat interface doesn't just offer an answer; it offers permission to stop reading before the answer would have arrived on its own. A separate review of generative AI in academic reading and writing describes this pattern as "diminished metacognitive engagement" — not an inability to comprehend, but a quieter erosion of the impulse to keep reading toward comprehension when a faster route is sitting one click away.

Two honest caveats belong here, because the panic narrative around reading and AI tends to skip both. First, the decline in reading for pleasure predates the chatbot era by two decades: a 20-year analysis of the American Time Use Survey found daily reading-for-pleasure participation falling from 28 percent in 2004 to 16 percent in 2023 — before ChatGPT existed, driven most plausibly by the broader displacement of leisure time onto screens generally. AI is not introducing a new phenomenon so much as accelerating one already well underway. Second, the "kids don't read anymore" framing has always run ahead of its evidence — reading habits have shifted in format and genre for generations without necessarily collapsing in substance, and PISA's cross-sectional score alone can't distinguish a generation reading less from one reading differently.

What the comprehension research actually supports is narrower and more useful than either the panic or its rebuttal: AI reading tools are not a verdict on reading, they're a lever, and which way the lever swings depends entirely on whether the person pulling it still has a comprehension process of their own to protect. For a reader still building that process, the tools can help. For a reader who already has one, the tools are a standing offer to stop using it — and the offer gets harder to decline exactly as thinking, not just answers, becomes the more expensive thing to produce.

Sources

1. PISA 2025: Students' reading and mathematics performance declined sharply across the OECD — OECD (Press Release)
2. Reading scores slide amid worries that AI is shortcutting learning — Journalism
3. Differential effects of GPT-based tools on comprehension of standardized passages — Frontiers in Education (Peer-Reviewed)
4. Distractions in digital reading: a meta-analysis of attentional interference effects — Frontiers in Psychology (Peer-Reviewed)
5. The impact of generative AI on academic reading and writing: a synthesis of recent evidence (2023–2025) — Frontiers in Education (Peer-Reviewed)
6. The decline in reading for pleasure over 20 years of the American Time Use Survey — iScience (Peer-Reviewed)