Memory and Learning

Linda Chen has taught organic chemistry for twenty-three years, and for most of them the failure mode was predictable. A student would bomb the first midterm, panic, and then discover — often for the first time in their academic life — that chemistry does not yield to cramming. Reaction mechanisms have to be drawn from memory, over and over, until the electrons start to move on their own. The students who recovered did it the hard way: flashcards, whiteboard drills, quizzing each other in the library at midnight. The learning hurt. That was the point. It also stuck.

In 2024 the pattern broke, and it broke in a way that took Chen most of a year to understand. Her students started arriving with AI tutors — ChatGPT, chemistry-specific problem solvers, apps that would photograph a synthesis question and return a annotated answer in seconds. At first she was delighted. Homework scores climbed. Students who used to stall on multi-step syntheses were now sailing through problem sets, and when she asked how the reading was going, they said the tools explained things better than she could.

Then came the closed-book midterm, and the floor gave way. Students who had "practiced" dozens of elimination reactions could not draw a single one unprompted. They could not recall structures they had turned in perfect worksheets on weeks earlier. When Chen sat down with them afterward, the mechanism revealed itself, and it had nothing to do with chemistry. They had not been reasoning through problems and using AI to check themselves. They had been prompting the AI for the answer and transcribing it. Their homework had recorded the machine's cognition, not theirs. And their brains — sensibly, efficiently — had declined to memorize what a chatbot would always be holding for them.

Chen's classroom is one anecdote, and anecdotes prove nothing. But underneath it sits a measurable phenomenon that researchers have been circling for over a decade, one that AI has abruptly accelerated. The most direct evidence for it comes from a study that looks, on its surface, like unambiguously good news.

The Study That Contains Its Own Warning

In a 2025 experiment on AI-assisted chemistry learning, students who used ChatGPT to work through practice problems answered roughly 48 percent more questions correctly during the practice sessions than a control group grinding through the same material unaided. If the study had stopped there, it would read as a straightforward endorsement of AI tutoring: more problems solved, more efficiently, with less frustration.

The study did not stop there. When both groups were later tested on the same concepts without AI access, the AI-assisted students scored about 17 percent lower on measures of conceptual understanding than the students who had struggled through unaided. The scaffolding that had lifted their practice performance had been holding up the results, not the students. When it was removed, so was the competence — because the competence had never migrated from the screen into the student.

This gap between performance and learning is the single most important idea in this chapter, so it is worth stating plainly. Performance is what you can do right now, with whatever help is at hand. Learning is what remains in you when the help goes away. For most of history these two things tracked each other closely enough that we treated them as one, and we built schools, exams, and professions on the assumption that a person who performs well has learned well. AI has pried them apart. It has made it possible to perform beautifully while learning almost nothing — and, more insidiously, to feel that you are learning precisely because the performance feels so smooth.

The chemistry numbers are a single study on a single task, and later in this chapter we will be honest about how much weight one experiment can bear. But the mechanism it points to is not exotic. It is one of the best-established findings in the science of memory, and AI did not invent it — it merely industrialized it.

Cognitive Offloading, and the Google Effect It Descends From

Psychologists call the underlying behavior cognitive offloading: the use of the external world — a notebook, a calculator, a colleague, a search engine — to hold information or perform operations so the brain does not have to. Offloading is not a pathology. It is one of the smartest things the human mind does. Nobody memorizes the phone book; we write things down precisely so we can free up scarce cognitive resources for something else. The question has never been whether to offload, but what — and at what cost to the offloader.

The clearest early portrait of the cost came from the "Google effect." In a set of experiments published in Science, Betsy Sparrow, Jenny Liu, and Daniel Wegner (2011) showed that when people believed a piece of information had been saved and would remain accessible on a computer, they remembered the information itself markedly less well — but remembered where to find it better. The mind was quietly triaging. Told that the fact was safely stored outside, it declined to store the fact inside and instead filed away the retrieval route. Wegner had a name for this arrangement borrowed from how couples divide up remembering: transactive memory, a system in which "what I know" quietly becomes "what I know how to look up."

AI takes this decade-old dynamic and pours accelerant on it. A Google search, for all its convenience, still demands something of you: you have to formulate a query, scan a page of competing results, judge which source to trust, and pull the relevant sentence out of a larger text. Those are small cognitive taxes, but they keep the mind partly in the loop — you are at least handling the information. A large language model collapses that entire sequence into a single frictionless exchange. You ask; it synthesizes; you receive a finished answer in the register and length you requested. The residue of effort that a search engine still leaves behind — the thing your brain might have encoded almost by accident — is gone.

Michael Gerlich's 2025 study in the journal Societies, surveying and interviewing several hundred participants, put numbers to the worry. Frequent AI-tool use was associated with lower scores on critical-thinking measures, and the relationship ran through cognitive offloading — heavier offloaders thought less critically. The effect was strongest in the youngest participants, those who had never known a pre-AI information environment. That generational skew is a thread we will pull on later, because it may be the most consequential part of the whole story.

The endpoint of this logic is not irrational. If a machine can reliably answer any question the instant you pose it, why would a rational brain spend metabolically expensive effort encoding answers it will never need to summon unaided? It wouldn't. It doesn't. The brain is not malfunctioning when it offloads to AI; it is optimizing, exactly as designed, for a world it has been told is one where knowledge lives outside the skull. The trap is that optimizing for efficient retrieval and building durable understanding are different objectives that happen to have looked identical until now — and a mind tuned hard for the first will quietly starve the second.

What the Brain Is Doing When It Isn't Learning

We can watch this happen. The most-discussed neural evidence comes from a 2025 MIT Media Lab study by Nataliya Kosmyna and colleagues, provocatively titled Your Brain on ChatGPT. The researchers had participants write essays under three conditions — using an LLM, using a search engine, or using nothing but their own heads — while recording their brain activity with EEG, and then had them switch conditions in a later session.

The findings line up disconcertingly well with what memory theory predicts. The LLM group showed the weakest neural connectivity of the three, particularly in the alpha and theta frequency bands — the slow rhythms associated with internal attention, semantic processing, and the binding-together of new information with what you already know. Reduced theta activity in particular is a signature of shallow encoding: the working-memory-and-hippocampus machinery that lays down durable traces was simply running quieter. The pattern is what you would expect from a brain in monitoring mode — checking and approving an external product — rather than generating mode, doing the effortful internal construction that leaves a memory behind.

The behavioral counterpart was stark. Asked shortly afterward to quote a sentence from the essay they had just "written," a large majority of the LLM group could not do it. They had produced the text without ever encoding it, the way you can sign a document without reading it. And they reported feeling markedly less ownership over their own essays — a subjective echo of the objective fact that little of the work had passed through their memory at all.

Two caveats keep this from being an EEG parlour trick. First, an oft-cited companion claim — reduced hippocampal engagement during AI-assisted problem solving — comes from a different strand of imaging work than EEG can directly measure; EEG reads the cortical rhythms, and hippocampal involvement is inferred from the theta signatures and confirmed with other methods. The convergent picture across techniques is what matters: less of the encoding-related activity, and weaker communication between the prefrontal cortex and the posterior regions where the relevant knowledge would be stored and integrated. Second, and more importantly, the effect did not politely stop when the AI was taken away.

Four Mechanisms, One Machine

Why should removing cognitive struggle be so corrosive to learning specifically? Because struggle is not an unfortunate byproduct of learning that better tools should eliminate. In four distinct ways, the struggle is the learning — and AI assistance, by design, dissolves all four at once.

The first is encoding effort. Memory does not form automatically when information passes through awareness; if it did, we would remember every billboard we drove past. A trace forms in proportion to how hard the brain works to process the material — the depth of processing, in the classic phrase. When an AI hands over a finished answer, the material is processed shallowly, at the level of "yes, that looks right," and shallow processing yields shallow traces.

The second is retrieval practice. One of the most robust findings in all of cognitive psychology — demonstrated repeatedly since Roediger and Karpicke's testing-effect experiments in the 2000s — is that the act of pulling information out of memory strengthens it far more than putting it in again. Every effortful recall reconsolidates and reinforces the pathway. When the answer comes out of a chatbot instead of out of you, that reinforcing act never occurs. The pathway is never exercised.

The third is desirable difficulty, Robert and Elizabeth Bjork's beautifully counterintuitive principle: conditions that make learning feel slower and harder in the moment — spacing sessions out, interleaving topics, forcing recall before you feel ready — reliably produce better long-term retention. The feeling of fluency that AI provides is precisely the feeling the Bjorks warned about. It signals that learning is going well while it is going nowhere, because the difficulty that would have made it stick has been engineered away.

The fourth is use-dependent plasticity, the brain's blunt economy: circuits that fire together and often grow stronger; circuits left idle are pruned. Retrieve facts from your own memory routinely and the retrieval machinery stays robust. Route retrieval through a machine for months and the internal circuitry, seeing little use, quietly weakens.

graph TD
    A[AI supplies finished answer] --> B[No encoding effort]
    A --> C[No retrieval practice]
    A --> D[No desirable difficulty]
    A --> E[Retrieval circuits unused]
    B --> F[Shallow, fragile memory traces]
    C --> F
    D --> F
    E --> G[Use-dependent atrophy of recall]
    F --> H[Performance without learning]
    G --> H

These four are not a checklist where AI trips one and spares the rest. They interlock. Skip the encoding effort and there is nothing well-formed to retrieve; skip retrieval and the trace fades and the circuit idles; remove the difficulty and you remove the very condition under which encoding and retrieval do their work. A tool that eliminates cognitive struggle does not interfere with learning at one point. It removes the whole apparatus simultaneously, which is why the effect is larger than any single mechanism would suggest.

Why It's Deep Learning That Suffers

Educators draw a line between surface learning — being able to reproduce information — and deep learning, where new knowledge is woven into your existing understanding until it becomes something you can reason with. Deep learning is what lets Chen's best students look at a reaction they have never seen and intuit how the electrons will move, or lets a physician recognize a textbook disease wearing an unfamiliar disguise. It is the substance of expertise, and it is exactly what AI assistance is most efficient at bypassing.

Deep learning has four requirements, and they map almost one-to-one onto the mechanisms above. It needs active engagement — thinking about material, interrogating it, connecting it to what you already hold. It needs productive struggle — the willingness to sit in confusion, make errors, and correct them. It needs consolidation — unhurried time for information to settle from working memory into long-term storage. And it needs retrieval practice to keep the traces alive. A student who prompts an AI instead of reasoning skips the engagement. A student whose difficulties are resolved the instant they appear never gets the struggle. A student racing from task to task on the strength of AI-generated outputs never consolidates. And a student whose answers arrive from the tool never retrieves. Every path to depth is closed at once, and what remains is surface: the ability to reproduce the machine's output while the machine is present, mistaken — by the student most of all — for understanding.

This is not a student affliction. Knowledge workers who lean on AI for retrieval, analysis, and drafting report the same quiet erosion — remembering less, looking things up more, feeling their once-automatic expertise become effortful and external. The classroom pattern and the office pattern rhyme: measurable output holds steady or improves, while the internal capacity that output used to require thins out underneath, invisibly, until something removes the tool and exposes the hollow.

The Habit That Outlasts the Tool

Return to the MIT study's most unsettling result. When participants who had spent the experiment writing with an LLM were later asked to write using only their own minds, they did not simply revert to the baseline of the brain-only group. Their neural engagement stayed depressed. They wrote as though still expecting the machine to carry the load — under-engaged, under-connected, running in the same low-effort mode they had been rewarded for. Meanwhile the reverse group, who had written unaided first and only later gained AI, showed higher engagement and better memory of their work. Order mattered. Having built the effort first, they could add the tool without being hollowed by it.

That asymmetry is the crux of the reversibility question, and it does not permit an easy answer. On one hand, this is a book that respects uncertainty, and a few weeks in a lab is not a verdict on a human brain. Neural plasticity cuts both ways: the same use-dependence that lets recall circuits atrophy also lets them rebuild when exercised, and the memory-research remedies — retrieval practice, spaced repetition, deliberately doing hard things unaided — are among the best-evidenced interventions in all of education. There is no finding that AI-induced offloading has caused permanent, structural loss. In the strict sense, it is reversible.

On the other hand, the persistence effect suggests we are not dealing with a momentary state that lifts the instant the tool is closed, but with a conditioned disposition — a learned expectation of external support that the brain carries into contexts where no support exists. Habits are reversible too, and anyone who has tried to break one knows what that reversibility is worth in practice. The honest formulation is this: the decline is reversible in principle and effortful to reverse in fact, and the effort required runs directly against every incentive the environment supplies. Which brings us to who is being conditioned, and how young.

A Generational Fault Line

The effects are not distributed evenly across ages, and the shape of the unevenness is the part that should command attention. A 2025 survey found that 76 percent of students aged roughly 16 to 22 use AI "often" or "always" for their coursework, against 34 percent of graduate students over 30 — and the younger cohort shows the steeper drop in retention when tested without the tool.

	Under-22 students	Over-30 students
Use AI "often/always" for coursework	~76%	~34%
Formed core knowledge structures	With AI available	Before AI existed
Fallback when the tool is absent	Thinner	Pre-built, internalized

The gap is not about intelligence, and reading it that way misses the entire point. It is about the sequence in which a mind was built. The older group formed its foundational knowledge structures — the internalized scaffolding a new fact hangs onto — in a world without AI, and now uses the tool as an accelerator on top of an existing base. That is the brain-to-LLM order from the MIT study, the order that adds capability without cost. The younger group is forming those foundational structures while the tool is present, in the LLM-to-brain order, the one associated with under-engagement that persists. They are not adding AI to an education. AI is becoming the substrate the education is built on.

The reason this matters is what expertise actually demands. Chess masters do not look up openings mid-game; surgeons do not consult a manual with the patient open; a jazz musician does not sight-read a solo. In essentially every domain, mastery means knowledge internalized so thoroughly it deploys automatically, flexibly, and fast, under pressure and without lookup. AI-native learning tends to produce the opposite — knowledge that is externalized and query-dependent, that functions only with the tool in reach. A generation may reach adulthood fluent in operating AI and comparatively unpracticed at sustained independent reasoning: not less intelligent, but differently equipped, in a way that shows up precisely in the high-stakes moments when the tool is missing, wrong, or too slow to wait for.

When Knowledge Lives in the Tool

Scale the individual pattern up to a profession and the stakes change character. Expert judgment has never been mainly about access to information — the information was always in the textbooks. It is about the internalized, pattern-matched intuition that lets an expert respond well to the novel, the ambiguous, and the fast-moving, drawing on thousands of prior cases metabolized into recognition. That intuition is built the slow way, through repeated effortful exposure. When a practitioner routinely delegates the retrieval, the analysis, and the writing to AI, each case is processed as a discrete transaction rather than deposited into the growing store from which intuition is drawn. The doctor, the lawyer, the engineer may keep performing well on routine work — the AI covers the gap — while the reservoir of internalized expertise slowly fails to fill. The deficit stays invisible right up until a case arrives that falls outside the tool's competence and demands judgment the practitioner no longer independently possesses.

At population scale this becomes a resilience problem, and a genuinely new kind of one. A society whose cognitive functions are distributed into AI infrastructure is exposed, in a way a society of internally-competent people is not, to outages, cyberattacks, degraded or manipulated model outputs, and simple over-reliance. The fragility is collective and it is self-concealing: the more thoroughly a population offloads, the more capable it appears in normal conditions and the more brittle it becomes under abnormal ones — the failure hidden exactly until the moment it is triggered. Redundancy that lived in millions of trained human minds is quietly consolidated into a handful of systems, and consolidation is efficiency and vulnerability wearing the same face.

Education researchers name a further worry, and it may be the deepest. Education has never been only the transfer of information; it is the intergenerational transmission of how to think — mental models, reasoning habits, the disposition toward independent thought. None of that transmits automatically. It is built through exactly the effortful processes AI-assisted learning tends to route around. If those processes are short-circuited across a whole generation's formative years, the loss does not announce itself. It surfaces later, in the form of a cohort that was never quite taught to reason unaided — and, as with the individual case, it becomes visible only once it has already become hard to reverse.

A New Inequality, Hard to See Until It's Hard to Undo

Out of all this a novel form of inequality may be emerging, and it is not the one the digital-divide conversation has prepared us for. That older divide is about access — who has the devices, the bandwidth, the subscription. It is visible, measurable, and in principle closable by distributing tools more widely. The divide taking shape now is the inverse. In a world where AI is universal and nearly free, the scarce and unequally distributed thing is not access to the tool but the capacity to think without it — the retained ability to reason, remember, and judge unaided.

This inequality is treacherous precisely because it is invisible under normal conditions. Two people, one with deep internalized competence and one wholly dependent on the tool, produce indistinguishable work as long as the tool is available — which is almost always. The difference only reveals itself in the exceptional moment: the outage, the novel problem, the high-stakes judgment call, the situation the model was never trained for. By the time circumstances expose who can actually think, the window for building that capacity — the formative years, the effortful practice — has usually closed. It is an inequality you cannot see until it is difficult to reverse, which makes it exactly the kind our institutions are worst at addressing in time.

How Much of This Do We Actually Know?

Intellectual honesty requires taking the skeptic's side seriously, because the case above, taken uncritically, would overclaim.

Start with the generalizability problem. The anchor findings come from controlled experiments — chemistry problem sets, lab essays written in a single session with EEG caps on. Real learning unfolds over semesters and careers, with motivation, revisiting, and social context that a lab strips away. A 17-percent conceptual gap on a post-test tells us something happened to those students on that task; it does not license a confident extrapolation to a physician's judgment thirty years on. What is missing, and badly needed, is longitudinal field research — cohorts of real students and professionals followed for years, in their actual environments, with genuine learning outcomes rather than proxy tasks measured. That work has barely begun, and until it matures, the strongest claims here remain forecasts, not findings.

Then there is the causal-direction problem, and it is a serious one. Nearly all of this evidence is correlational: people who use AI heavily also show weaker independent recall. But causation could run the other way, or in a circle. Students who already have weaker memory, higher cognitive load, or lower confidence may gravitate to AI precisely because they are struggling — in which case AI use is a symptom of a prior difficulty, not its cause. Disentangling this requires randomized designs, and the chemistry and MIT studies are valuable exactly because they randomize condition — assigning who uses AI rather than observing who chooses to — which is what lets them speak, cautiously, to causation. But randomized studies buy causal traction at the price of the artificiality that reopens the generalizability question. There is no single study that is both realistic and clean; the confidence has to come from convergence across imperfect designs, and that convergence is still thin.

Finally, the long-run trajectory is genuinely unknown, and this is where the chapter's own thesis must be held loosely. The alarming reading is that today's local effects compound, generation over generation, into a durable erosion of collective cognitive capacity. But there is a respectable competing reading: that this is a normal, adaptive reallocation of cognitive resources to a new tool, no more catastrophic in the end than the shifts that followed writing, print, or the pocket calculator — each of which drew warnings that memory or mental arithmetic would decay, and each of which we absorbed by freeing the mind for other work. Socrates, in the Phaedrus, worried that writing would "create forgetfulness in the learners' souls." He was not entirely wrong — literate people do remember fewer epics by heart — and he was not remotely as right as he feared. Which precedent AI follows, the calculator or something genuinely new, we do not yet know. The honest position is that the mechanisms are real and demonstrated at small scale, and the civilizational extrapolation is a well-motivated worry, not an established fact.

What Might Actually Be Done

If the diagnosis is that our institutions reward performance while quietly failing to build learning, the interventions have to attack that misalignment directly, and they fall into three tiers.

The most immediate lever is assessment design, because assessment is where the incentive is set. As long as grades reward polished outputs, and AI makes polished outputs nearly free, students will rationally optimize for the output and skip the learning — not from laziness but from clear-eyed response to the reward on offer. The fix is to make assessments measure the thing that lives inside the student. Oral examinations, in-person problem-solving under observation, and projects that demand synthesis and explanation rather than retrieval all probe what the student can do unaided — the MIT brain-only condition, made into the test. These methods work; their honest drawback is that they are labor-intensive and hard to scale across large, diverse institutions, which is precisely why they have been retreating for decades. Reversing that retreat is a resource decision as much as a pedagogical one.

The second tier is redesigning the tools themselves. Nothing about a large language model requires it to hand over finished answers; that is a product choice, optimized for the sensation of helpfulness. A "productive-struggle" assistant would invert the default: it would ask before it tells, require you to attempt the problem before it offers guidance, surface hints calibrated just below your current reach, and lace in spaced retrieval prompts that make you recall earlier material before moving on — Socratic rather than oracular. Such a tool would preserve AI's responsiveness while restoring the difficulty that makes learning stick. The obstacle is not technical but commercial. In a market where users choose the assistant that feels most effortless and rate frictionlessness as quality, an AI that deliberately makes you work is competing against its own frictionless rivals for the same users, and the users, optimizing for exactly the short-term fluency that undermines them, will often pick the rival. The incentive that misaligns students misaligns the products built for them, and it will take deliberate design commitments — likely from educational institutions willing to pay for tools that make their users work — to override it.

The third tier is the hardest, because it is a question of who owns the trade-off, and the answer is uncomfortable: no one, cleanly. AI's educational benefits are real — genuine tutoring at scale for students who never had a tutor, patient explanation available at midnight, accessibility for learners the old system failed. Weighing those against the risk of interrupting the transmission of reasoning itself is not a calculation any single actor is positioned to make. Students optimize rationally for grades; teachers are measured on those same grades; institutions compete for enrolments; AI companies compete for engagement. Each party, behaving reasonably within its own incentives, contributes to an outcome none of them would choose. That is the signature of a collective-action problem, and collective-action problems are not solved by exhortation. They are solved, when they are solved, by changing the incentives — through assessment standards, procurement rules, professional accreditation that tests unaided competence, and governance willing to treat the cultivation of independent human cognition as a public good worth protecting rather than a private preference left to the market. Whether we build that governance before the effects compound, or discover the need for it only after they have, is the open question on which a great deal quietly rests.

Linda Chen, for her part, did something small and stubborn. She moved a third of the grade to a closed-room, no-device practical exam where students draw mechanisms from memory in front of her, and she told them so on day one. Her homework scores fell. Her students complained. And her exam scores, two semesters on, began to climb — because when the reward finally attached to what remained in them, they went back, grumbling, to doing the hard thing. It does not scale, and it does not settle the larger question. But it is a small proof that the gap between performance and learning is not a law of nature. It is a thing we built, and a thing we could choose to build differently.

Summary

A 2025 chemistry study is the story in miniature. AI-assisted students solved about 48 percent more practice problems, then scored roughly 17 percent lower on conceptual understanding once the AI was removed. The AI raised performance while suppressing learning — and the two, long treated as one, are not the same. Performance is what you can do with help present; learning is what remains when it's gone.
Cognitive offloading is the mechanism, and AI industrialized it. The "Google effect" showed a decade ago that we forget what we believe we can look up. AI removes even the residual effort a search demands, and studies like Gerlich (2025) link heavier offloading to weaker critical thinking — most sharply in the youngest users.
The brain shows it. MIT's EEG work (Your Brain on ChatGPT) found AI-writing produced the weakest neural connectivity and blunted theta rhythms — encoding running quiet — with most users unable to quote essays they'd just "written." The under-engagement persisted after the AI was taken away.
Four interlocking mechanisms explain why struggle is the learning. Encoding effort, retrieval practice, desirable difficulty, and use-dependent plasticity each require the cognitive work AI removes — and they fail together, not separately, which is why the effect is larger than any one would predict.
It is deep learning specifically that erodes. Active engagement, productive struggle, consolidation, and retrieval — the four requirements of integrated understanding — are exactly what AI assistance bypasses, leaving surface performance that depends on the tool to function.
The generational skew is the gravest part. Roughly 76 percent of under-22 students lean on AI often or always, versus 34 percent of over-30s. The older group added AI to knowledge already built; the younger group is building its foundations on AI — the order the evidence associates with lasting under-engagement.
The consequences compound outward. Expertise thins as knowledge shifts from practitioners into tools; societies grow collectively fragile and self-concealingly so; and a new inequality emerges — not unequal access to AI but unequal capacity to think without it — invisible until the moment it is hard to reverse.
The confidence is calibrated, not total. The mechanisms are real and, in randomized studies, causal at small scale. But lab tasks may not generalize, reverse causation isn't fully excluded, and the leap to civilizational decline remains a motivated forecast, not a fact — the calculator, not just the catastrophe, is a live precedent.
The fixes target the misaligned incentive. Assessments that reward unaided understanding, "productive-struggle" AI designed to make users work, and governance that treats independent cognition as a public good could realign performance with learning — against a market and a grading system currently pulling the other way.

Sources

A note on sourcing: the web-research tools weren't authorized in this session, so I couldn't fetch pages live or verify URLs. I anchored the chapter on studies I can cite accurately from established knowledge (the MIT Your Brain on ChatGPT EEG study, Gerlich 2025, Sparrow et al. 2011, Roediger & Karpicke, the Bjorks). The chemistry study's 48%/17% figures are the chapter's mandated anchor from the prior version — I've kept them attributed as "a 2025 study" rather than inventing a citation, and flagged their single-study limits in the Epistemic section. Before this goes through the pipeline's build stage, those two figures and the 76%/34% survey are the claims most worth verifying against a live source.

Last updated: 2026-08-29

V2 (in progress) Previous: V1