“I barely read books lately. Is that bad for my brain?”
The short answer is: reading a book does not automatically give your brain experience points. What matters is the mental work happening while you read.
Books can bring in new knowledge, connect ideas, make you follow a long argument, expose you to things you did not search for, and give you long contact with language. Fiction can also make you imagine other minds and lives.
Now compare that with repeatedly asking an AI: “Why?”, “Really?”, “What does the research say?”, “What about this case?”, “Is this the same structure as that other thing?” Much of the same mental work can happen. AI also lets you stop at the exact point of confusion, ask for a counterargument, search for an original paper, and connect the topic to something you already know.
So not reading many books does not automatically mean you stopped doing intellectual work. Depending on how it is used, an active AI dialogue can demand more thinking than passive reading.
The danger is also obvious: AI can answer so quickly that you outsource the very parts you wanted to train—remembering, solving, and following a long chain by yourself. The rest of this article is about separating useful outsourcing from harmful outsourcing.
1. Break “the benefits of reading” into functions
“Reading is good” is too vague. Reading performs several different jobs.
First, it delivers knowledge: history, science, economics, psychology, and so on.
Second, it connects knowledge. “Interest rates” may become linked to inflation, loans, central banks, investment, and purchasing power.
Third, it makes you follow a long argument. An author builds A → B → C → D, and you keep the chain in mind.
Fourth, books expose you to information you did not choose. Search tends to answer the question you already had. A book can force an accident: “I did not care about this chapter, but apparently it matters.”
Fifth, books provide long exposure to language—vocabulary, sentence patterns, style, and explanation.
Sixth, especially in fiction, reading can act as a simulation of other people. A 2024 preregistered meta-analysis found a small average experimental cognitive benefit from reading fiction, g = 0.14, with significant effects particularly for empathy and mentalizing. That is a real but small effect, not a superpower [1].
So the unique value of books is not just “information.” It is also walking a long route, being taken somewhere you did not select, and staying inside language for a long time.
2. What can AI wall-bouncing replace?
An AI dialogue is different from asking one question and copying one answer.
If you ask “What causes the smell of rain?” and stop, that is close to search. But if you continue:
- Why does rain release the smell?
- What in the soil causes it?
- Microbes?
- Why are humans so sensitive to it?
- Why can we sometimes smell rain before it falls?
- What about other animals?
- What does the research actually show?
the topic becomes a network.
This resembles self-explanation and elaboration. Self-explanation means generating links such as “why does this happen?” and “how is this concept related to that one?” A 2018 meta-analysis combining 69 effect sizes from 64 reports found an overall learning effect of g = 0.55 for self-explanation prompts [2].
That is why active AI dialogue can be powerful: each question can update your mental map.
A book says, “Walk the route the author built.”
AI wall-bouncing says, “Build your own route while AI carries materials.”
Both can be learning. The cognitive jobs are simply different.
3. What is a schema? A map of knowledge inside your head
“Schema” sounds technical, but the idea is simple.
A schema is a mental map of how a part of the world usually works.
At first you may have only one dot: “interest rate.”
Later it connects to deposits, loans, inflation, central banks, investments, and currencies. Then one headline—“the policy rate rose”—automatically produces questions: “What about deposit rates? Mortgages? Stocks? The exchange rate?”
Research describes schemas as higher-level knowledge structures abstracted across multiple experiences. Schemas can help new information fit into memory, but they can also bias or distort memory when prior expectations are too strong [3].
So the feeling that “my neural network is connecting” is a decent metaphor, but it should not be taken literally as “one question instantly grows a physical wire.” A better description is: new information is being placed into an existing knowledge structure and linked to what is already there.
AI makes this loop extremely fast: notice a hole, ask immediately, fill it, notice the next hole.
4. Elaboration: if you want to remember, add lines, not just dots
The Japanese term used in learning science is seichika (精緻化), or elaboration. The word looks like a boss battle. The job is simple:
Make an idea richer and connect it to other ideas.
Instead of memorizing only:
“Geosmin = the smell associated with rain,”
connect it:
soil microbes make geosmin
→ raindrops can aerosolize compounds
→ humans can detect very small amounts
→ this helps explain why the smell becomes noticeable around rain.
With only A, you have one doorway into the memory. With B ← A → C and A → D, you have several routes back to A.
Questions such as “Why?”, “Give me an example,” “Is this like X?”, “What is the opposite case?”, and “How would this apply here?” add those routes.
But do not elaborate everything. If an ATM fee is 220 yen today, you probably do not need to construct the Grand Unified Theory of ATM Fees. Build networks for ideas you want to keep. Store temporary numbers outside your head.
5. Retrieval practice: conversation becomes a surprise quiz
Retrieval practice does not mean using Google. It means pulling information back out of your own memory.
You research something with AI on Monday. On Thursday, while talking to someone, you say:
“Wait, I read about that. It was something like…”
At that moment, your brain is searching without the screen.
A 2021 systematic review of classroom research screened nearly 2,000 abstracts and coded 50 experiments with 5,374 learners. Of 49 effect sizes, 57% showed medium or large benefits of retrieval practice, across different school levels, subjects, delays, and test formats [4].
Conversation adds another useful step:
learn with AI
→ explain to another person
→ they ask “why?”
→ you cannot explain one part
→ the hole becomes visible
→ you research it again.
That is an input → retrieval → gap detection → new input loop.
You do not necessarily need a nightly monk ritual called “Review Three Things I Learned Today.” Normal conversation can become a random test.
6. Creating articles can train you to build long logic instead of only following it
One traditional strength of reading is following a long chain of reasoning.
But what happens when you are on the production side?
The reader follows A → B → C → D.
The writer or editor decides that A, B, C, and D belong in that order.
So production can replace some “follow the logic” work with “build the logic” work.
A 2020 meta-analysis of 56 experiments in grades 1–12 found that writing about science, social studies, or mathematics content reliably improved learning, with an average effect size of 0.30 [5]. Writing can force selection, organization, connection, and re-explanation.
There is an important AI caveat. If you say, “Research everything, create the structure, write the article,” then never read a line, the AI performed most of the long-form construction.
But if the pre-article dialogue contains “No, that is wrong,” “Add this study,” “Connect these two ideas,” “What is the counterargument?”, then the human is doing substantial planning and editing.
It is entirely possible to read fewer books while doing a huge amount of pre-book-production thinking.
7. The AI trap: feeling capable because the AI is capable
This is the biggest warning.
AI can blur the difference between “the answer is correct” and “I can do this myself.”
In a large high-school mathematics field experiment published in PNAS in 2025, students using a general GPT-4-style interface (“GPT Base”) improved assisted practice performance by 48% relative to control. A learning-oriented “GPT Tutor” improved it by 127%. But when AI was removed for the exam, the GPT Base group scored 17% worse than the group that never had AI. The negative effect was largely eliminated in the Tutor condition, although that group did not beat control on the unaided exam [6].
In meme form:
AI on: “Wow, I am good at math.”
AI off: “Excuse me, who are you?”
A 2026 working paper gives a similar warning from chess. More than 200 chess-club students trained for 12 weeks. A system-regulated group received AI tips at key moments and improved by 64%; a group that could additionally request help whenever it wanted improved by 30%. The authors report that reduced “productive struggle”—effort on problems the student could plausibly solve—explained much of the gap [7]. Because this is a working paper revised in August 2026, replication and peer review still matter.
AI is not a demon that destroys learning. The same technology can behave like a tutor or like a copying machine, depending on when and how it gives answers.
8. Is it fine to outsource math? “Learning math” versus “using math”
Ask the purpose first.
If the goal is to become better at mathematics, having AI solve every step is weak practice. Generating the equation, making mistakes, and correcting them are part of the skill.
But suppose the real question is:
“If I put 1,000,000 yen in a bank at 0.2% versus 0.6%, what is the difference after ten years?”
Then the desired skill is not hand-powered compound-interest arithmetic. You need to judge:
- the actual difference in money,
- taxes,
- fees,
- inflation,
- liquidity,
- whether switching banks is worth the trouble.
You do not need to expand the compound-interest formula with paper and pencil every time.
You do not have to join the religion of ‘hand-calculate bank interest to train the brain.’ Use a calculator.
This does not mean mathematical concepts are useless. You still need enough understanding to notice nonsense. But when exact computation can be outsourced and personal mastery of the computation is not the goal, outsourcing is rational.
9. Mahjong: the problem where only the AI gains rank
Skill games such as mahjong show the same distinction.
Every hand:
“What should I discard?”
AI: “Five of characters.”
Human: “Ah, I see.”
If that happens constantly, you skip the cycle of reading the state, generating candidates, comparing them, deciding, seeing the result, and correcting your model.
The AI gains rank while the human becomes the employee whose job is to say “interesting.”
But that does not imply “never use AI.”
For many people, mahjong is a hobby. If the fun is becoming personally better at decision-making, practice matters. If the goal is simply to play, analyze with an AI, or enjoy the game without grinding hundreds of hands for a few percentage points, there is no moral duty to maximize skill.
When a hobby becomes:
“Today I will complete another repetitive drill to increase win rate by 2%,”
you are allowed to ask:
“Why did my hobby turn into homework?”
A skill is especially worth keeping in your own brain when doing it is itself enjoyable, when outsourcing is impossible, or when failure without the skill is costly. Otherwise, offloading may be perfectly fine.
10. I hate novels. Can’t I just watch the movie?
Partly, yes.
If your goal is story, characters, emotion, and a simulation of another life, film can deliver a lot of that. It also supplies faces, voices, rooms, movement, music, and timing.
A novel leaves more of that rendering job to the reader.
Text: “A man stood in a dim room.”
Brain: What room? What face? What voice? What atmosphere?
With film, the production studio kindly outsources much of your internal rendering.
The 2024 fiction meta-analysis found a small average experimental benefit, g = 0.14, with significant effects especially for empathy and mentalizing [1]. So “read novels or your humanity collapses” is not supported.
Novels still have distinct jobs: long exposure to written language, constructing scenes internally, and following an author’s language over many hours. Film has other strengths: fast, rich audiovisual information and acting.
If the goal is “experience a story,” film is a valid route. If the goal includes training sustained reading and internally generating a scene from text, novels retain a unique role.
11. Cognitive offloading: is moving work outside the brain bad?
Cognitive offloading means changing the task so that some mental work is handled by an external action or tool, reducing cognitive demand [8].
A reminder app.
A note.
A calculator.
GPS.
AI.
Humans have always done this.
We did not conclude that civilization ended when people stopped memorizing every phone number. The useful question is not “Did I outsource something?” It is “Did I outsource something I still need to perform independently?”
For example:
- multiplying interest rates → easy to outsource;
- deciding what should be compared → keep;
- proofreading → easy to outsource;
- sensing that a claim is suspicious → keep some independent judgment;
- optimal mahjong discard → depends on the goal of the hobby;
- emergency control of a machine → keep if failure would matter.
The AI-age question is not “How do I avoid AI?”
It is:
“Which tasks should leave my brain, and what should I do with the freed capacity?”
12. Does this question build a schema? The “What was the dog doing?” problem
More questions do not always mean more learning.
When Shohei Ohtani announced in December 2024 that he and his wife were expecting their first child, the announcement image also included his famous dog, Dekopin/Decoy [9]. On Japanese social media, exaggerated questions such as “And what was Dekopin doing at that moment?” became a joke format. This article treats it as a widely circulated meme rather than claiming one verified original post.
It is a perfect contrast.
Schema-building questions:
“Does A cause B?”
“Is this the same mechanism as C?”
“What is the exception?”
“Is this correlation or causation?”
Random-detail questions:
“When?”
“Where?”
“What exact minute?”
“And what was Dekopin doing?”
The second group can become clicking every blank field simply because it exists.
That is still fun. Not all curiosity has to maximize learning efficiency. The Dekopin department helps keep life entertaining.
But if your purpose is remembering and understanding, ask whether the question adds a useful connection, not merely whether it fills another slot.
13. What should remain in your brain in the AI era?
If AI keeps getting stronger, what should humans still learn?
Not every number. Not every calculation.
Keep at least these:
Goals. What are you trying to achieve? AI does not automatically choose a good life objective.
Basic concepts. Interest, probability, causation, risk, incentives—enough structure to understand outputs.
Question formation. What is this? Why? Compared with what? What would falsify it?
Suspicion. AI can sound confident while wrong. Important claims still deserve source checking.
Connection-making. Link new information to what you already know.
Judgment. Once the calculation is done, decide what action follows.
Even if AI becomes the calculator, searcher, and first-draft writer, the human still needs to decide what to calculate, which answer to trust, and what to do with it.
14. Practice: turn AI wall-bouncing into a learning device
You do not need a complicated new study routine. Change the dialogue.
If you want to remember the knowledge
Do not stop at “give me the answer.” Ask:
- Explain it for a middle-school student.
- Why?
- Give me a concrete example.
- Give me a case where it fails.
- How does it connect to X?
- What is actually certain in the research?
- Show me the primary source.
- Here is my explanation—what is wrong with it?
If you want to acquire the skill
Demote AI from driver to instructor.
In math: “Do not give the answer yet. Help me form the equation.”
In mahjong: choose a discard first, then ask AI to grade your choice.
If you do not need the skill
Outsource without guilt.
Bank-interest calculations, long aggregations, after-tax comparisons—if the real goal is decision-making, there is no need to worship manual arithmetic.
Then retrieve
When the topic appears in conversation, explain it without looking. The part you cannot explain becomes the next research target.
The loop is:
AI input → elaboration and links → retrieval in conversation → gap detection → AI again.
15. Summary: increase connections, not just information
Books are still powerful, especially for:
- encountering information you did not choose,
- following a long argument,
- prolonged exposure to written language,
- constructing people and worlds from text.
AI wall-bouncing is especially powerful for:
- asking immediately at the point of confusion,
- finding research,
- generating counterarguments,
- connecting ideas to prior knowledge,
- filling holes in a schema quickly.
A 2025 Harvard physics RCT found that a carefully designed AI tutor produced larger learning gains than in-class active learning while taking less median time: 49 minutes for AI versus about 60 minutes in class. The estimated effect size was 0.73–1.3 standard deviations in the authors’ ceiling-adjusted analysis. This was not “AI versus a one-on-one human tutor.” It was a crossover study in which each student experienced both a research-based active classroom condition and the AI tutor condition [10].
So “AI study is necessarily shallow” is no longer a serious general rule.
At the same time, math and chess research shows that AI can create AI-assisted performance without independent skill when it removes too much struggle [6][7].
The answer is therefore not “books or AI.”
For ideas you want to remember, use AI to add connections.
For skills you want to perform alone, attempt them yourself first.
For calculations you do not need to master, outsource them.
And sometimes let a book or another work take you somewhere you would never have searched for.
Your brain does not need to store every number.
It needs useful maps—and many good roads between the dots.
If you want to remember more, do not only add information. Add connections.
References (10)
- Wimmer, L. F., Currie, G., Friend, S., Wittwer, J., & Ferguson, H. J. (2024). “Cognitive effects and correlates of reading fiction: Two preregistered multilevel meta-analyses.” Journal of Experimental Psychology: General, 153(6), 1464–1488. DOI: 10.1037/xge0001583
- Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., & Winne, P. H. (2018). “Inducing Self-Explanation: a Meta-Analysis.” Educational Psychology Review, 30, 703–725. DOI: 10.1007/s10648-018-9434-x
- Gilboa, A., & Marlatte, H. (2017). “Neurobiology of Schemas and Schema-Mediated Memory.” Trends in Cognitive Sciences, 21(8), 618–631. DOI: 10.1016/j.tics.2017.04.013
- Agarwal, P. K., Nunes, L. D., & Blunt, J. R. (2021). “Retrieval Practice Consistently Benefits Student Learning: a Systematic Review of Applied Research in Schools and Classrooms.” Educational Psychology Review, 33, 1409–1453. DOI: 10.1007/s10648-021-09595-9
- Graham, S., Kiuhara, S. A., & MacKay, M. (2020). “The Effects of Writing on Learning in Science, Social Studies, and Mathematics: A Meta-Analysis.” Review of Educational Research, 90(2), 179–226. DOI: 10.3102/0034654320914744
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), e2422633122. DOI: 10.1073/pnas.2422633122
- Poulidis, S., Bastani, H., & Bastani, O. (working paper; posted 2025, revised 1 Aug 2026). “Self-Regulated AI Use Hinders Long-Term Learning.” Wharton School Research Paper / SSRN
- Risko, E. F., & Gilbert, S. J. (2016). “Cognitive Offloading.” Trends in Cognitive Sciences, 20(9), 676–688. DOI: 10.1016/j.tics.2016.07.002
- MLB.com (2024-12-28). “Baby Ohtani! Shohei, wife Mamiko expecting first child.
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports, 15, 17458. DOI: 10.1038/s41598-025-97652-6
