From “Summer Companion” Workbooks to Generative AI: What 825 Parents Reveal About Learning to Learn

Cicadas call beyond the window. A mathematics workbook sits to the left, manuscript paper in the center, and a tablet to the right. The cursor blinks in a chat box. Type “make an outline for my book report,” and paragraphs arrive in seconds. Ask why an insect moves at night, and the screen can produce something that looks like a hypothesis for an independent-research project. The assignment may be easier to finish. But whose thinking has it become?

A survey released August 19 by Duolingo catches this new summer desk through parents’ eyes. The company collected online responses on August 3 and 4 from 825 parents in their 30s through 50s who have children in elementary or junior-high school. Some 83.2% said it was important for children to choose and pursue learning beyond the homework assigned by school. Before the break, 75.8% had hoped their children would do such additional study.

More options did not automatically produce more autonomy. The most commonly reported difficulty was that motivation or concentration did not last, at 32.5%. Forming a continuing habit followed at 28.6%; 27.3% said the child did not begin independently. The household bottleneck was less a shortage of content than the work of starting and continuing.

What the survey actually found

Among the 625 parents who had expected additional learning, 71.2% said the child had met or exceeded the expectation. In 447 households where expectations were met or exceeded, “it is a daily habit” was the most common reason at 29.1%, just ahead of “the child has a goal or something they want to learn” at 26.6%. This suggests that a dependable time and place may be more realistic than demanding a fresh act of will every morning.

It does not show that the habit caused success. A child already eager to study may form habits more easily. Parental time, income, the child’s age, tutoring and school policy could affect both the routine and the outcome. The survey detects associations, not the causal effect that a controlled or carefully adjusted study would estimate.

In 1924, summer already came with “do this every day”

It is tempting to imagine a pre-digital summer in which children roamed outdoors and freely invented their own investigations. Archival records complicate that nostalgia. On July 11, 1924, the Education Ministry’s school-health section asked Tokyo Prefecture to examine six elementary schools that assigned vacation homework. The inquiry classified review work, school-set tasks, summer practice books and diaries, the time required by subject, and special cautions given to students.

Itabashi Jinjo Higher Elementary School reported using vacation practice books for every child, encouraging students from third grade upward to read in the school library, and bringing pupils back twice to check health and progress. The school estimated 45 minutes of daily homework in first grade and two hours in fifth and sixth grades and the higher course. A mimeographed notice from Hachioji Third Jinjo Elementary told second graders to review in the cool morning, read and write every day, and complete arithmetic in regular portions.

The three problems in the 2026 survey—beginning, continuing and building a habit—were therefore already central to school instructions a century ago. The paper practice book has become a tablet; the teacher’s stencil has become an app notification. The contradiction remains: how does an adult externally produce a child who learns from within?

Postwar “free research” and the paradox of assigned freedom

The 1947 general curriculum draft introduced jiyū kenkyū—literally “free research”—as part of the new course of study. Its explanation began from a child’s activity in an ordinary subject. Interest might generate a further activity that could not fit inside the scheduled class. Free-research time was meant to extend that interest with guidance when needed. Instrumental music, calligraphy, crafts, science experiments and painting were among the examples. The document warned against using the period to make every child learn the same thing.

That formal subject is not identical to the summer project familiar today. Yet its principle remains useful. A question should grow from what a child notices, not arrive fully formed from a teacher, parent or AI. Guidance is not forbidden, but it should not outrun and replace the child’s interest. The hardest part of a required “free” project is not unlimited freedom. It is asking all children to deliver individualized work by one deadline and often judging the polish of the object. Household support inevitably enters the product.

Parents have long selected themes, bought materials, corrected prose and improved display boards. Generative AI can be understood as a faster, cheaper and always-available version of that older adult substitution. The problem is not uniquely artificial intelligence. Assignment design makes substitution attractive when it values the finished display but ignores why the question was chosen, what failed and how the student changed direction.

A book report is about the reader, not just the book

Japan’s National Youth Book Report Contest began in 1955. The Japan School Library Association and the Mainichi newspapers run it through schools: work passes from school judging to district, prefectural and national levels. Its longevity helped make the dokusho kansōbun a signature form of summer work, although not every assigned report is entered in the contest.

A book report appears perfectly suited to generative AI. A model can summarize a plot, extract themes and produce a smooth introduction, body and conclusion. But the core of a response is not generic information about a title. It is the place where a scene disturbs the reader’s memory or values. That fact exists in the student’s life, not in a language model. An AI can ghostwrite plausible emotion, but fluency cannot fill the resulting emptiness.

The question for an AI-age submission is not simply whether the prose flows. Which observations came from the student? Where did a tool help? What was checked? What did the student finally change?

GIGA School made the screen ordinary

Generative AI did not fall suddenly into a paper-only classroom. The GIGA School program, launched in fiscal 2019, sought a device for every compulsory-school student and high-speed networks. COVID-19 school disruption accelerated deployment, and by fiscal 2021 most compulsory schools had begun using one device per child.

A device brings searching, photography, collaborative writing, visual explanation and adaptive exercises into one place. When it travels home, the boundary between school infrastructure and summer learning becomes thinner. In the Duolingo survey, 34.9% of parents said children spent more learning time with digital devices than the parents had, while 28.7% said the choice of learning methods had grown. Those impressions sit on top of the GIGA infrastructure.

MEXT has repeatedly cautioned that distributing devices does not itself guarantee educational effects. Turning a paper drill into a PDF only relocates repetition to a screen. If a child measures neighborhood temperatures, photographs shade and pavement, shares data and revises a hypothesis, the same device can expand an investigation. Technology’s effect depends on the intellectual action an assignment requires.

Why 47.1% is not the same as the government’s figures

In the Duolingo survey, 9.9% of parents said their child “often” used generative AI in ordinary learning and 37.2% said “sometimes,” producing the eye-catching 47.1%. The Children and Families Agency’s fiscal 2025 survey, released in February 2026, produced very different self-reported figures among internet-using youth: 8.6% for elementary students age 10 and older, 30.8% for junior-high students, and 46.2% for high-school students.

The gap does not prove that either survey is wrong. One asks parents and the other students; one narrows attention to learning and the other lists internet activities; one combines two school levels while the other separates them; sampling differs. A parent may include an AI-assisted search or a feature inside a learning service. A child may use an AI feature without recognizing it as generative AI. The defensible conclusion is that use changes sharply by school level and definition. There is no single context-free “AI use rate for Japanese children.”

Parents prefer an AI that makes a child think

Among the 389 children reported to use generative AI, the single best-fitting use was “ask for explanations of unclear parts” at 16.2%, followed by “explore an interest or expand learning” at 15.7%. Ten percent were said to use an AI answer or text directly. Another 8.5% of parents did not know how the child used it.

Across all 825 parents, 32.4% worried that copying an answer would reduce opportunities to think; 27.9% feared overdependence would weaken independent study; 26.9% worried that the child could not judge correctness. Their preferred uses followed the same distinction. Explaining why an answer was wrong or how to reason led at 23.3%, and giving a hint without the answer followed at 20.4%. The central divide is less pro-AI versus anti-AI than substitution versus scaffolding.

MEXT’s Version 2.0 guidance similarly treats generative AI as a tool that may support and extend human ability, not as the purpose of learning. An output is one reference; a person makes the final judgment and accepts responsibility for the resulting work. Schools must also consider age limits and parental consent, personal information, copyright, bias and false output. The 2023 interim guidance had already said that submitting generated material as one’s own work for a vacation assignment such as a book report could be inappropriate or dishonest.

The 88.9% family figure is not a magic causal recipe

One of the survey’s largest contrasts concerns family learning. In households where summer study exceeded expectations, 88.9% said the family learned together at least weekly. Among those far below expectations, the share was 31.6%, while 48.1% never learned together. In the exceeding group, 32.1% made plans or goals together, 30.2% selected materials together and 27.2% agreed on study time or rules. In the far-below group, a prominent response was simply telling the child to study, at 51.9%.

It would be a mistake to translate that into “sit beside the child and success follows.” A child progressing well may make participation pleasant and feasible. Parents with more discretionary time or money may be overrepresented; younger students may receive more help. The release does not publish adjustments for household income, parental work hours or detailed school stage.

Still, the difference between accompaniment and command is consistent with other evidence. A 2026 longitudinal study using the Japanese Longitudinal Study of Children and Parents found that autonomy-supportive involvement benefited children’s vocabulary, while parental control and direct instruction could hinder it. That study used different outcomes and is not a test of summer homework. It does reinforce a broader point: the method of involvement matters, not only its amount.

AI can widen access—and amplify household differences

A generative AI can answer at night, restate a difficult explanation and make practice items. A household without specialist subject knowledge may gain a first rung on the ladder. In that sense, access to support can widen. Safe use, however, requires someone to read terms, avoid personal data, suspect mistakes, locate a source and understand school rules. Adult time and digital literacy matter again.

Paid features, reliable connectivity, a quiet desk, library access and opportunities for real-world experiences are also unequal. AI may resemble a free tutor, yet results differ depending on whether a household has an editor who can help frame a good question, detect a wrong answer and avoid giving too much away. Technology can lower the price of information without erasing inequality in attention, time and judgment.

If a school permits AI, “use it appropriately at home” is not an adequate equity policy. Schools need to provide a common safe environment and basic AI literacy, preserve a non-AI route, and design work whose evaluation does not depend on which subscription a family can afford. Fairness also means not outsourcing all supervision to parents.

In the AI age, submit the process as well as the product

Combining notes, observation photos, drafts, conversation and a brief oral explanation is more educational than trying to police every submission with an AI detector. It also avoids treating a probabilistic detector as proof. Connect a project to neighborhood temperature, household electricity, local plants or a grandparent’s memory, and an AI may assist—but cannot substitute for the student’s access to reality.

Assessment must move beyond polish. A project whose first hypothesis failed but whose author measured again and explained why may contain more learning than flawless prose delivered on the first try. If failure only loses points, a child has an incentive to ask AI for a smooth success story. If revision earns credit, hiding the process becomes less attractive.

Seven agreements for a household

A parent’s most productive question is not only “Are you finished?” Try: “What surprised you most?” “Where was the AI wrong?” “How did you check that number?” “What would you change tomorrow?” The adult does not need to know the answer. Listening to an explanation and asking a child to make judgment visible can itself be accompaniment.

The useful friction to preserve

The Duolingo survey does not prove that digital study caused independent research and book reports to decline. It compares parents’ recollections with their current children, finding the traditional forms more than 20 percentage points lower and app/tablet use at 23.4%. Changes in school homework policy, age, region, household conditions and the pandemic may all contribute.

The sponsor is also a language-learning app company, and the latter part of its release promotes features designed to build habits on its own service. That does not invalidate every response. It does require readers to distinguish a company’s commercial interest from independent academic evidence and to notice the wording, sample and unpublished cross-tabulations.

Even with those limits, the parental dilemma is recognizable. Adults want children to learn by themselves. The child may not start. The routine may collapse. A screen may tempt the child into working—but the screen may also do the thinking. A century ago, a teacher printed instructions to work every day. Postwar educators tried to create time for individual interest. In 2026, families face the same question between notifications and conversational AI.

Education contains useful friction. A student waits for the insect to appear, rereads a chapter, struggles to phrase a thought, locates the line where a calculation went wrong, explains an idea aloud and discovers it was not understood. That friction is not merely inefficiency. It is the time in which information becomes the learner’s knowledge.

Generative AI can remove that time or support it. If it explains a difficult step, offers a counterexample or suggests the next observation, it can be a good companion at the desk. If it fills the space from title to conclusion and the result is submitted as the child’s own, the homework ends before the learning begins. The new summer question is not simply “Was AI used?” It is how much of the thinking between the first question and the final judgment still belongs to the child.

Survey and policy at a glance