AI writes the plan, teachers make the decision: a small Japanese company tests human control in special education

A special-education plan writes a child’s year in the future tense. What can they do? Where do they struggle? Who will help? What kind of life do they want after school? Done carefully, the document is a map connecting classroom, home, health care and welfare. Copied under pressure from last year, it can erase growth, new anxieties and the student’s own voice.

Soshi Sozo Research Institute, a small company in Fukuoka’s Chuo Ward, wants to put conversational AI into that gap. Founded about a year ago, it says teachers can speak or type through a dialogue instead of filling dozens of fields. The system then drafts Japan’s two central individualized planning documents. Teachers may stop the interview when they have enough information. AI-written passages are marked, a review box must be checked before finalization, and the original generation, teacher edits and cited records go into an audit log.

This sounds like modest workflow software. It sits, however, at the leading edge of a 150-year change in Japan’s question—from where to place children with disabilities to what each child needs. Whether AI deepens that question or merely industrializes elegant paperwork will depend less on speed than governance.

Access to school was not always guaranteed

The Kyoto school for blind and deaf pupils opened in 1878. A 1923 ordinance created a national framework for such schools. The 1947 School Education Act formally established schools and classes for children with disabilities. Yet nationwide compulsory schooling in what were then called schools for “handicapped” children did not begin until 1979. Before that, postponement and exemption kept some children outside school.

Creating a guaranteed seat was a profound achievement. The same system also sorted children by disability type and severity into “special places.” Postwar history therefore carries two movements at once: ending exclusion from education and challenging the boundaries built by separation.

In 2007, the question moved from place to need

In April 2007, Japan’s education ministry instructed schools to promote “special needs education.” The earlier system centered instruction in special settings according to disability category and severity. The new approach said schools should identify individual educational needs and provide support to improve learning and daily-life difficulties, including for pupils with developmental disabilities in regular classrooms.

Japan signed the UN Convention on the Rights of Persons with Disabilities that year and ratified it in 2014. Article 24 calls for non-exclusion from the general education system, reasonable accommodation and individualized support. A plan is therefore not merely an administrative form. It can be an instrument of access, participation and continuity across schools and services.

Two similar names, two different clocks

The Individual Education Support Plan—a descriptive translation—is the long map. It connects education, health, welfare and employment across life stages, recording the student’s and family’s wishes, reasonable accommodations, transitions and interagency support. The Individual Instruction Plan is the nearer itinerary: concrete school- or term-level goals, content, method and evaluation. The 2017 curriculum revision required both to be created and used for each child receiving resource-room instruction.

The company uses the English labels ESP and IIP. Readers should not assume those are nationally standardized official acronyms. What matters is the relationship: a short-term teaching target should advance, rather than contradict, the longer plan.

About 680,000 children—and many different classrooms

Education-ministry data put the compulsory-age population receiving special support at roughly 680,000 in 2024, or 7.3% of all pupils: about 87,000 in special schools, 395,000 in special classes and around 200,000 receiving resource-room instruction. Official resource-room totals vary between roughly 196,000 and 201,000 because source documents use different years and counting methods. Rising totals reflect access, identification, enrollment and methodology; they do not by themselves prove that disability is becoming more common.

A 2022 teacher questionnaire estimated that 8.8% of elementary and junior-high students in regular classes showed marked learning or behavioral difficulty. That was neither a diagnosis nor an individual clinical assessment. It does show why planning can no longer belong only to a small specialist office.

A tiny company chose work that wins no applause

Soshi Sozo Research Institute was established July 25, 2025, led by Yoshimasa Sugimori. It lists management support and software development; public materials do not disclose capital, head count or revenue. In May it expanded Sukusuku toward special classes and resource rooms, with role-based sharing among homeroom teachers, specialists and administrators.

For the long support plan, the new feature asks about student and guardian wishes, strengths, participation and environment, agencies, reasonable accommodation, transition, long-term goals and consent. For the instruction plan, it traverses the previous evaluation, current state, Japan’s six domains and 27 items of independent activities, alignment with the higher plan, short-term goals, interventions, generalization, evaluation and family cooperation. Users may attach photographs, video and audio for transcription or image analysis.

Forms are fast when the author already knows the answer. Dialogue can give sequence to someone still learning what to observe. A good question does not replace professional judgment; it helps a teacher notice a relationship. The small company’s choice to attack unwritten, unglamorous documentation time has a practical logic.

Can ICF rescue the child from the diagnosis?

The company says its questions draw on the World Health Organization’s International Classification of Functioning, Disability and Health. ICF places body functions alongside activity, participation and environmental factors. The same difficulty with reading can become a different educational experience depending on voice input, teaching material, peers, attitudes and classroom noise. The frame shifts attention from “fixing the child” alone to the fit between person and environment.

That is promising, but ICF is a classification, not an oracle that chooses the right intervention. WHO practical guidance stresses that child and parent views are especially important in educational planning and may differ from professional views. An AI can sprinkle ICF language over a document and still produce an ornate deficit inventory.

What is measurable is not automatically valuable

Sukusuku is designed to prompt SMART short-term goals—specific, measurable, achievable, relevant and time-bound. Evaluation matters. “Be calmer” is less actionable than a goal describing how a student can choose a break card during noise and return when ready.

But measurability is not value. A system optimized for minutes seated or number of utterances can push autonomy, safety, friendship and self-determination into the background. Japan’s six domains and 27 items are a curriculum menu, not a diagnostic machine. Generative systems may favor behaviors that are easy to count rather than experiences that matter.

Color, confirmation and a diff: useful friction

Several announced choices deserve credit. AI passages are visually marked. A teacher must tick “I reviewed AI-generated content” before finalization. The original output, editing diff and source records remain. The manual form remains available. The system carries forward the annual goal and prior evaluation, then suggests continuation, higher difficulty or review.

The diff is especially valuable: reviewers can see what the machine proposed and what the professional changed. It can support learning, not merely blame. Yet the company has not said whether logs are tamper-resistant, who may view them, how long they live, whether families can obtain them, or whether an old output can be reconstructed after a model update.

“Human in the loop” is not the same as human control

Japan’s education-ministry guidance on generative AI assigns final judgment and responsibility to humans and warns about error, bias, confidentiality and personal data. But if a teacher only clicks at the end, a process can be formally human-in-the-loop and substantively automated. Automation bias—the tendency to trust a confident system—grows dangerous when prose is polished and time is scarce.

Meaningful oversight requires an easy path to reject suggestions, nearby evidence, uncertainty signals, alternatives, error reporting and enough time to think. Schools should measure not approval rate but what teachers changed, whether student and family dialogue increased, and whether plan quality improved.

The teacher is not the only human who matters

Putting the teacher above the algorithm is necessary, not sufficient. The long support plan centers student and guardian wishes and accompanies the child across school stages. Ministry guidance calls for care and student or guardian understanding and consent when information is shared. Inclusion under the disability convention is not simply a benevolent professional choosing well.

A child who does not use speech, or needs time to communicate, can participate through choices, pictures, assistive technology and supported observation. If the AI interviews only teachers, the student becomes an object described again. The announcement does not explain student-facing summaries, parent access, consent workflow, correction or appeal.

Photographs, voices and video are not ordinary attachments

Disability, medical history and health records may constitute Japan’s “special care-required personal information.” Classroom video may capture other children’s faces, voices, behavior and names. Audio may record family circumstances; backgrounds reveal location. Because a plan follows a child for years, an error can stick to the next decision.

The company claims TLS 1.3, field-level database encryption, role controls and separate tenants for each school operator—useful safeguards. A claim of alignment with ministry security policy is not a substitute for independent audit. Before procurement, a school must learn purpose, minimization, consent, whether data trains models, residency, subprocessors, retention, deletion and breach response.

An audit trail can become another exposure

A log makes drafting visible, but permanent storage of erroneous diagnostic speculation can magnify harm after a breach. Access should be granular; retention tied to purpose; corrections unmistakable; deletion requests balanced against lawful recordkeeping.

Model updates also matter. Reproduction requires the model and prompt versions, timestamp, settings, references and system configuration. For a young vendor, buyers should require data portability, a standard export, business-continuity terms and a clear path if the service closes.

Where does saved time go?

The company estimates that work formerly taking 60–120 minutes could take 25–45. It also describes 10–20 minutes of dialogue and roughly 10 minutes of review and correction; that component arithmetic does not plainly match the stated total. No sample, comparator, median or quality score is disclosed. “Designed to reduce” is therefore more accurate than “reduced.”

Still, amid persistent teacher workload, returning half an hour would matter. The test is whether it goes to observing the child, talking with student and family and making material—or is absorbed by another compliance demand. Success should include dialogue time, goal quality, student assent and teacher burden, not minutes alone.

Questions a school should ask before buying

Can AI hold the pen lightly?

“Copy last year” is not a story about lazy teachers. It is the structural result of asking for individualization with limited time, staff and training. Soshi Sozo Research Institute chose a job no one finds glamorous and offered concrete friction: ordered questions, visible AI text, diffs and a manual route. Its small size may let it return classroom feedback to the product quickly.

But good intentions and good interface choices do not guarantee rights. “Teachers decide” becomes real only when student and family are co-decision-makers, the entire data life cycle is governed, errors can be corrected, versions are traceable and independent pilots are published.

The center of the plan is neither AI nor teacher. It is a child who is changing now, who has hopes and who may see the world differently from every professional in the room. AI should not impersonate that voice. Its best role is quieter: offering the next good question so humans do not fail to hear it.