Half a year after accessing AI, an autism rehabilitation institution has dismissed its part-time supervisor.

By Liu Lingguo

Edited by Hai Ruojing

How long does it take to teach a four-and-a-half-year-old child what an apple is? For children with autism, this is an unknown question.

Huanhuan sits on a small bench, while the therapist holds up flashcards printed with apple patterns one by one in front of her: a red apple, a Red Delicious apple, a cut apple, a half-eaten apple, and a simple line drawing of an apple. She needs to identify and answer whether all of these are apples. A correct answer will be marked as "+ (plus)" by the therapist, and an incorrect answer will be marked as "- (minus)". For typical children, this relatively simple cognitive generalization ability can be acquired naturally, but some children with autism may need hundreds of practice sessions.

After a 30-minute class, the therapist will complete at least 30 question-and-answer rounds, each of which can be regarded as a piece of behavioral data. Huanhuan's mother told 36Kr that her child needs to take seven classes a day, five days a week, with almost no long breaks throughout the year. This means that a large amount of data with a clear corresponding relationship of "teaching-behavior-feedback" is formed in this process.

In the past, these data were mainly used for teachers to judge the completion status of a class. At present, they are used to train AI, and have become a new business.

"We turn the scarce supervision experience into standardized, scalable products for wider output." Wang Zhanwei, Vice President of R&D at Dami & Xiaomi, a rehabilitation enterprise for children with special needs, said that in the second half of 2023, as various foundation models were launched one after another, it became feasible to use the company's accumulated data from children's rehabilitation processes to train models, and embed them into institutions to support work such as assessment, class scheduling, and teaching plan generation. Subsequently, Dami & Xiaomi partnered with Amazon Web Services to customize an AI system. Currently, this system has been used in nearly 100 directly-operated stores and more than 300 partner institutions.

In Pingdingshan, Henan Province, a rehabilitation institution for children with special needs accessed AI in January this year. Six months later, Song Wentao, the person in charge, said that he had dismissed the previously hired part-time external supervisor, because AI can now take over part of the supervisor's work at a lower cost.

In autism rehabilitation institutions, therapists are mostly executors of supervision plans, who are responsible for teaching children directly; supervisors need to have years of frontline therapist work experience, and possess "high-level" capabilities such as assessing children's conditions and determining intervention courses.

It takes at least three to five years for a therapist to grow into a qualified supervisor. However, rehabilitation institutions lacking professional support cannot afford to wait that long. At present, AI is entering the long understaffed special rehabilitation track in the role of "supervisor".

Third- and Fourth-tier Cities Start to Pay for the Service

Huanhuan, who lives in Pingdingshan, Henan Province, was diagnosed with autism by the hospital at the age of three. Her family once considered sending her to Zhengzhou, the provincial capital, for rehabilitation training, but "it costs more than 10,000 yuan a month (after medical insurance reimbursement)". Huanhuan's mother mentioned that in addition to the class fees, plus the costs of renting a house, transportation, and full-time care, the family could not afford the total expense.

At the same time, she said that the top autism rehabilitation institutions in Zhengzhou are short of therapists and supervisors, and children can only arrange three to four classes a day. "We also have to find small additional institutions for extra classes, each of which costs 300 to 400 yuan. I reckon it would cost no less than 30,000 yuan a month." After weighing the options, Huanhuan finally stayed in Pingdingshan for rehabilitation training.

Staying was a choice forced by reality, but the family's demand for professional rehabilitation training remains strong. In fact, this demand has continued to rise in recent years. As the coverage of screening and assessment expands, more and more children with autism who were not identified in the past are entering the rehabilitation system.

"Since our outpatient department opened in 2019 up to now (2025), the number of visits to our specialized autism outpatient clinic has increased year by year, and now the annual outpatient volume has increased from about 300 visits to about 3,000 visits." Qi Yanjie, Associate Chief Physician of the Department of Child Psychiatry at Beijing Anding Hospital, once mentioned in an interview with the media.

In April 2026, the Chinese Center for Disease Control and Prevention cited screening data from the National Health Commission, stating that the prevalence of childhood autism in China is about 7‰, and the total number of people living with autism exceeds 14 million.

This demand is transmitted to the supply side, prompting rehabilitation institutions to urgently hope to fill the manpower gap and improve service quality. But professional competence is built on experience, which "cannot be rushed", so the talent gap in the industry has long been difficult to fill. "One therapist teaches 4 to 6 children on average every day, and one supervisor may supervise 50 to 60 cases." Zheng Tian, a senior supervisor at Dami & Xiaomi, mentioned.

Before introducing AI, Song Wentao once hired a part-time external supervisor holding a BCBA (Board Certified Behavior Analyst) certificate. According to his recollection, the external supervisor provided 1-2 online teaching sessions per month, with an annual employment cost of more than 100,000 yuan. However, due to the limited teaching time, the number of covered cases was insufficient, and it was impossible to continuously track the changes of individual children, so the effect was not satisfactory.

The "supplementary role" of AI comes at just the right time.

Filling the gap is one aspect, and the other significance is to make the services of institutions more differentiated. Song Wentao introduced that in third- and fourth-tier cities such as Pingdingshan, restricted by the supply of professional talents, the service quality of various children's rehabilitation institutions is similar, and they used to compete mainly by cutting prices. After introducing AI, Song Wentao added AI courses priced at 80 yuan per class on the basis of the original 40-yuan and 60-yuan classes, and currently about 30% of parents have signed up for them.

In the past two years, Song Wentao's institutions have expanded to ten. He said that all institutions will be equipped with "AI supervisors" later, to help new therapists get started quickly, standardize teaching, and make the service quality of each store more consistent and stable.

Reducing costs, increasing customer unit price, widening service gaps, and facilitating standardized expansion together form the admission ticket for AI to enter rehabilitation institutions.

Continuing down the payment chain, parents are the final payers. Parents' willingness to pay mainly depends on whether AI can bring substantial rehabilitation progress to their children. At present, according to a controlled study provided by Dami & Xiaomi, after 65 children from three third-party institutions received three months of intervention, the children in the AI intervention group performed better than those in the ordinary intervention group in five areas including social interaction, language communication, cognition, and imitation.

How do these changes happen?

AI "Understands" Children Earlier

The first meeting between AI and the child takes place during the assessment.

Assessment is similar to a "physical examination" for children's ability development. The difference from hospital diagnosis is that it does not end with a diagnostic conclusion, but figures out exactly what the child can do, what they cannot do, and why they cannot do it. The purpose is to guide the next step of intervention.

Zheng Tian explained that a child who "cannot speak" may be due to insufficient communication willingness, inability to imitate mouth movements, inability to understand instructions, or being able to understand but not knowing how to express themselves. Only after a detailed assessment can we "prescribe the right medicine".

Previously, therapists usually spent several weeks testing each of the child's abilities one by one, and then the supervisor analyzed the behavioral data and manually completed a report as long as more than ten pages. Now, with the participation of AI, the report writing time alone has been shortened from at least three hours to a few seconds: the therapist enters the child's feedback into the tablet synchronously during the assessment, and the report comes out when the assessment is over, and the supervisor only needs to review and improve it on this basis.

Huanhuan's mother mentioned that through AI assessment, it was found that Huanhuan, who is four and a half years old chronologically, only reached the developmental level of a three-and-a-half-year-old. This one-year gap is what Huanhuan needs to catch up with through rehabilitation training.

Wang Zhanwei explained that AI will start from the child's actual age, test whether the child can complete the projects corresponding to the corresponding stage; if the child cannot complete them, AI will search down level by level until the ability baseline is determined, and then formulate the child's IEP (Individualized Education Program for Children with Special Needs) accordingly.

There are similar explorations overseas. In 2024, the team of American rehabilitation technology company TheraDriver used the assessment forms, diagnostic reports, parent interviews and other materials of 617 children with autism aged 2 to 14 to train AI to identify behaviors that need intervention and skills to be learned, and generate personalized treatment plans. Subsequently, 6 child cases were randomly selected for testing, and it was found that the AI-generated plans were highly close to those made by supervisors.

But the IEP is more like a to-do list of "skill points", and the specific order of learning what first and what later is also critical. A "perfect" class schedule must follow the law of children's development, and take into account multiple factors such as the child's long-term and short-term goals, weaknesses, and advantages, which is obviously very difficult. "Even if the child comes for three hours a day, five days a week, it is impossible to require the child to master dozens or even hundreds of skills in the IEP all at once. AI can tell you which of these skills should be taught first, and how many rounds of practice are recommended per day for the child to master them as soon as possible."

In operation, the therapist can directly enter basic information such as the child's class frequency and training duration in the dialog box, as well as personal preferences such as liking toy cars, blowing bubbles, tickling, and being lifted high, and even "off-site factors" such as traits that parents focus on. After comprehensive analysis by AI, a rehabilitation plan will be issued.

After the formal training course starts, the generation of personalized teaching plans and teaching aids is even more the "absolute domain" of AI. Still taking "recognizing apples" as an example, children with autism may have just learned to recognize apples on trees, but then they cannot recognize picked apples. In order to confirm that the child has truly mastered a concept, in the past, teachers needed to use search engines to obtain different images, then print, cut, laminate and organize them. Now, AI makes the acquisition of examples extremely convenient, and all kinds of apples can be obtained with one click.

However, this is also a function that general large models can fully realize. In this regard, Wang Zhanwei explained the difference: the general large model provides general and universal examples, while the Dami & Xiaomi AI always gives a "customized example" exclusive to this child based on all the child's past intervention conditions and professional database information. For example, AI will judge which change makes the child suddenly unable to recognize the apple again, and provide targeted training.

Similarly, AI can also customize picture books according to the child's photos and rehabilitation goals; for children who are afraid of unfamiliar environments, teachers can integrate the institution's gate, classrooms and toilets into stories to help them get familiar with the environment in advance; in group classes, AI can continuously design new games according to the number of people, ability base and training goals.

The rehabilitation cycle of children with autism is often measured in years. After each class, the child may have subtle changes. These tiny, imperceptible changes are captured by AI, which will continuously strengthen its "understanding" of the child; then, like a "real-time positioning" navigation, it tells the therapist what to do next. In addition, even if the therapist leaves the job, the tracking records of the child have been completely retained by AI.

Another group that badly needs "navigation" is parents. Since children with autism receiving intervention are generally young and not good at communication, parents used to lack sufficient understanding of their children's specific conditions. At the end of last year, after signing Huanhuan up for AI-involved courses, Huanhuan's mother mentioned that the most intuitive feeling was that "the goals are clearer". For example, parents can see on the tablet that the child has an 80% mastery of the pronunciation of the initial consonant "k" and other information.

Back to the Human World

For this industry that has long relied on manpower and experience, AI has proven that it can assist in undertaking multiple tasks. But this is a fill from scratch under the premise of insufficient supply of professional talents, which does not mean that it is necessarily better in all aspects than professional supervisors who have rich contact and interaction experience with children.

"AI can provide rich interactive games and activity ideas, allowing teachers to further optimize their teaching methods on this basis." Zheng Tian mentioned, but a set of carefully polished courses by supervisors may not lose to AI in terms of innovation and on-site adaptation. After all, there is always a "time lag" between AI and the child, and the therapist is the one who interacts with the child in real life and builds emotional connections. This also fits the goal of rehabilitation itself: to allow children to return to the "human world" more freely and comfortably.

At present, the more direct value of AI is still to improve the work efficiency of therapists and supervisors, acting as a "super assistant". With less desk work, supervisors who are "in charge of multiple children" can squeeze out more time to accompany children who have difficulty adapting to classes, guide therapists' teaching methods in person, and soothe parents who are anxious because their children have not made obvious progress recently, so that human professional abilities can return to the most irreplaceable links.

The upper limit of this "super assistant" depends on the continuous "new data feeding" by therapists. When Zheng Tian trained small and medium-sized institutions, she found that because many therapists did not have the habit of recording round by round before, even if the institution introduced AI, it was difficult to achieve complete data entry. In the long run, AI may "lose track" of the target.

Then, with sufficient data, can AI really replace therapists to make judgments?

At present, the answer is still no. Wang Zhanwei said that all content generated by AI will be manually reviewed and checked before entering the actual process. Once, AI judged that Huanhuan had mastered a certain training item and could move on to the next goal, but the therapist thought she still needed to continue consolidating.

Huanhuan's mother said that she feels that AI does not really sit opposite the child and have eye contact, language and emotional communication with the child, so she trusts the teacher's judgment more, "AI contributes a little, and the supervisor teacher adds a little more, I think that's even better."

Daily life such as eating, sleeping, playing games, and traveling is also a natural "training ground" for children. At present, many companies are developing targeted home intervention products. For example, let the child use a tablet to have a question-and-answer interaction with AI, practice recognizing objects from pictures repeatedly, and learn social skills in simulated scenarios such as greeting, borrowing things, and taking turns playing games; embed AI into plush smart robots, so that when children cry, are irritable or depressed, the robot will actively initiate communication and give timely emotional support; let parents use the APP to directly ask AI for advice on tricky problems online, and carry out home training by themselves according to the guidance.

Huanhuan's mother said that she is very willing to try such products, but her main concern is that when she gives Huanhuan classes at home, "she ignores me at all, doesn't think I am a teacher at all, and just thinks I am still her mother." AI entering autism rehabilitation can be the "teacher of teachers" and the "teacher of parents", but it is still difficult to replace humans in building real relationships with children.

In the future, robots may also sit in front of children, hold up a small flashcard, and ask "Is this an apple?" But will the child be willing to look up, face the gaze, and give an answer? This is still unknown.

(At the request of the interviewee, Huanhuan is a pseudonym in the article.)