Describe Your Child to a Chatbot, and It Will Almost Always Find a Diagnosis
Researchers gave three AI models the stories of 220 children at a psychiatry clinic. With the child's own words the AI missed most problems. With the parent's words it found one in almost everyone, including the children who had none.
It is eleven at night. Your nine-year-old has finally fallen asleep after another evening of tears over homework, and you are sitting with your phone. You open a chatbot and type three sentences about what worries you. Then you ask: what is this?
A lot of parents have done some version of this. Now a group of child psychiatrists has tested what comes back.
They took 220 real children on their first visit to a clinic and gave their stories to three of the best-known AI models. The same child, told in four different ways. The result says less about how clever the machines are and more about something else: the answer depended on who was telling the story.
One question, written down word for word
The study took place at two university hospitals in Turkey in the spring and summer of 2026. Every child who came to the child psychiatry clinic for the first time was invited, from toddlers to seventeen-year-olds. The average age was nine.
Before anything else happened, the parent was asked one open question: What is the complaint that brought you here today? The child was asked separately: Why did you come here today? What bothers you the most?
The answers were written down as close to word for word as possible. Then the child psychiatrist did what child psychiatrists do: a full assessment, a talk with the family, an examination, and in the end a diagnosis, or no diagnosis.
Afterwards, the researchers gave each case to the AI models four times. Once with only what the child had said. Once with only what the parent had said. Once with both. And once with both plus the doctor's own notes, with every diagnosis word removed.
The first thing to notice is how little there was to go on.
- The child's own words9 words
- The parent's words32 words
- Child and parent together46 words
- Child, parent and the clinician's notes122 words
Nine words. That was the typical length of what a child said about why they were there. Some said nothing at all. The parents said about three times as much, which is still shorter than this paragraph.
The same child, four different answers
The models were asked to do two things: first decide whether this sounded like a diagnosis at all, and if so, list the three most likely ones.
- The child's own words36–44%
- The parent's words66–69%
- Child and parent together69–72%
- Child, parent and the clinician's notes79–83%
With only the child's words, the AI's first suggestion matched the doctor in roughly four cases out of ten. With the parent's words, it was close to seven out of ten. With the doctor's notes added, around eight out of ten.
Two details are worth stopping at.
Adding the child's words to the parent's made almost no difference. Statistically, the models did no better with both than with the parent alone.
It barely mattered which AI was used. The three models were OpenAI's GPT-5.6, Google's Gemini 3.5 Flash and Anthropic's Claude Sonnet 4.6. They differed clearly only when they had the child's words alone. As soon as a parent was in the picture, the differences were too small to tell apart from chance. What went in mattered far more than which machine it went into.
Two ways of being wrong
This is where the study gets interesting. A doctor who is "right 70 per cent of the time" can be wrong in two very different ways: by missing a child who is struggling, or by putting a label on a child who is fine. The researchers looked at both.
The child had a diagnosisHow often the AI said that something was there. 199 children.
- The child's own words41–59%
- The parent's words94–98%
- Child and parent together94–99%
- Child, parent and the clinician's notes93–98%
The child had no diagnosisHow often the AI said that nothing was there. 21 children.
- The child's own words62–86%
- The parent's words24–38%
- Child and parent together14–43%
- Child, parent and the clinician's notes52–62%
When the AI only heard the child, it was cautious to a fault. In roughly half of all cases it answered that nothing seemed to be wrong. But 199 of the 220 children did have a diagnosis. The AI missed about half of them.
When the AI heard the parent, it flipped. Now it found something in almost everyone. It caught more than nine in ten of the children who had a diagnosis.
It also "caught" most of the children who did not.
Think about what that means. A worried parent describes a child. The doctor, after a full assessment, concludes that this is within the normal range. The AI, reading the same parent's words, suggested a diagnosis for most of those children.
That is not as strange as it sounds. A parent who has made an appointment at a psychiatry clinic is describing a problem. That is why they are there. The words are full of worry, and the AI answers the words. It has not seen the child play, heard the child talk, or asked the follow-up question that makes the worry smaller.
Why nine words are not enough
Why was the child's own account so much less useful?
Partly because children say little. But it also depended on what was wrong. From the child's words alone, the models did fairly well with anxiety and obsessive-compulsive problems. Those are things that happen on the inside, and a child can sometimes say them: I am scared all the time. I have to check the door.
ADHD, autism and language difficulties were different. A seven-year-old does not say "I am inattentive in several settings." The people around the child see it. For those diagnoses, the AI needed the parent.
Age mattered too, in the way you would expect. For each year older the child was, the odds that the AI got it right from the child's own words rose by about 16 per cent. A sixteen-year-old can describe their inner life. A five-year-old says "I don't know."
Even with everything on the table, some things stayed hard to see.
- OCD83–92%
- Autism77–96%
- ADHD78–81%
- Anxiety69–85%
- Language and speech disorders58–78%
- Depression50–60%
- Intellectual disability15–23%
Depression was caught in about half of the children who had it. Intellectual disability was almost never suggested, which makes sense: it is established with tests, not with a description.
What this does not show
- Nine out of ten children here had a diagnosis. That is what a specialist clinic looks like. Among children in general, most have none. An AI that leans towards finding something would then be wrong far more often. That last step is our reasoning, not something the study measured.
- Only 21 children had no diagnosis. The figures about them are rough, and the researchers say so.
- The doctor's notes were written by the doctor who made the diagnosis. The diagnosis words were removed, but the notes may still have pointed in the right direction. So the best result may flatter the AI.
- The doctor is not a perfect answer key. Each diagnosis was made by one psychiatrist in ordinary clinical work, without a second opinion.
- Everything was in Turkish, and the stories were short. A long conversation with a chatbot at home is a different situation, for better or worse.
- Nobody measured what happens next. The study compared answers. It did not follow any family that acted on what an AI said.
If you are the parent with the phone
The researchers' own conclusion is sober: the models did best with a professional's observations in hand, and they often gave a diagnosis to children who did not have one.
From that, three things seem fair to say. They are our reading, not medical advice.
An AI answers your description, not your child. If you write when you are most worried, you will get an answer that fits your worry.
"It could be ADHD" from a chatbot is a weaker signal than it feels like. In this study the models said something like that about most children whose parents were concerned, including those where the specialist found nothing.
Writing it down is still useful. What the parents said in 32 words was the single most valuable piece of information in the whole study. Notes about what you see, when it happens and how long it has lasted are exactly what a GP or a school nurse will ask for.
The same rule holds for the screening tests on this site, including the ones for ADHD and autism. A high score is a reason to talk to someone who can assess the whole person. It is never a diagnosis.
The child in the study who said nine words was not wrong, and neither was the parent who said thirty-two. They were each telling the part they could see. The job of putting the parts together still belongs to someone who is in the room.
Sources and further reading
- Kayış, H., Tahıllıoğlu, A., Hira Selen, A. T. & Gedizlioğlu, Ç. (2026). Large language model diagnostic performance across informant sources in child and adolescent psychiatry: a prospective multicenter study. Scientific Reports. Published online 20 November 2026 as an early version of the accepted article.


