Is AI maths help safe for kids?

What did the research find?

The best recent evidence is a pre-registered randomised trial at one large high school in Turkey, run in 2023, with nearly 1,000 students in Years 9 to 11.¹ Classes were assigned to one of three setups:

• A general chatbot, like ChatGPT, built on GPT-4.
• A tutor version of the same chatbot, set up with the correct solutions and notes on common mistakes, and told to give hints, not answers.
• No chatbot, just the textbook and notes.

Students with either chatbot did better during practice. Then the chatbot was taken away for a closed-book exam on matched problems. Students who had used the general chatbot scored 17% lower than students who had only their textbook and notes. Students who had used the hint-giving tutor scored about the same as the no-chatbot group: no harm, but no benefit either.¹

This was a short-term measure: four sessions, with the exam at the end of the same session, in class with a teacher. So it can’t tell us about learning months later, or about a child at home.

Why would getting answers lower learning?

In the study, students using the general chatbot mostly asked it for the answer, and the pattern of errors suggested they were copying answers rather than reading the working. The authors call it using the AI as a crutch.

There’s an older reason to expect this. Decades of memory research suggest that working something out yourself makes it stick better than reading it.² That research used word lists with adults, so it is a reason the pattern makes sense, not proof of what happened in the trial.

There’s one more finding to know about. Students couldn’t easily tell that the help had stopped them learning. Those who used the general chatbot didn’t think they had done worse or learned less, and those who used the tutor thought they had done significantly better, but they hadn’t. So a child’s own sense of “I understood that” is not a reliable check.

Can a chatbot give the wrong answer?

Yes. In the trial, the 2023 version of GPT-4 often gave wrong answers or wrong working on these problems. Newer models will differ, so we can’t give you a current error rate. It’s also our reasoning, not a research finding, that if teenagers didn’t catch the mistakes, younger children are unlikely to.

Does the reading matter for younger children?

We expect it does, but this is reasoning from established science, not a tested finding. We haven’t found a study measuring chatbot maths explanations against children’s reading levels. What we do know is that working memory holds only about four chunks at once.³ A young child who has to read and hold a long, wordy reply while also doing the maths is using the same limited space for both. A child who can’t read the reply can look like a child who can’t do the maths. Our article on reading and maths goes into that.

So is AI bad for maths learning?

Not in general, and the research doesn’t say so. The hint-giving tutor in the trial caused no harm. A living review that pools many studies, which is a preprint and not yet peer reviewed, found positive effects on average, but mostly where AI was used alongside a teacher.⁴ It covered 34 studies and 4,349 participants, and it changes every couple of months. The same review found the evidence that a student using AI on their own gains more is weak, and that is the case closest to a child alone with a chatbot at home.

What should I ask about any AI maths tool?

• Does it give the answer, or does it make your child do the last step?
• What does it send to the AI, and does that include your child’s name, school or email?
• Is the help written at a level your child can read?
• Where is your child’s data kept, and for how long?

How Workings does this

The AI in Workings writes words. It never decides what the right answer is: every answer is computed by code, and the answer is never sent to your child’s browser. Workings never gives the answer, in any mode, and not after any number of tries. When a child is stuck, the help is a question aimed at the mistake, a hint, help with the reading, or a smaller question, and the child takes the last step.

Here is what the AI does, from our privacy policy and FAQ:

• A language model drafts question wording ahead of time, into a bank checked by software and approved by a person. No model writes a question while a child waits.
• Apart from the Pip trial below, there is no open chat. Anything a child types is screened for safety first.
• Some help is written live by a model, about the question in front of the child only. Software checks every reply before they see it: never the answer, and only the question’s own numbers.
• Talking with Pip is a trial that a parent switches on for each child. Pip talks only about the question in front of the child, and knows their year, reading level and up to three interests, never their name. Parents don’t see the chat, and chats are deleted after 30 days. Children who use Workings through a school are never in the trial.
• A child’s name, email, school, avatar or any identifier is never sent to a model.
• Our model provider is OpenAI. It receives the text described above, with no child’s identifier, and processes it in the United States.

The design choice that matters most to this research is hints rather than answers. That’s the setup that avoided harm in the trial. It did not, in that trial, produce extra learning. We haven’t yet measured Workings’ effect on learning. We don’t publish efficacy percentages, and we plan an independent, pre-registered evaluation. Read the privacy policy or see how we keep the AI from deciding the answer.