China's AI homework apps don't just give the answer

A mother in Henan province told ABC News in December why she now leaves her children’s homework to an app instead of sitting with them herself. “My children behave better in front of AI,” she said. “They think I talk too much.”

It is a funny line until you notice how much it gives away. The AI is not winning on patience, or on explaining long division better than she can. It is winning because it does not sigh, does not repeat itself, does not have a day behind it. A journalist in Guangdong quoted in the same piece put it more plainly: he uses the app, called Dola, “when I’m too busy”, and prefers to sit with his son himself whenever he has the time.

That is the honest starting point for a story that could easily turn into either cheerleading or panic. It is neither. China’s homework-AI habit is now large enough to measure, and what the apps actually do, when you look past the marketing, is a useful thing to copy, with or without the app.

The habit, by the numbers

ByteDance’s Dola, a chatbot built to check on children while they study, has 172 million users a month, ABC News reported in December, with a dedicated learning version, Dola Aixue, drawing another 8.76 million. Its sibling app Doubao, described by The New York Times (syndicated via The Star) as China’s most popular AI chatbot, has a camera function parents use to photograph a problem on the page. DeepSeek gets used for the same homework, and sometimes to build a learning game out of it. One mother bought her daughter a dedicated translation gadget, at around US$375, so the child could talk to it in English from an early age. The Times put the Chinese edtech market this habit now sits inside at more than US$43 billion.

None of this is a niche curiosity. A 2025 KPMG survey found more than 90 percent of Chinese respondents optimistic about AI technology, against just over half of Americans. Caixin Global, reporting on the industry from the inside, found token consumption for K-12 tutoring through ByteDance’s Volcano Engine grew twelvefold in six months to May last year, and that roughly a fifth of all queries on DeepSeek and Doubao now concern the school curriculum. A fifth of everything two of the country’s biggest chatbots are asked is, in some form, homework.

What the industry admits is the hard part

Here is the part worth borrowing. Caixin’s reporting on the sector names the thing that actually separates a good homework app from a bad one, and it is not speed or price. It is whether the thing explains itself. Doubao Aixue is built to use image recognition and reasoning to walk through a problem rather than just solve it; a rival from NetEase Youdao, a model called Ziyue-o1 released last January, was built specifically to generate step-by-step explanations instead of a final number. Caixin’s own framing of the industry’s struggle is blunt: complex junior-high maths needs “nuanced, stepwise explanations rather than simple answers”, and getting a model to do that reliably is a harder problem than getting it to produce the right number at all.

That is worth sitting with. The companies building these tools, at enormous scale, have concluded that an answer on its own is close to worthless, and an explanation is the product. A Shanghai mother quoted by ABC News, who works in international education, gave the honest counterweight: the apps help, “but it can also reduce the process of thinking they do.” Both things are true of the same tool, depending entirely on how it gets used.

The three-step version, for whatever you already have

You do not need Dola or Doubao to borrow the move, and in Singapore you likely cannot get the under-18 version of either in any case. Any chatbot that can read a photo, which by now includes most of the major ones, can be made to do the same three jobs. The trick is in what you ask for, not which app you open.

1. Photograph the problem with the wrong answer still on the page. Not a fresh copy of the question. The actual working, mistake included.
2. Ask the chatbot to name the specific error and explain the idea behind it, and tell it explicitly not to give you the corrected answer first. Something close to: “Don’t tell me the right answer yet. Tell me what went wrong in this step and why.”
3. Once your child has followed the explanation, ask for two new problems testing the same skill, with different numbers, and let your child try those alone.

The part that matters is the order. Step two without step three is just a nicer way of being told you were wrong. Step three is the bit that tells you, and your child, whether the explanation actually landed, which is the only question that was ever worth asking.

The thing the app can’t do

Go back to that Henan mother for a second. Her children behave better in front of the AI because the AI has no memory of yesterday’s argument and no opinion about whether they are trying hard enough. That is a real advantage, and also the limit of it. An app can watch the working, catch the error type, and hand back two fresh problems without getting tired. It cannot tell you, afterward, whether your child actually understood or just got lucky on easier numbers. That second look, five minutes at the end, is the part no amount of patience from a chatbot replaces, and it happens to be the one thing a parent is still better positioned to do than the tool that is doing everything else.