DeepSeek shows its working. Teach your child to mark it
Anyone who has sat beside a Primary 5 child at the dining table knows the refrain. Show your working. The answer at the bottom is the least interesting line on the page; the teacher wants to see how you got there.
In January 2025, the Chinese lab DeepSeek released a chatbot built around roughly the same idea. Its R1 model, which DeepSeek said performed “on par with OpenAI-o1” on maths, code and reasoning, was a reasoning model: it writes out a chain of thinking before it commits to an answer. DeepSeek invited people to “Try DeepThink” on its website and released the model under the MIT licence, so anyone could build on it.
Here is the useful part for a family. A chatbot that shows its working gives your child something better to do than copy. They can mark it.
Why marking beats asking
When a child asks an AI for help with a maths problem, the usual pattern is: question in, answer out, answer copied. Nothing about that builds the skill the homework was set to build.
Turn it around and the child becomes the teacher with the red pen. They are no longer asking “what is the answer?” but “is this step right, and how do I know?” That second question is the whole of maths, and it is also the single most useful habit anyone can have with AI. If you want the broader version of this habit, our earlier piece on teaching kids to fact-check AI covers words and facts. This one is for numbers.
The 20-minute routine
You need a chatbot, one maths word problem your child has already solved on paper, and a pencil. The prompts below work in DeepSeek and in the other mainstream chatbots, because they ask for numbered steps in plain words rather than relying on a particular button.
Before you start, use your own account, not your child’s. DeepSeek’s terms say the service is “primarily intended for adults” and that anyone under 18 should use it “only with the consent of your legal guardian”. Its privacy policy says it does not knowingly process children’s personal data, and that it collects, processes and stores personal data “in People’s Republic of China”. So: your login, your child beside you, and nothing typed that identifies them. No name, no school, no photo of a worksheet with a name on it. Type the problem out.
- Open the chatbot. If you use DeepSeek, its April 2026 release notes say the web chat at chat.deepseek.com offers an Expert Mode and an Instant Mode. Either is fine for primary school maths.
- Type this, then paste the problem underneath: “Solve this step by step. Number every step. Explain each step in one sentence a 10-year-old could follow. Put the final answer on its own line at the very end.” You should get back a numbered list of steps with the answer on its own line at the bottom.
- Cover the last line. Literally, with a hand or a sticky note on the screen. The answer is the least interesting part, remember.
- Your child marks each step. A tick or a cross on paper, one per numbered step. For every step, two questions: does it use the numbers that are actually in the question, and does it follow from the step before? If their own working took a different route, that is fine. Different is not wrong. They just have to say whether the AI’s step is true.
- Ask “why” about one step. Pick the step your child found hardest and type: “Why is step 3 true?” (with the right number). Your child decides whether the explanation actually explains anything or just repeats the step in longer words.
- Now plant a mistake. Type: “Solve it again, but make exactly one mistake somewhere in the steps. Do not tell me where.” Your child hunts for it. If the chatbot points out its own mistake anyway, say “Again, and this time keep the mistake hidden.” In our view this is the best round of the six, because finding an error in a computer’s working feels like winning.
Then uncover the answer, compare it with what your child got on paper, and stop. One problem done properly beats ten copied.
[[photo:1]]
The catch: the working is not a confession
There is one thing to tell your child honestly, and it is the reason this routine marks steps rather than admiring them.
Showing the working does not mean the working is how the AI actually got the answer. In an April 2025 study, researchers at Anthropic slipped hints to two reasoning models, their own Claude 3.7 Sonnet and DeepSeek’s R1, and checked whether the models admitted using the hint in their visible reasoning. Across the hint types, Claude 3.7 Sonnet mentioned the hint 25% of the time and R1 mentioned it 39% of the time. “A substantial majority of answers, then, were unfaithful,” the researchers wrote. They also found the unfaithful explanations were, on average, “substantially longer than the faithful ones”.
Longer, in other words, and less honest. A tidy, confident list of steps can be a story written after the fact.
That is exactly why the child marks each step against the question, not against how sure the steps sound. A step is right because the numbers work, never because the chatbot wrote it neatly. It is the same lesson behind why chatbots make things up: fluent and correct are two different things.
A word on DeepSeek itself
We picked DeepSeek because, in our view, it did more than anyone to make visible reasoning ordinary, not because it is the best home for your child’s homework. Some governments have pulled it off their own devices: the Victorian Government in Australia, following a Commonwealth direction, told its public bodies to remove DeepSeek products from government systems because the Commonwealth judged them to pose “an unacceptable level of security risk”. That is a ruling about government devices, not families, but read it next to the privacy policy’s line about data stored in China and you have a decent reason to keep the account yours and the details vague.
The routine works the same in whichever chatbot you already use. What matters is the swap at the centre of it. Tonight, instead of the AI checking your child’s working, your child checks the AI’s. Twenty minutes, one problem, a pencil. And, if it goes well, a child who has learnt that a computer’s answer is something you mark, not something you copy.
Was this useful?
Sources
- DeepSeek-R1 Release · DeepSeek
- DeepSeek V4 Preview Release · DeepSeek API Docs
- DeepSeek Terms of Use · DeepSeek
- DeepSeek Privacy Policy · DeepSeek
- Reasoning models don't always say what they think · Anthropic
- Direction on the use of DeepSeek products, applications and web services · Victorian Government