Ask ChatGPT or Claude to solve a maths problem and, by default, they will — that's what a general-purpose assistant is built to do. A small number of AI tools are built the opposite way: designed specifically to withhold the final answer and question a student toward finding it themselves.

Why does "won't give the answer" matter for learning?

A worked answer solves tonight's homework question but teaches nothing that transfers to the next one. Genuine learning happens in the struggle just before the answer — trying an approach, getting it wrong, adjusting — which a tool that hands over the solution short-circuits entirely. This is the core argument for guided ("Socratic") tutoring over direct-answer tools, and it applies whether the tutor is a human or an AI.

How do AI tools compare on this specific behaviour?

Tool Default behaviour on a homework question Can be prompted to guide instead Built-in structural limit against giving the final answer
ChatGPT / Claude (general chatbots) Solves it directly unless told not to Yes, with explicit prompting each time No — depends entirely on how it's asked
aitutors.me (Pi, Quill, Darwin, Curie, Newton, Harari, Mercator) Guides with a graduated hint ladder N/A — this is the default, not an opt-in Yes — the tool that reveals hints has no "give final answer" level; the top rung is a fully worked similar problem, never the child's own answer
Video-lesson libraries (Oak National Academy-style) Explains a worked example, then quizzes Partially — depends on format No — the point is often to show a worked method
Practice/quiz platforms Shows correct answer after submission No — feedback is typically answer-first No — designed for retrieval practice, not withholding

What does "won't give the answer" actually look like in practice?

For a purpose-built Socratic tutor, it's not a setting a parent has to enable — it's how the tool is designed to behave by default, at every level of difficulty. aitutors.me's professors use a four-level hint ladder: gentle nudges first, then more specific guidance, then a strong steer, and — if a student is still stuck — a fully worked similar problem followed by "now redo yours." There is no fifth level that reveals the answer to the actual question in front of the child. That's a structural design choice, not a suggestion the model happens to follow most of the time.

Can a general chatbot be prompted to behave the same way?

To some extent, yes — telling ChatGPT or Claude "don't give me the answer, ask me questions instead" genuinely changes its response for that turn. The limitation is consistency: a general assistant has no persistent instruction locking in that behaviour across a whole session, so a differently worded follow-up question, or simply forgetting to repeat the instruction, can result in the direct answer being given anyway. A tool built around Socratic teaching as its core design doesn't have that inconsistency, because withholding the answer isn't a prompt the user has to remember to add.

Does "won't give the answer" mean the tutor is less helpful?

The opposite is usually true for actual learning, even though it can feel less immediately satisfying than a quick answer. The EEF's tuition evidence consistently shows that support structured around genuine understanding — not just task completion — produces stronger outcomes over time. A tutor that makes a student do more of the thinking is doing more of the teaching, even when the homework takes ten minutes longer to finish that evening.

Frequently asked questions

Does aitutors.me ever give the final answer to a problem?

No. Its hint ladder tops out at a fully worked similar problem, then asks the student to redo their own — it is designed never to give the answer to the exact problem the child is working on.

Will ChatGPT give my child the answer to their homework?

By default, yes, if simply asked to solve it. It can be prompted to guide instead, but that instruction needs to be given (and often repeated) — it isn't the tool's built-in default behaviour.

Is a Socratic AI tutor slower to use than a direct-answer chatbot?

Often, yes, in the moment — working through hints takes longer than reading a solution. The trade-off is that the understanding built through guided questioning tends to transfer to future problems, whereas a copied answer usually doesn't.

Can my child just keep asking for more hints until the answer slips out?

A well-designed hint ladder is structured so the top rung is a worked similar problem, not the child's own answer, specifically to close this loophole — no amount of asking should surface the exact solution.

Are there other AI tutors built around this same Socratic principle?

Purpose-built tutoring products increasingly design around guided questioning rather than direct answers, though the strength and consistency of that design varies — check specifically how a tool behaves at its most "stuck" hint level before trusting the claim.

For an AI tutor built around a hint ladder that never gives the final answer, see aitutors.me.