The Answer Machine Problem: Why AI Tutors Should Ask More Than They Tell By Rhys Higgs, founder of Discourse AI, classroom teacher. There’s a moment every teacher knows. A student is stuck. Their hand goes up. And you have a choice: give them the answer, or ask them a question that helps them find it themselves.The first option feels kind. It’s faster, the student is relieved, the lesson moves on. The second option feels slower and sometimes uncomfortable — for both of you. But every teacher learns, usually within their first year, which one actually produces learning.Now scale that moment up. Millions of students have an AI in their pocket that will always, instantly, take option one. And most AI tools in education are being built to do exactly that: answer faster, answer better, answer more.I think we’re optimising the wrong thing.What the research has always told usNone of this is new, which is what makes it strange that we keep forgetting it.The “generation effect” — the finding that we remember information better when we produce it ourselves rather than receive it — has been replicated for decades. Retrieval practice consistently outperforms re-reading and passive review. And work on “desirable difficulties” shows that a degree of struggle during learning, the kind that feels inefficient in the moment, produces stronger long-term retention than smooth, frictionless instruction.In other words: the discomfort of not being told the answer isn’t a bug in learning. It’s the mechanism.An AI that removes all friction removes the very thing that makes knowledge stick. A student who pastes a question and copies the response has completed a task. They haven’t learned anything — and worse, they feel like they have, because fluency is easily mistaken for understanding.The homework problem is really a design problemMuch of the current anxiety about AI in schools centres on cheating: essays written by chatbots, maths homework solved by photo. Schools respond with detection tools and bans, and I understand why. But detection is an arms race schools will lose, and bans just push use out of sight.The deeper issue is that most general-purpose AI is built to be maximally helpful in the immediate sense — to satisfy the request in front of it. For most contexts, that’s the right design. For a learner, it’s often the wrong one, because what a student asks for (“give me the answer”) and what a student needs (“help me get there”) are different things.This is a design problem, and design problems have design solutions.What a Socratic AI actually looks likeWhen we built Discourse AI, the non-negotiable was that the tutor behaves like a good teacher, not a good search engine. In practice that means:It asks before it tells. When a student is stuck, the first response is a question — “What do you already know about this?” or “What happened when you tried?” — because diagnosing the gap matters more than filling it.It scaffolds in steps. Hints come in layers. The full worked answer exists, but it’s the last resort, not the first response.It tolerates being ‘less helpful’. This is the hard part. A Socratic tutor is measurably slower at resolving a query than an answer engine. If your metric is time-to-answer, it looks worse. If your metric is what the student can do a week later, it looks very different.It knows when to just answer. Socratic method isn’t dogma. Sometimes a student needs a definition, a date, a formula — friction there is pointless. The judgement about when to withhold is exactly the judgement good teachers exercise a hundred times a day, and it’s the thing worth building into the machine.The uncomfortable trade-offHere’s my honest admission: building this way is commercially harder. Students, given the choice, will often prefer the tool that just gives them the answer — in the same way we’d all prefer the gym where the weights lift themselves. An EdTech product that adds friction is swimming against the current of what makes consumer software successful.But teachers and school leaders aren’t choosing tools to maximise student satisfaction in the moment. They’re choosing tools to produce learning. And I’d argue that in the long run, the AI tools that survive in education will be the ones that earn the trust of the profession — by being built around how learning actually works, not around how quickly a query can be closed.A question for the professionIf you’re a teacher: where’s your line? When does an AI helping a student cross into an AI replacing the learning? And if you’ve used AI tutoring tools with your classes — did they push students to think, or just to prompt?I’d genuinely love to hear your experiences, particularly the messy ones. That classroom judgement — when to tell, when to ask, when to let a student sit in the struggle a moment longer — is the hardest thing to encode, and the most valuable. Rhys is the founder of Discourse AI (thediscourse.ai), a UK-based platform that generates structured, interactive courses with a Socratic AI tutor. Before founding Discourse, he was a classroom teacher and Digital Lead at a multi-academy trust.