[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"blog-ai-tutor-conversation-styles":3},{"id":4,"title":5,"author":6,"body":7,"date":166,"description":167,"extension":168,"image":169,"meta":170,"navigation":171,"path":172,"seo":173,"stem":174,"tags":175,"__hash__":181},"blog\u002Fblog\u002Fai-tutor-conversation-styles.md","An AI tutor that won't just hand over the answer: conversation styles in LearnBuilder","Peter",{"type":8,"value":9,"toc":155},"minimark",[10,14,17,30,41,46,49,83,86,90,93,100,103,110,114,117,121,124,128,131,135,138,142,145,152],[11,12,13],"p",{},"Drop an AI tutor into a lesson and the obvious behaviour is: the learner asks a question, the tutor answers. It's the default everyone reaches for, and it's genuinely useful — a patient helper sitting next to the material, available at two in the morning, never sighing at the third re-phrasing of the same question.",[11,15,16],{},"It's also, on its own, the behaviour the research is most sceptical about.",[11,18,19,20,24,25,29],{},"The most useful study I know on AI tutoring is Bastani et al.'s ",[21,22,23],"em",{},"Generative AI Can Harm Learning"," (Wharton, 2024). Students given an unrestricted, answer-everything AI during maths practice scored ",[26,27,28],"strong",{},"worse"," on a later exam than students with no AI at all. A constrained \"tutor mode\" version — one that guided rather than solved — did better. The mechanism is the point: a too-helpful tutor removes the productive struggle that builds durable knowledge. Frictionless answers feel like learning and often aren't.",[11,31,32,33,36,37,40],{},"So when we rebuilt the AI Tutor block, the headline change wasn't a smarter model. It was giving authors control over ",[21,34,35],{},"how"," the tutor teaches — a ",[26,38,39],{},"conversation style"," — so \"just answer everything\" stops being the only option.",[42,43,45],"h2",{"id":44},"five-ways-to-run-a-tutor","Five ways to run a tutor",[11,47,48],{},"Each style changes how the tutor drives the conversation. Pick one per block:",[50,51,52,59,65,71,77],"ul",{},[53,54,55,58],"li",{},[26,56,57],{},"Answer questions."," The classic. The learner asks, the tutor answers from the material. Best for reference-heavy lessons, glossaries, and \"I just need to look something up\" moments.",[53,60,61,64],{},[26,62,63],{},"Quiz the learner."," The tutor asks one question at a time, checks each answer, and gives feedback before moving on. It opens the conversation itself — the learner arrives to a first question, not a blank prompt.",[53,66,67,70],{},[26,68,69],{},"Socratic coaching."," The tutor never gives the answer away. It responds with hints and counter-questions, pushing the learner to reason toward it. This is the \"desirable difficulty\" style made literal.",[53,72,73,76],{},[26,74,75],{},"Practice scenario."," The tutor plays a role so the learner can rehearse applying the material — an upset customer with a billing complaint, a patient describing symptoms, a stakeholder pushing back on a recommendation. It stays in character until the learner works through it.",[53,78,79,82],{},[26,80,81],{},"Custom."," Describe exactly how the tutor should behave and it follows your instructions instead of a preset.",[11,84,85],{},"Any preset can be refined with a line of extra instruction. \"Focus on the safety rules; start easy and get harder; ten questions, then summarise.\" \"Push the learner to justify every claim with something from the material.\" \"Play the customer as impatient but fair.\" You're not choosing from four fixed personalities — you're choosing a starting posture and then shaping it.",[42,87,89],{"id":88},"the-pedagogy-is-the-feature","The pedagogy is the feature",[11,91,92],{},"It's worth being explicit about why this matters, because it's easy to read \"five styles\" as a cosmetic menu.",[11,94,95,96,99],{},"Two of these styles — Socratic coaching and Quiz — are deliberately ",[21,97,98],{},"withholding",". They are designed not to do the thing an AI is most eager to do, which is to be maximally, instantly helpful. That restraint is the whole point. Robert Bjork's work on desirable difficulties and John Sweller's cognitive load theory both predict the same thing from different directions: a learner who is handed the answer before wrestling with the problem retains less. The Bastani result is that prediction showing up in a live classroom.",[11,101,102],{},"An answer-everything tutor optimises for what the learner asks for in the moment. A Socratic or quizzing tutor optimises for what the learner will remember next week. Those are different goals, and until now the block only supported the first one.",[11,104,105,106,109],{},"The practice scenario style is the other end of the same idea — retention through ",[21,107,108],{},"use",". Reading about de-escalation is not the same as having to de-escalate an AI that's genuinely playing difficult. The tutor holds the scenario open so the learner has to perform, not just recognise.",[42,111,113],{"id":112},"it-stays-inside-your-material","It stays inside your material",[11,115,116],{},"None of this is a general-purpose chatbot bolted onto a lesson. Each tutor is grounded in the content you point it at — the lesson it lives in, and the course around it. Ask it something the material covers and it answers from the material; ask it something off-topic and it steers back rather than free-associating from the open internet. For a compliance course or an internal-knowledge course, that boundary is the difference between a tool you can ship and one you can't.",[42,118,120],{"id":119},"it-remembers-and-the-learner-can-leave","It remembers, and the learner can leave",[11,122,123],{},"A tutor that forgets everything the moment you navigate away isn't a tutor, it's a search box. Conversations now persist: a learner can work through half a quiz, close the tab, come back tomorrow, and pick up where they left off — with the tutor still holding the thread. There's a history sidebar for returning to earlier conversations, and voice input for learners who'd rather talk than type. Guests on a shared course get the tutor too, so a preview link is a real trial of the experience, not a disabled shell of it.",[42,125,127],{"id":126},"what-the-teacher-sees","What the teacher sees",[11,129,130],{},"Anything that talks to learners privately raises an obvious question: what is it actually saying, and what are they actually asking? The block feeds an activity view on the course's AI tab — the questions learners ask, how they're doing, where they get stuck. Learners are told their tutor conversations are visible to the course owner, so this is transparency by design, not surveillance. For an author, the aggregate is quietly valuable: the questions learners ask a tutor are a direct readout of where your material is unclear.",[42,132,134],{"id":133},"setting-one-up","Setting one up",[11,136,137],{},"Add an AI Tutor block to a lesson, open it, and pick a conversation style from the dropdown. Add a line of extra instruction if you want to shape it, optionally set the opening message, and publish. For quiz and practice-scenario styles the tutor makes the first move on its own, so the learner lands on a question or a scene rather than an empty box. That's it — the grounding, persistence, and activity reporting come along automatically.",[42,139,141],{"id":140},"whats-still-open","What's still open",[11,143,144],{},"I'll keep the same honesty here that I try to keep about all of this.",[11,146,147,148,151],{},"The conversation styles change the tutor's ",[21,149,150],{},"posture",", but they can't guarantee the learner engages with the harder ones in good faith — a determined learner can still try to extract answers from a Socratic tutor, and sometimes succeed. The block reports what learners ask and how a session went, but it doesn't yet measure durable, transferable mastery — no chat interface does, and claiming otherwise is exactly the trap the research warns about. And auto-detecting when a scenario has actually been \"resolved\" well, versus merely ended, is still a judgement the tutor makes imperfectly.",[11,153,154],{},"What I'm confident about is the direction. A tutor whose default is to withhold when withholding helps, grounded in your material, visible to you, and able to remember a learner across sessions, is a meaningfully better teaching tool than a chat box that answers everything as fast as it can. The five styles are how you tell it which kind of teacher to be.",{"title":156,"searchDepth":157,"depth":157,"links":158},"",2,[159,160,161,162,163,164,165],{"id":44,"depth":157,"text":45},{"id":88,"depth":157,"text":89},{"id":112,"depth":157,"text":113},{"id":119,"depth":157,"text":120},{"id":126,"depth":157,"text":127},{"id":133,"depth":157,"text":134},{"id":140,"depth":157,"text":141},"2026-07-15","The AI Tutor block can now teach in five different ways — answer, quiz, Socratic coaching, practice scenario, or fully custom. Here's why 'just answering' is often the wrong default, what each style does, how it stays grounded in your material, and what teachers can see.","md","\u002Fblog\u002Fai-tutor-conversation-styles.webp",{},true,"\u002Fblog\u002Fai-tutor-conversation-styles",{"title":5,"description":167},"blog\u002Fai-tutor-conversation-styles",[176,177,178,179,180],"AI","AI tutor","instructional design","authoring","LearnBuilder","33AMTcHOBo6OsKBg74DTWBW8Kpve4rM4dlzhWVBTCmo"]