Same question. Four different loops with different memories, trust models, and guarantees about what the AI knows.
Your student has a question about decision trees: was the four-feature limit pedagogical or real? Too specific for chapter prose, perfect for office hours. Where she asks depends on where she reads. Canvas gets IgniteAI. Medhavy gets a native tutor with course memory. Kindle has no embedded AI—just a companion prompt she pastes into Claude. React has a chat component wired to an API. Same question. Four different loops, four different memories and trust models, four different guarantees about what the AI knows about this reader in this course.
The fluency trap migrates from the author who outsources judgment to the student who outsources hers.
The author's job across all surfaces is identical: keep AI as provocation, not vending machine. The fluency trap does not stop at the author's desk; it migrates to the student. When the answer arrives polished and complete with no loose threads, it stops the student's thinking at the exact moment her thinking should begin. That is the trap. The alternative is surfaces where the AI prompts the student back into her own reasoning—where assistance raises questions instead of closing them. This distinction between provocation and performance is the dividing line. It determines whether Ask-AI anywhere improves learning or degrades it.
Each surface carries the same question through a different context, with different author control points and different AI knowledge about the reader.
Four surfaces, four different mechanics. Canvas is the LMS instructors live in. IgniteAI is embedded when licensed; your lever is the course structure in the `.imscc`. Medhavy is an AI-native tutor via LTI. Its Ask-AI function accumulates context across sessions—which concepts generate questions, which phrasing confuses. That memory makes it the richest loop. React's AskAI component is a chat interface wired to API with system prompt and guardrails you write. No course memory unless you build session state. Finally, Kindle and PDF have no embedded AI, but ship with a companion prompt—the reader pastes chapter text plus prompt into Claude or ChatGPT, instantiating a tutor loop. Four routes. Four memories. Four different scopes of what the AI can know.
IgniteAI draws on what you put into the `.imscc` export. Better course design means richer AI responses.
IgniteAI is embedded in Canvas when institutions license it. Your lever is the course structure you gave it. Outcomes statements, aligned objectives, and context notes aren't decoration—they're material IgniteAI draws on when answering. A thin `.imscc` gives IgniteAI thin context. Answer quality reflects your course design, not just the underlying model. Sparse outcomes and no alignment leave IgniteAI with little to anchor on. Thick structure with clear objectives and rich context notes about what each section aims to do lets IgniteAI answer with precision. Better scaffold, richer loop.
A student asking about decision trees is asking inside a system that already knows her prior questions about overfitting.
Medhavy is an AI-native tutor via LTI, and its distinguishing feature is persistence. Medhavy's Ask-AI accumulates context across student sessions. It knows which concepts generate questions, which phrasing confuses, which exercises get most re-reads. That memory makes it the richest loop. A student asking about decision tree feature counts on Medhavy is asking inside a system that already knows her two prior questions about overfitting last Tuesday. The response reflects that context. The author's job mirrors IgniteAI: the course structure you give Medhavy is the scaffolding it leans on. Richer course design yields richer AI context and better service to each student.
The system prompt is your voice in the embedded model. It travels into every conversation the student has through AskAI.
Chapter 18's build script generated an AskAI.tsx placeholder—a chat component wired to an API, with system prompt and guardrails you specified. The developer activates it. You wrote what the component says about itself, what it refuses, how it frames uncertainty. On the React site, Ask-AI has no course memory unless you build session state into the API. It knows your system prompt and what the student types. That is the scope. The author's job is not implementation; it is the system prompt—the document that tells the model who it is, what this course is for, what students must bring, and critically, what it should refuse even when asked.
The prompt text is the author's leverage point where the format offers nothing but the chapter itself.
This path has no native AI. EPUB or PDF contains no embedded model, no API, no LTI. What they can contain: a companion prompt—the reader pastes it alongside chapter text into Claude or ChatGPT, instantiating a tutor loop in their own LLM. The companion prompt is the "+1" static formats cannot host natively. It's the author's intervention where the format otherwise offers just text. For Kindle and PDF readers, this is the only path to an interactive loop. The prompt itself, ready to paste in Appendix 97, is the author's lever when native AI cannot exist on the platform.
The scope of memory shapes what the AI can infer about the reader's progress and needs.
The memory dimension separates the surfaces. Canvas stores course-session context—what happened in this course, this session. Medhavy stores persistent, per-student memory accumulating across sessions, and aggregate patterns across cohorts. React stores session-only memory without additional backend state. Kindle and PDF have no persistent memory; the companion prompt instantiates a fresh loop each paste. Each memory scope shapes what the AI can infer and answer. A Medhavy student gets an answer referencing her prior five questions. The same student on React gets an answer knowing none of that. Neither is wrong; they're different leverage points, different tools for different teaching aims.
Same job, four different levers. Each is where the author shapes what the AI knows about this course and this reader.
The author's job shifts by platform, but the principle is constant: you control what the AI knows. Canvas: outcomes, alignment, context in the `.imscc`. Medhavy: course structure and configuration. React: system prompt and guardrails. Kindle/PDF: the companion prompt itself. Four platforms. Four control points. In each case, the AI's delivery quality depends on what you put into the surface. You cannot outsource this to the platform. The platform doesn't know your course better than you do. Your structure, framing, careful choices about what to include and refuse—these make the Ask-AI loop work. This is non-delegable work.
Better scaffold. Richer loop. The same principle holds across all four surfaces, each with its own author leverage point.
The artifact meets the student in four ways, but the principle is identical across all. Whether Canvas or Medhavy or React or Kindle, the author structures the context the AI draws on—making the surface rich enough that AI provokes rather than performs. On Canvas: invest in course outcomes and alignment. On Medhavy: design course structure for clear scaffolding. On React: write a system prompt reflecting your teaching voice and guardrails. On Kindle: craft a companion prompt activating the reader's reasoning. Same job, four levers. Better structure yields better context and richer loop. Poor structure yields thin answers and performance instead of provocation.
The fluency trap migrates from the author to the student. Your job is to build surfaces where the polished answer invites more thinking, not fewer questions.
Keep the AI a provocation, not a vending machine. This principle threads through all four surfaces. The fluency trap doesn't stop at the author's desk; it migrates into the student's hands. An author outsourcing judgment to AI creates students who outsource theirs. The polished, complete answer with no loose threads stops the student's thinking at the exact moment her thinking should start. Your job is building surfaces where this doesn't happen—where the answer provokes a better question rather than closing conversation. That distinguishes teaching with AI from teaching through AI. The author's choice. The author's responsibility. Across Canvas, Medhavy, React, and Kindle, there is one job: keep the AI as a tool making the student think harder, not one that thinks for her.
AI 1 · Chapter 20 · Ask AI Everywhere