How Canvas LLM Assignments Let Professors Control AI Interactions
Professor Maria Chen stared at her laptop screen at 11 PM on a Sunday, reading through forty-two essays on Shakespeare’s use of irony in The Merchant of Venice. Something felt off. Not wrong, exactly - the essays were competent, well-structured, even polished. Too polished. Every third paper read like it came from the same voice, the same careful hedging, the same predictable paragraph rhythm. She’d been teaching English Composition at a mid-sized state university for nine years, and she knew what student writing looked like. This wasn’t it.
Her department chair had sent around a memo the previous week about Canvas’s new LLM-powered assignment tools. Maria had skimmed it and moved on. But now, sitting with a stack of suspiciously uniform essays, she pulled the memo back up.
What Canvas Actually Built
Canvas LMS, the learning management system used by thousands of universities, rolled out a set of features that let instructors build assignments where students interact with an AI - but on the professor’s terms. Think of it like giving students access to a workshop with power tools, except the instructor decides which tools are plugged in, what safety guards are attached, and what the students are allowed to build.
The system works through OpenAI’s API, integrated directly into the Canvas assignment workflow. When a professor creates an LLM-enabled assignment, they don’t just toggle AI on or off. They configure the interaction. They write custom system prompts that define how the AI should behave during that specific assignment. They can restrict the AI to asking Socratic questions rather than providing direct answers. They can instruct it to refuse to generate full paragraphs, or to only give feedback on drafts the student has already written.
Maria spent that Monday morning experimenting. She created a new assignment for her Tuesday section - same Shakespeare prompt, but this time she configured the LLM to act as a “writing tutor who never writes for the student.” Her system prompt told the AI to ask probing questions about the student’s thesis, point out logical gaps, and suggest they look at specific scenes in the play. It could not produce essay text. It could not outline for them. If a student asked it to “write my introduction,” it would redirect them to articulate their own argument first.
This distinction matters more than it might seem. The difference between “AI does the work” and “AI coaches the process” is the difference between a calculator that spits out answers and one that shows you where your arithmetic went wrong.
Setting Up the Guardrails
The configuration process lives inside the Canvas assignment creation flow. After selecting the LLM-enabled assignment type, professors see a prompt engineering panel. It’s not as intimidating as that phrase sounds. Canvas provides template prompts for common scenarios: Socratic tutoring, draft feedback, brainstorming support, citation checking, and a few others.
Maria started with the Socratic tutoring template and modified it. She added a line: “If the student submits text that appears to be AI-generated or copy-pasted from an outside source, ask them to explain their reasoning in their own words before continuing.” She added another constraint telling the AI to reference only Acts 1 through 3, since her class hadn’t covered the rest yet. That last part surprised her - she could scope the AI’s knowledge to match where her students actually were in the curriculum.
Professors can also set interaction limits. Maria capped each student at fifteen exchanges with the AI per assignment. Enough to get meaningful feedback on a draft, not enough to iterate endlessly until the AI had essentially shaped the entire paper through suggestion alone. Some of her colleagues in the STEM departments set tighter limits - five exchanges for a problem set, where students had to show their work between each AI interaction.
The system logs every student-AI conversation. This is where things get genuinely useful for assessment. Maria could see that one student, Jake, asked the AI six times to “make my thesis stronger” without ever revising it himself. Another student, Priya, used her exchanges to test three different interpretations of Shylock’s monologue before committing to one. Same tool, radically different engagement. The logs made that visible in a way that final submissions alone never could.
Canvas also gives professors a dashboard showing aggregate patterns. Average number of exchanges per student. Most common types of requests. How many students hit the interaction cap. Maria noticed that students who engaged more deeply with the AI’s questions - actually revising between exchanges rather than just asking again - produced measurably stronger essays. Not because the AI wrote better content for them, but because the back-and-forth forced them to think.
Where It Gets Complicated
Three weeks into using the system, Maria ran into its limits. One student figured out that while the AI wouldn’t write essay paragraphs, it would happily generate “example sentences that demonstrate strong analytical writing about Shakespeare.” The student then stitched these examples together with minor modifications. Technically, the AI followed its instructions. Practically, the student found a workaround.
Maria revised her system prompt to close that gap, adding a constraint against generating example sentences longer than fifteen words. But the incident highlighted something fundamental about these tools: they’re only as good as the prompts that govern them. A professor who writes a vague system prompt - “help students but don’t do the work for them” - will get vague enforcement. Specificity matters. It’s the same principle behind writing good exam questions. Ambiguity creates loopholes.
Not every department adopted the tools with equal enthusiasm. The Computer Science faculty worried about students using the AI to debug code in ways that bypassed learning. The Philosophy department saw potential for structured argument analysis. The Math department remained skeptical that language models could meaningfully assist with proof-based courses. Each discipline needs its own approach to prompt configuration, and Canvas’s one-size-fits-all templates only go so far.
There’s also the equity question. Students with prior experience using AI tools - often those from better-resourced high schools - navigate these interactions more effectively. They know how to ask good questions, how to push back on unhelpful responses, how to extract value from a conversation with a language model. Students encountering AI for the first time need more scaffolding. Maria started dedicating fifteen minutes of class time to demonstrating effective AI interaction strategies, which helped, but added to an already packed syllabus.
The data privacy dimension deserves attention too. Student conversations with the AI flow through OpenAI’s API. Canvas’s data processing agreement covers FERPA compliance, but some faculty remain uneasy about student intellectual work passing through third-party servers. The university’s IT security team reviewed the implementation and approved it, though they required that no personally identifiable information be included in AI prompts - a restriction that sometimes made the interactions feel stilted.
By the end of the semester, Maria had a clearer picture of what these tools could and couldn’t do. Her students’ writing improved, but not uniformly. The students who were already strong writers used the AI as a genuine thinking partner. The struggling writers benefited most from the structured Socratic questioning, which forced them to articulate ideas they might otherwise have left vague. A small group - maybe 10% - continued to look for shortcuts regardless of the guardrails in place.
She submitted her course evaluation data along with a recommendation to her department: adopt Canvas LLM assignments broadly, but invest in faculty training for prompt engineering. The tools work. But they work like any teaching tool - they require a skilled instructor behind them, making deliberate choices about what students should and shouldn’t have access to. The AI isn’t the teacher. The professor still is. Canvas just gave professors a new set of dials to turn.