AI in the Classroom Without Doing Students' Homework

TL;DR: The difference between AI that helps students learn and AI that does their homework is instruction design. Configure your classroom AI to ask what students understand before explaining, to give hints rather than answers on problems, and to explicitly refuse to complete graded work. Those three instructions change everything about how the tool behaves.

Every educator who's considering classroom AI has the same worry: what stops students from using it to skip the thinking entirely? It's a reasonable concern — unconfigured AI tools will absolutely write an essay, solve a problem set, or summarize a reading for a student who asks.

But "AI could be misused" and "AI will be misused" are different statements — and the gap between them is instruction design. A knowledge-based agent built on your course materials and configured with the right behavioral instructions behaves fundamentally differently from a general AI tool students access on their own.

Here's what that design actually looks like in practice, and why it changes the academic integrity calculus significantly.

The academic integrity problem with AI isn't that students use it — it's that most AI tools make completing work for students the path of least resistance. A well-designed course AI agent inverts that: it's more useful for understanding than for getting answers, which means students who use it as intended come away having learned. The design choices are specific and implementable, which is what this post covers.

The Core Design Principle: Ask Before You Explain

The single most effective instruction for academic AI integrity is also the simplest: ask the student what they already understand before providing any explanation.

This one pattern shifts the interaction from passive consumption to active engagement. A student who wants a copy-pasteable answer needs to work harder to get it — they have to answer questions about their own understanding first, engage with the concept to give any coherent response, and produce the thinking the agent is looking for before it provides information.

Students trying to shortcut the process typically disengage within 2–3 exchanges when the agent keeps asking questions rather than providing answers. Students genuinely studying engage more — they get more out of the conversation because the agent is building on what they know.

The instruction to write: "When a student asks about a concept, ask them what they already understand about it before explaining. Build on what they know rather than presenting information from scratch every time."

Setting Explicit Scope Limits

An AI agent configured with explicit refusal language handles academic integrity at the instruction level — before it ever reaches the content.

Write these instructions explicitly, not aspirationally. Not "try to avoid doing students' work" but "Do not complete graded assignments, write essays for submission, provide answers to graded problem sets, or complete any work a student will submit for a grade. If a question sounds like it's from a graded submission, redirect: 'I can help you understand the concepts involved — what part are you working through?'"

Specificity matters here. A vague "don't help with cheating" instruction produces inconsistent behavior. An explicit list of what the agent won't do produces consistent behavior every time.

Scope limits work in both directions. An agent that's too broad will drift into territory you didn't intend — answering questions about adjacent courses, making claims about things not in your materials, or engaging with topics that have nothing to do with the course. An agent that's too narrow frustrates students with legitimate questions. The sweet spot: document your intended scope in the instruction set, test a set of edge-case questions, and refine until the agent handles both clearly in-scope and clearly out-of-scope questions correctly.

Knowledge Boundaries as a Second Layer

A correctly configured knowledge-base AI only answers from what you've uploaded — it doesn't synthesize from the broader internet. That constraint is a feature, not a limitation, for academic use.

When a student asks something outside the course material, a well-configured agent says: "I don't have information about that in the course materials — I'd suggest checking the textbook or asking during office hours." That response is better for learning than a broad AI answer would be, and it keeps the student focused on the course rather than falling into general internet research.

The retrieval instruction: "Only answer questions based on the uploaded course materials. If a question is outside the scope of those materials, say so clearly and suggest where the student should look."

The retrieval instruction is your second line of defense after the behavioral instruction. Even if the behavioral instruction says "only help students understand concepts," a permissive retrieval instruction can still allow the agent to draw on general internet knowledge rather than your course materials. The combination of a behavioral instruction (what the agent does) and a retrieval instruction (where it gets information) is what creates genuinely constrained, curriculum-aligned behavior.

How This Differs From General AI Tools Students Already Have Access To

A crucial question: if students have access to ChatGPT and Claude on their own, does building a course-specific AI change anything?

For academic integrity purposes, it changes quite a lot. A general AI tool optimizes for giving a good answer — it will complete an essay prompt, solve a problem, summarize any reading. It has no awareness of your course's academic integrity requirements and no instruction to ask questions before explaining.

A course-specific AI agent configured with Socratic instructions and explicit refusal language is a categorically different tool. It doesn't compete with the student's own thinking — it supports it. The instruction design makes misuse significantly harder without making legitimate use harder at all.

The students who benefit from a well-designed course AI are the ones studying at 1am who need a concept explained. The students trying to skip the work hit a system that asks them questions instead of answering theirs — and most give up.

Deploying Transparently With Students

The framing you use when introducing the tool to students matters significantly. The right framing isn't "here's an AI that can help you" — it's specific about what the AI does and what it doesn't.

Sample student-facing language: "I've built a study companion trained on our course materials. It can explain concepts, help you work through practice problems, and prepare you for exams. It won't do your assignments for you — it'll ask you questions and guide your thinking instead. Think of it like a really patient teaching assistant who's read everything in the course and is available any time."

Students who understand what the tool is designed to do use it well. Students who try to misuse it quickly discover the Socratic design isn't helpful for that purpose. Transparent framing handles both scenarios simultaneously.

Ready to build a classroom AI with academic integrity built in? Start free on Alysium — your course materials are your build materials.

For the full build process, read How to Build an AI Study Buddy From Your Textbook. For the student perspective, see Can AI Help Students Learn — Not Just Cheat?.