Blog

An AI that only answers from your course materials

The concern faculty most frequently raise about AI in education is that students will receive and trust incorrect information. This concern is entirely valid when applied to general-purpose language models, which generate plausible-sounding text regardless of its accuracy and cannot distinguish between well-supported claims and things they've made up.

The response from most institutions has been either to ban AI or to accept the risk and hope students develop the critical thinking skills to evaluate the output. Neither approach addresses the underlying problem, which is that the wrong kind of AI is being used for the wrong purpose.

An AI tutor that answers exclusively from the student's own course materials works on a fundamentally different architecture, and the difference eliminates the hallucination problem at its source.

How it works

When a student uploads their course materials to Fabric, the system indexes everything: the syllabus, the assigned readings as PDFs, the lecture slides, textbook chapters, lecture recordings (which are automatically transcribed and timestamped), and any notes the student has written.

When the student asks a question, the AI assistant retrieves the answer from this indexed content. It searches across every document by meaning, finds the relevant passages, synthesises an answer, and provides a citation for each claim pointing to the specific source: the page number in the PDF, the slide in the deck, or the timestamp in the lecture recording.

The student can click any citation and see the original text or hear the original explanation. The answer is verifiable against materials the institution itself assigned.

If the student's materials don't contain the answer, the AI says so. It does not fall back to generating from its general training data. This constraint is the single most important architectural decision, because it means the system cannot produce the kind of confident hallucination that makes general-purpose chatbots unreliable in academic settings.

What this means for faculty

For professors who are sceptical about AI in education, and with good reason, a course-grounded tutor changes the risk profile entirely.

Every answer is auditable. A professor can ask the same question the student asked and see exactly which sources the AI drew from and whether its synthesis is reasonable. This is a fundamentally different accountability structure from a general-purpose chatbot whose reasoning is opaque and whose sources are unverifiable.

The AI reinforces your teaching rather than replacing it. When a student asks the tutor to explain a concept, the explanation uses the framing from the professor's lecture and the assigned reading, because those are the sources the AI has access to. The student is engaging more deeply with the materials the professor chose, which is the opposite of what happens when a student asks ChatGPT and gets a generic explanation that may contradict the course's approach.

The quality of the tutor reflects the quality of the curriculum. This is perhaps the most interesting implication for course designers. A course with carefully selected readings, well-structured lectures, and clear supporting materials produces an AI tutor that gives excellent, well-grounded answers. A course with vague readings and disorganised slides produces a tutor that reflects that disorganisation. The AI holds up a mirror to the course materials, which creates a positive incentive: investing in better materials directly improves the AI support students receive.

Students learn to work with sources. Because every answer points back to a specific passage in a specific document, students develop the habit of tracing claims to their sources. This is the opposite of the ChatGPT dynamic, where students learn to accept generated text without verification. The tutor teaches source-based reasoning as a side effect of its architecture.

The distinction that matters for academic integrity

The core academic integrity concern with AI is that students will submit AI-generated work as their own. This concern maps onto a specific use case: a student asks an AI to write their essay, produce their problem set answers, or generate the text they submit for assessment.

A course-grounded tutor is architecturally ill-suited to this use case, by design. It explains concepts from the assigned materials. It finds relevant passages. It generates practice questions. It helps students connect ideas across readings and lectures. What it does not do is produce polished, submittable text, because that's not what it's built for.

The useful test, mentioned in the ChatGPT comparison: would a professor be comfortable if a student used a human tutor to do exactly the same things? Human tutors explain concepts, help students find relevant sources, quiz them on material, and connect ideas across readings. No professor considers this cheating. The AI version is the same set of activities, delivered at scale and available at any hour.

The AI policy framework for institutions can make this distinction explicit: AI that helps students understand assigned materials is an appropriate study tool. AI that generates work the student submits as their own is not. The architecture of the tool, not just the policy around it, should enforce this boundary.

For students: how to get the most from it

If your university provides access to a course-grounded AI tutor, or if you set one up yourself, a few practices make it significantly more useful.

Upload everything. The tutor can only answer from what it has. Your syllabus gives it the course structure. Your readings give it the arguments. Your lecture recordings give it what your professor actually said, including the emphasis, clarifications, and tangents that don't appear in the slides. The more complete your library, the better the tutor's answers.

Ask it to explain using the lecturer's framing. If you didn't understand a concept from the lecture, ask the tutor to explain it drawing from the lecture transcript and the relevant reading. The explanation will use your course's terminology and approach rather than a generic one.

Use it for retrieval practice. Ask the tutor to quiz you on the last three weeks of material, or to generate practice questions from a specific reading. Testing yourself is the most effective study technique, and the tutor generates questions grounded in your actual content.

Ask it to find connections. "How does the concept from the week 2 reading relate to the case study from week 7?" is the kind of question that produces genuine insight, because the tutor can search across your entire library and surface connections you might have missed.

Annotate as you go. Your highlights and annotations on readings become part of the searchable library. When you're writing an essay and need "the passage about institutional bias in the week 3 reading," the search finds both the original text and your annotation about why it matters.

Frequently asked questions

What if the AI's synthesis is wrong even when the source is right? This can happen, and it's why the citation architecture matters. The student can always click through to the original passage and evaluate the AI's interpretation against the source text. A wrong synthesis with a visible citation is self-correcting in a way that a hallucinated answer without a source never is.

Can the AI access materials from other students' courses? No. Each student's library is private and separate. The AI only accesses the materials that specific student has uploaded to their workspace. There is no cross-contamination between students' course materials, and data is encrypted and never used to train the AI models.

What about copyrighted course materials? The materials remain in the student's private workspace, the same way a physical copy of a textbook remains in the student's room. The AI reads from the student's uploaded content to help them study; it does not distribute, reproduce, or make those materials available to anyone else.

Does the AI work with handwritten notes? If the student photographs their handwritten notes and uploads them, the system can process and index them. For the best searchability, typed notes or voice memos that are automatically transcribed are more reliable, but handwritten notes are supported.