Does Canvas detect AI? What's actually running under the hood

The short answer

No. Canvas — the LMS itself — does not detect AI writing. As of this writing, no Instructure documentation describes a native AI-detection feature anywhere in Canvas, including SpeedGrader and New Quizzes. When an "AI score" appears on a submission inside Canvas, it was produced by a third-party tool plugged into Canvas through its integration platform, and shipped back into the gradebook view.

That distinction sounds like pedantry. It is actually the most important fact about the score — because it tells you whose documentation governs what the number means, whose error rates apply, and who you should be quoting when a dispute starts.

How detection actually gets into Canvas

Instructure's own developer documentation is unambiguous about the architecture. The Canvas "plagiarism detection platform" exists, in its words, to provide "a standard way for LTI2 tool providers (TPs) to seamlessly integrate plagiarism detection tools with Canvas" — with newer integrations moving to the LTI Asset Processor ("Document Processors") model. In other words: Canvas built the socket, and vendors like Turnitin and Copyleaks build the thing that plugs into it.

The accepted answer on Instructure's own community forum says the quiet part plainly: "Canvas itself does not check for plagiarism or provide a similarity score, rather that is done by 3rd party integrations with Canvas (like Turnitin or Copyleaks)… I think it's unlikely that Instructure would provide any kind of AI checking itself."

So the flow behind a flagged essay is: student submits in Canvas → Canvas hands the file to the integrated vendor → the vendor's model scores it → the score renders inside SpeedGrader as if it were part of Canvas. The LMS is the picture frame, not the painter.

Why the distinction matters when a score appears

The score inherits the vendor's caveats, not Canvas's authority. A number displayed inside official university software borrows an air of institutional fact. But if the integration is Turnitin, the number comes with Turnitin's own published fine print: the under-1% false-positive claim is scoped to documents with over 20% AI writing; scores below 20% are now shown as an asterisk because they are "less reliable"; the model skips lists, tables, references, and code; and — the vendor's own words — the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors." We walk through every one of those caveats, with quotes, in what your Turnitin AI score actually means.

"Canvas flagged it" is never an accurate sentence. In an integrity meeting, in a report, or in a syllabus, name the actual tool. It forces precision about which product, which model version, and which documented limitations are in play — and it avoids implying the LMS independently corroborated anything. Two tools in the same Canvas instance can disagree about the same essay; "Canvas" didn't say anything.

Your institution chose this, and can un-choose it. Because detection is an integration, it is a switch an administrator controls — which is exactly how a growing list of universities turned AI detection off without replacing anything else they run. Disabling a detector is not ripping out your LMS; it is unplugging one appliance.

What about quizzes, proctoring, and "Canvas knows"?

Students trade folklore about what Canvas can see, and instructors sometimes inherit it. The architecture above is the corrective: capabilities live in the specific tools your institution licensed, not in Canvas as such. If a proctoring tool is installed, its powers are in that product's documentation. If no AI detector is integrated, nothing in a standard essay submission is scoring it for AI — there is no latent detector waiting in the LMS. Policy conversations go better when they start from what is actually installed, not from what the software is rumored to do.

If you're the instructor reading the score

Three habits keep an integrated score from becoming a procedural problem:

1. Find out what is actually integrated in your instance, and read that vendor's documentation — particularly its stated false-positive behavior and its instruction about scores being the sole basis for action. The vendor's caveats are the floor of fair use, not optional reading.
2. Treat the score as a reason to look, never a finding. The strongest tool you have is not in the LMS at all — it is a well-run conversation about the work, backed by the student's process evidence.
3. Decide your policy before the score arrives. The alternatives to AI detectors — assignment design, baselines, process artifacts, and consent-based provenance — all work better than a score precisely because they don't depend on a vendor's error bar.

The bottom line

Canvas doesn't detect AI; a product plugged into Canvas might. The number on the screen belongs to that product — with that product's error rates, that product's fine print, and that product's own warning not to act on the score alone. Name the tool, read its label, and build your process on evidence that would survive without it.