Does Canvas Detect AI? What the LMS Really Sees

In August 2023, Vanderbilt switched off Turnitin's AI detector across its whole campus, having worked out that even a 1% false-positive rate would wrongly flag roughly 750 of the 75,000 papers it processes each year. That decision was made by a university, not by a learning management system. The distinction sounds like a technicality and turns out to be the entire answer to this question.

Canvas does not detect AI. It is a learning management system: it stores submissions, timestamps them and logs access. AI detection appears only when your institution enables a third-party integration such as Turnitin, so what your instructor sees depends on choices made by administrators and IT staff, not by Canvas itself.

Does Canvas detect AI writing on its own?

No, and it is worth being precise about why, because the reason predicts everything else in this article.

An LMS is a distribution and record-keeping system. It hands out assignments, collects files, stores grades, routes announcements and keeps an audit trail of who touched what. Nothing in that job description involves analyzing the content of a submission. Canvas does not read your essay any more than a filing cabinet reads the folders in it.

What Canvas does have is an extension mechanism. Institutions connect external tools through standard integration points, and some of those tools do analyze content. When an AI percentage shows up next to a student's name, it arrived from one of them and Canvas simply displayed it. The window is not the weather.

What Canvas actually records

Canvas keeps a reasonable amount of activity data, and this is the part that generates the most folklore. In broad terms, and depending on what your institution has configured, an instructor or admin can see when a submission arrived, whether it was late, when course pages were accessed, how many times a file was resubmitted, and interaction data from quizzes including per-question timing.

None of that is authorship evidence. Here is the honest mapping:

The submission-time myth is the most persistent of these and the easiest to dismantle. Students worry that turning in an essay four minutes after opening the assignment page looks damning, and instructors sometimes read it that way. But almost nobody drafts inside the LMS. The normal workflow is to write somewhere else for days and then spend ninety seconds uploading a file, which produces exactly the same log entry as pasting in something you generated on the walk to class. The signal cannot distinguish between those two stories because it never observed either of them.

Read the right-hand column and you have most of the arguments that go wrong in integrity meetings. The gap between "the system recorded an event" and "the event means what I think it means" is where students get hurt and where instructors overreach in good faith.

How detection actually gets into Canvas

Three layers have to line up before an AI score reaches a gradebook.

First, the institution licenses a detection product. That is a procurement decision made centrally, usually annually, and it costs real money. Second, an administrator enables the integration and sets defaults, including whether the AI indicator is visible at all, which assignment types it runs on, and which courses inherit it. Third, an instructor turns it on for a specific assignment, or accepts the default that was set for them.

Any of those three can be a no. A university can hold a license while an administrator hides the AI indicator, exactly as several institutions chose to do after 2023. A department can enable it while an individual instructor never checks the box. This is why "does Canvas detect AI" has no single answer even within one campus: the honest answer is that it depends on three decisions you cannot see from the student side.

There is a fair objection to all this: if the score lands in Canvas and the consequence lands on you, who cares which box produced it? The distinction feels pedantic from the receiving end.

It isn't, for three practical reasons. The error rates you can cite belong to the detector, not the LMS, so knowing the tool tells you which published numbers apply to your case. The threshold that triggered a referral was set by your institution, so knowing the policy tells you what the flag was supposed to mean. And the appeal route runs through academic integrity, not through IT, so knowing the chain tells you who can actually change the outcome. A student who says "Canvas flagged me" has already lost track of all three.

It also explains a common confusion. Similarity checking and AI detection are separate functions of the same product, and they can be switched on independently. A similarity report full of green does not mean an AI check ran, and an AI percentage does not mean anything was matched to a source. Our breakdown of what Turnitin does and does not catch walks through that split in detail.

What your instructor sees in the gradebook

In practice, a grader working through SpeedGrader with an integration enabled sees the submission, a similarity percentage, and, if the indicator is visible, a separate AI writing percentage with a link to a report showing which passages were flagged. They also see the submission time and any late marker.

They do not see your drafting process, your revision history, which tools you had open, or where the text came from before it entered the box. If they want that, they have to ask you for it, which is exactly why the strongest thing a student can do is arrive with a record already in hand. A Google Docs revision timeline answers the question the gradebook cannot, and we cover that in our piece on what Google Docs actually records.

The asymmetry is stark when you lay it out. Your instructor has one number produced by a probabilistic tool. You have, potentially, a complete account of how the document came to exist. Most students never realize they are holding the better hand.

Watch it play out. A community college instructor teaching two sections of composition opens SpeedGrader and sees a 71% AI indicator on a 900-word literacy narrative submitted eleven minutes before the deadline. Everything about the display invites a conclusion: high number, late submission, formal prose. He emails the student, who replies with a link to the document and an offer to walk through it. The revision history shows nine days of intermittent work, a paragraph rewritten five times, and an entire opening section abandoned on day four. The conversation lasts a quarter of an hour and ends with no referral.

Nothing in that story required either party to be an expert on detection. It required the instructor to ask before acting and the student to have kept a record. Both of those are habits, not skills.

Why institutions turn detection on and off

The Vanderbilt arithmetic from the top of this article is worth sitting with, because it is the clearest public example of an institution doing the math rather than the marketing. One percent sounds like a rounding error. Against 75,000 papers it is about 750 students, most of whom did nothing wrong, each of whom would have to be investigated by people whose time is finite and whose judgment is not calibrated for statistics.

Add the evidence that arrived alongside it. The Washington Post tested the same detector in April 2023 on 16 mixed student samples and got over half at least partly wrong. Turnitin itself has claimed under 1% false positives at the document level while disclosing roughly 4% at the sentence level, which matters enormously when a report highlights three sentences in a 2,000-word essay. And the RAID benchmark showed commercial detectors degrading sharply under paraphrase attacks, meaning the students most likely to be caught are the least sophisticated ones.

Set against that, some institutions concluded the tool was worth keeping with strict guardrails, and others concluded it was not worth the false accusations. Both are defensible readings of the same evidence. Neither is a fact about Canvas.

Find out what your school actually enabled

Stop guessing. One email to your instructor or department administrator resolves it, and asking before there is a problem is much easier than asking during one:

Could you tell me whether this course uses an automated similarity or AI writing check on submitted work, and if so which tool? I would also like to know whether the AI indicator is visible to graders and what the department's policy is on how those results are used. I am asking so I understand the process, not because there is an issue.

Then do the second half, which is building your own record. Draft in a versioned editor, keep your notes, and check your own text before you submit. AI Detector 360's AI essay checker returns a sentence-level heatmap and an explicit confidence level rather than a single percentage, so you can see whether one formal paragraph is carrying the whole score, and the downloadable PDF report gives you a dated artifact to keep with your drafts. The free scanner covers 5,000 characters with no sign-up and five scans a day.

If a flag does land, the sequence for responding is procedural and worth knowing in advance. Our step-by-step appeal guide covers requesting the evidence, assembling the timeline and escalating when the process itself goes wrong.

What nobody outside your IT department can tell you

Anyone who tells you what Canvas sees at your specific institution is guessing, including us. Integration settings, visibility of AI indicators, referral thresholds and retention of log data are all local configuration choices, and they are not published. They also change between semesters without announcement, because a license renewal or a policy review can flip a default that nobody communicates downward.

Two things follow. Never generalize from a friend at another university, or even from a different department at your own. And treat any confident claim about "what Canvas can see" from a forum post as folklore until an administrator confirms it.

Whatever your school runs, you can always get an independent read of the same text. AI Detector 360 scores with multiple engines and surfaces disagreement between them rather than averaging it into a single confident-looking number, which is usually the most informative thing you can bring to a conversation about a flag.

What holds everywhere is the structural point: the LMS is a window, the detector is a separate licensed product with published error rates, and the policy governing both is your institution's. Getting those three straight in your head is worth more than any rumor about what the gradebook knows.