Can Canvas Detect ChatGPT? What Canvas Actually Runs

Canvas does not detect ChatGPT on its own. Instructure built Canvas as a learning management system, and its built-in plagiarism check runs through third-party integrations your institution switches on. If your submission gets an AI score, that number came from Turnitin, or from a detector your school wired into the assignment — not from Canvas itself. What Canvas shows you is the report a detector handed back, attached to your submission.

That distinction matters, because it changes what you can do about a flag. You are not arguing with Canvas. You are looking at a probability estimate from a detector, and you can work on the sentences that carried the signal before anyone reads them.

Key takeaways

• Canvas is the delivery system, not the detector. Turnitin or a similar integration does the scoring, and Canvas displays the result.
• A flag is a probability estimate, not proof of authorship. Instructors are told this too, though not all of them act on it.
• Detection runs at sentence level in most modern tools, so a whole-document score hides which lines caused the flag.
• Surface habits — uniform sentence length, hedged transitions, tidy paragraph shapes — push a score up even when the ideas are yours.
• You can fix the flagged sentences without rewriting the argument, and you can verify the meaning survived.
• Our own measurement of 46 hand-written DeAIze guides scored a mean AI score of 27%, with 17 of 46 flagged at a 30% threshold. Human writing flags too.

How does Canvas actually handle AI detection?

Canvas itself does not ship a ChatGPT detector. What it ships is an assignment workflow with an external check attached. Your instructor enables a plagiarism and AI service on the assignment; when you submit, Canvas passes the file to that service and displays the returned report inside SpeedGrader.

The common integration is Turnitin, which has its own AI writing indicator. Turnitin’s report shows a percentage of the submission its model considers likely AI-generated, plus a breakdown across segments of the text. Other schools use different vendors entirely, and some use none.

If you want the fuller picture of how those scores are produced, our guide to how AI detection works walks through the signals. The short version: detectors look at how predictable each word is given the words before it, and how much the sentence lengths and structures vary. Flat, predictable prose scores higher.

Why does Canvas show an AI score on my own writing?

Because the detector is measuring statistical regularity, not authorship. Text that moves in even, well-behaved sentences with predictable transitions looks the same whether a person wrote it carefully or a model produced it.

That is not a hypothetical. In a first-party measurement we ran on 2026-09-19, we scored 46 hand-written DeAIze guides with our own detector at a 30% flag threshold. The mean AI score was 27%, the median 25%, and 17 of the 46 were flagged. Every one of those documents was written by a person, start to finish.

We also scored 12 unedited model outputs from fixed prompts. Those had a mean AI score of 35% and a median of 34%. Notice what that means: the gap between careful human writing and raw model output was narrower than most people assume. Detector output overlaps. That is the honest picture, and it is why a single number should never be treated as a verdict.

If you want to see which specific lines are carrying the signal rather than guessing from one number, run the draft through the AI detector — it scores the document and highlights each sentence by AI probability, so you can edit the lines that matter instead of rewriting everything.

What does the Turnitin AI indicator inside Canvas actually report?

Turnitin’s AI writing indicator reports an estimated percentage of qualifying text that its model considers AI-generated, and it segments the submission so you can see where. It does not report a named author, and it does not report intent.

A few practical consequences:

1. Short submissions produce unstable readings. A few hundred words gives a detector very little to work with, and the score swings more.
2. Mixed documents produce mixed segments. If you drafted with a model and then edited heavily by hand, you may see one segment flagged and the rest clean.
3. Quoted and cited material can still be scored. The detector does not reliably exclude block quotes or reference lists from its segment analysis.
4. The percentage is not a percentage of your essay that is fake. It is a model estimate over qualifying text. The wording on the report is narrower than the way people repeat it.

If you have already received a report and need to respond to one, our walkthrough of the Turnitin AI report covers what each part says and how to reply to an instructor without sounding defensive.

Which sentences push a Canvas AI score up?

Mostly the boring ones. Here is a comparison of the patterns that carry signal, alongside the fix that costs you nothing in meaning.

Pattern that raises the score	Why it reads machine-made	What to do instead
Every sentence 18–25 words long	Uniform length is a strong statistical regularity	Break one sentence in each paragraph into two, or fuse two into one long one
Transition openers on every paragraph	“Moreover”, “Furthermore”, “Additionally” in sequence is a template	Delete the opener and let the first clause carry the link
Balanced three-part lists	Triadic structures are a model habit	Cut to two items, or expand one item into a full sentence
Hedge stacking	“It is important to consider that there may be”	Say the thing. “Consider X.”
No concrete nouns	Abstract summary with no names, dates or objects	Add the specific tool, file, or number you actually used
Identical paragraph shapes	Claim, support, restate — three times over	Vary it. Start one paragraph with an example

None of these are moral failures. They are editing problems, and they are fixable in a single pass.

Can I fix the flagged sentences without rewriting my argument?

Yes, and that is the right scope. Rewriting from scratch risks changing your claims, which is a worse outcome than a flag.

The workflow that holds up:

1. Score the draft first. Get a baseline so you know which sentences are actually carrying signal. Editing blind wastes effort.
2. Rewrite only the flagged sentences. Leave the clean ones alone. If a paragraph scores well, touching it can only make it worse.
3. Check that the meaning survived. This is the step people skip. A rewrite that reads better but shifts your claim is not an improvement.
4. Re-score and compare. If the number did not move, the rewrite did not address the actual signal.

On the workflow question, our draft editor runs this as a closed loop: it scores, rewrites only the flagged sentences, keeps a rewrite only when semantic similarity says the meaning held, then re-scores and shows you before and after. On the example paragraph on our site, a 92% AI score becomes 4%. The site reports an average AI-score reduction of 60%+ after humanizing, with a best case of 71% — those are our own predicted scores, not an official verdict from Turnitin or anyone else.

That caveat is not modesty. It is the actual limit of the tool. No humanizer controls what Turnitin returns.

Does humanizing count as cheating?

This is the objection worth taking seriously, so here it is without a sales pitch.

If you generated the ideas, the analysis and the conclusions, and a model helped you phrase them, tightening the prose is editing. Every writer does it. Grammarly does a version of it. That is the case for fixing surface patterns.

If you generated the argument and are trying to make a submission pass as your own work, that is a different act, and no tool changes what it is. Detection scores are not the thing being violated. Misrepresenting authorship is.

Most academic integrity policies address this directly, and the wording varies more than students expect. Our breakdown of what your university actually allows covers the common structures, and the academic integrity hearing guide explains what panels look at when a case goes that far. Read your own policy. It is usually shorter than the rumour mill suggests.

What should I do before I submit?

A short routine, in order:

1. Read the assignment’s detector settings. If the instructor enabled an AI check, you know which report will appear next to your name.
2. Score your own draft before submitting. Finding out privately beats finding out in SpeedGrader.
3. Fix the flagged sentences, not the whole document. Preserve the passages that already score clean.
4. Re-read the rewritten sentences aloud. If you stumble, a reader will too.
5. Keep the version history. If you drafted with a model and edited by hand, that history is your evidence, and it is the most useful thing you can bring to a conversation with an instructor.
6. Be ready to explain your process. “I drafted with a model and edited it heavily” is a defensible sentence. “I don’t know why it flagged” is not.

The fear underneath this search is usually not about the tool. It is about being accused of something you did not do, with no way to prove it. The defence is documentation plus clean prose, and both are things you control.

Ready to see which sentences read machine-made?

You do not need to guess which lines caused the flag. Run the draft through the AI detector and look at the sentence-level scores, fix the ones that carry the signal in the draft editor, then check the before-and-after numbers. If you are working through a backlog of documents, pricing is pay-as-you-go with credits that do not expire, and the free tier gives you 200 words on signup with no credit card. When you are ready to start, create an account and test it on the draft that is worrying you.

Text is processed for the current task only, never used for training and never resold. Your drafts stay yours.