How Does Canvas Detect AI? The Real Mechanics Behind the Score

How Does Canvas Detect AI in a Submission, Step by Step?

How does Canvas detect AI in a submission? The process begins the moment a student clicks submit, not when an instructor opens the assignment. If the assignment was created with a Turnitin-linked or similar LTI integration, Canvas sends the submission to the connected tool through LTI 1.3, the current version of the Learning Tools Interoperability standard that lets external applications operate inside an LMS as if they were native features. That handoff carries the submission's text content, the assignment ID, and enough metadata to route the result back to the correct student and grade column — but nothing about how the student wrote it. The external tool extracts readable text from whatever file format was submitted, queues the extracted text for analysis, and runs it through a trained classification model. Once scoring finishes, the tool sends the result back to Canvas through the same LTI connection using an outcome service call, and the score appears in the instructor's SpeedGrader view next to the plagiarism similarity report, typically within minutes for short submissions and up to an hour during high-traffic periods like finals week. Canvas does not store a copy of the analysis logic, does not perform any part of the scoring itself, and has no visibility into how the external model reached its conclusion — it is a courier for text going out and a score coming back.

1. Student submits a file or pasted text through the Canvas assignment page
2. Canvas launches an LTI 1.3 connection to the linked detection tool (commonly Turnitin)
3. The external tool extracts readable text from the submitted file
4. Extracted text is queued and run through the tool's classification model
5. The resulting score is sent back to Canvas via an LTI outcome service call
6. The score appears in SpeedGrader alongside the plagiarism similarity report

What Actually Happens During the Scan: Tokenizing, Scoring, and Aggregating Text

The scanning step itself is not a keyword search or a plagiarism-style match against a database — there's no fixed text to compare against because the writing might be completely original. Instead, the classification model breaks the extracted text into overlapping segments, often a few hundred words each, and evaluates each segment for statistical patterns associated with language-model output. Two properties drive most of these classifiers: perplexity, which measures how predictable each word choice is given the words before it, and burstiness, which measures how much sentence length and structure vary across a passage. Large language models tend to generate text with lower perplexity because they are trained to select high-probability next words, and with lower burstiness because their sentence rhythm stays relatively even. The model assigns each segment a probability score, then aggregates those segment-level scores into the single percentage that appears on the report, along with sentence-level highlighting that flags the specific passages driving the overall number. This is a trained classifier making a probabilistic judgment, not a lookup — which is exactly why the same underlying mechanics that catch AI-generated text also catch human writing that happens to share those statistical properties, like formal academic prose or heavily edited drafts.

"The model isn't matching your sentence to a database of ChatGPT output — it's asking how statistically unsurprising your word choices are compared to typical human variation." — NLP researcher describing classifier-based AI detection, 2025

Which Canvas Submission Types Can the Pipeline Actually Analyze?

Whether the detection pipeline can analyze a Canvas submission at all depends entirely on whether the tool can extract usable text from it. This is the practical answer to how does Canvas detect AI at the file-type level — it depends on what the classifier can read, not on what a student actually wrote. Typed text entries, pasted text, and standard document uploads — Word files, text files, and PDFs with a genuine text layer — extract cleanly and move through the pipeline without issue. Scanned PDFs and photographed handwritten pages are a different story: if the file is essentially an image with no embedded text layer, the detection tool has nothing to tokenize, and the submission passes through without any AI analysis unless the institution also runs optical character recognition first, which most standard Canvas-Turnitin integrations do not do automatically. Canvas quizzes built from multiple-choice, true/false, or matching questions produce no analyzable prose at all — there's no sentence-level text for a classifier to evaluate, which is why these question types sit entirely outside AI text detection regardless of what tool a school has connected. Short-answer and essay-style quiz questions can be analyzed if the institution has configured that integration, though this is less common than assignment-level detection. Code submissions run through most detection pipelines unreliably, since the classifiers are trained on natural-language patterns and code follows entirely different statistical rules. Group submissions get scored as a single document, meaning the pipeline has no way to attribute which portions came from which contributor.

1. Typed or pasted text entries — reliably analyzed
2. Word documents and text-based PDFs — reliably analyzed
3. Scanned or photographed pages without a text layer — generally skipped
4. Multiple-choice, true/false, and matching quiz questions — not analyzable
5. Short-answer or essay quiz responses — analyzed only if separately configured
6. Code submissions — passed through the pipeline but classified unreliably
7. Group submissions — scored as one document with no per-author attribution

How Does Canvas Detect AI Reliably — and Where Does It Fall Short?

Even when a submission is fully text-based and moves cleanly through the pipeline, several structural limits shape how much the resulting score can actually tell an instructor. Most detection tools require a minimum word count — commonly around 300 words — before they'll generate a score at all, because shorter passages don't provide enough statistical sample for the model to reach a stable probability estimate; submissions below that threshold typically return an 'insufficient text' notice rather than a percentage. Reliability also drops for non-English submissions, since most classifiers were trained primarily on English-language text pairs, and for content mixing multiple languages within a single document. Canvas itself does not natively track keystroke-level history or timestamped revision data for most assignment types, so the detection tool has no drafting timeline to compare against the final submission — it only ever sees the finished text, with no way to verify whether that text was typed over three hours or pasted in one motion, unless a separate proctoring extension is layered on top. The classifiers are also version-specific: a model trained to recognize patterns from one generation of AI writing tools can lag behind newer models that produce more naturally varied output, which is part of why detection accuracy shifts over time as both AI writing tools and detection classifiers keep updating. None of these limits mean the score is meaningless, but they do mean it is a probability estimate built on incomplete information, not a verified record of how a document was written.

Why Does the Same Pipeline Produce Different Scores for Similar Text?

Students sometimes notice that two passages that read as equally formal or AI-like to the eye come back with very different scores, and the mechanics explain why. The classifier scores statistical patterns at the segment level, so a document with a few AI-influenced paragraphs mixed into otherwise human-written sections can produce a moderate blended score rather than a uniformly high one — the aggregation step averages across segments, which means heavy editing of just the flagged portions can shift the overall number substantially even if most of the document stays untouched. This is part of why the question of how does Canvas detect AI doesn't have a single fixed answer — the same pipeline can return different numbers for similar text depending on timing, configuration, and which portions happen to fall into a flagged segment. Paraphrasing tools complicate this further: light paraphrasing that only swaps synonyms tends to preserve the low-perplexity, low-burstiness signature that triggered the original flag, while paraphrasing that restructures sentence order and varies length more meaningfully can lower the score without necessarily changing the underlying meaning. Institutional configuration adds another layer of variation — some schools set their integration to display a raw percentage, while others apply a threshold that only surfaces a flag above a certain cutoff, so the same underlying score can look different depending on how an instructor's Canvas instance is set up. And because the classifiers themselves get retrained and recalibrated periodically, the same exact text submitted months apart through the same institution's pipeline can return a measurably different score, simply because the model doing the scoring has changed.

What Happens to the Score After the Scan Finishes?

Once the score lands in SpeedGrader, the pipeline's job is done — everything after that point is a human decision, not a mechanical one. Canvas does not auto-flag, auto-fail, or auto-report anything based on the number; it simply displays whatever the connected tool returned, the same way it displays a plagiarism similarity percentage. Some institutions have set threshold policies where scores above a defined cutoff trigger an automatic notification to an academic integrity office, but that policy lives at the institution or department level, configured separately from the detection tool itself. In the absence of a threshold policy, interpretation is left entirely to the instructor, who typically weighs the score alongside other context: the student's prior writing samples, the nature of the assignment, and whether the flagged passages match patterns the instructor already associates with that student's voice. Because the underlying mechanics produce a probability rather than a verdict, most institutional guidance — including frameworks published by academic integrity offices since 2024 — treats the score as a starting point for a conversation with the student rather than standalone evidence of misconduct.

How Can You Check Your Writing Against the Same Mechanics Canvas Uses?

Since the detection pipeline is scoring statistical patterns rather than searching for a specific 'tell,' the most useful thing a student can do before submitting is see how their own draft performs against a similar kind of analysis. Running a draft through a detection tool that scores perplexity and burstiness at the sentence level shows which specific passages read as statistically uniform — the same passages a Canvas-integrated tool would likely flag — while there's still time to revise before the deadline. NotGPT's text detector applies this kind of sentence-level analysis and highlights exactly which portions are driving the overall probability score, which is more useful than a single blended percentage because it shows where to focus. If a flagged passage genuinely reflects your own formal writing style rather than AI assistance, the Humanize feature can adjust sentence rhythm and phrasing at Light, Medium, or Strong intensity to reduce the statistical uniformity that triggers false flags, without changing the substance of what you're saying. The goal isn't to game a score — it's to understand, before an instructor does, which parts of a submission share the statistical fingerprint these classifiers are built to catch.

1. Finish your draft with enough time to review it before the Canvas deadline
2. Run the full text through a sentence-level AI detection tool
3. Note which specific passages score highest rather than only the overall percentage
4. Check whether flagged passages are heavily edited, highly formal, or unusually uniform in sentence length
5. Revise flagged sections for natural variation, or use a humanizing tool if the flagged style is genuinely your own
6. Re-check the revised draft before submitting through Canvas