Canvas AI Detection 2026: What Canvas Actually Detects (and What It Cannot)

The Short Answer: Canvas Has No Native AI Detection

Canvas does not detect AI writing on its own. Instructure, the company that builds Canvas LMS, ships zero AI-classification logic inside the core product. There is no checkbox in the admin panel labeled “AI detection.” There is no built-in algorithm that reads your submission and assigns it an AI-likelihood score.

Every AI flag you have ever seen inside a Canvas course came from somewhere else. An external detector vendor, integrated via the LTI (Learning Tools Interoperability) standard or installed as a third-party scaffold, ran the scan and pushed the result back into the Canvas SpeedGrader interface. The score you see in the rubric was produced outside Canvas and displayed inside it.

Use Case Disclosure: Who This Guide Is For

Students checking before submission. You have an essay drafted, the assignment uses Turnitin or Copyleaks, and you want to understand what the detector will see before you submit.

Educators evaluating institutional policy. You are deciding whether to enable AI detection on your assignments, or your dean is pushing back on a flag.

ESL and non-native English writers facing false positives. Stanford 2023 (Liang et al., DOI 10.1016/j.patter.2023.100779) found that detectors flagged 61.3% of authentic TOEFL essays as AI-written.

To be explicit about scope: if you are looking to submit AI-written essay you did not write to a class that prohibits AI, that is academic dishonesty and this guide is not for you. The rest of this page assumes you are working under your institution's rules in good faith.

What AI Detector Does Canvas Use? The Three Integration Paths

Path 1: Turnitin LTI. A learning-tool-interoperability integration that Turnitin maintains as a Canvas Partner Listing. The institution licenses Turnitin separately, the IT administrator enables the LTI at the account or sub-account level, and individual instructors then toggle “Turnitin enabled” on each Canvas assignment. This is the dominant path: the large majority of US higher-ed Canvas deployments run Turnitin LTI.

Path 2: Copyleaks LTI. A similar partner-listing integration that Copyleaks maintains. Slightly more common at enterprise and corporate-training Canvas instances than at undergraduate-focused universities.

Path 3: GPTZero Scaffold. A real LTI integration, built with K16 Solutions, that an administrator enables in the Canvas app store before instructors switch it on per course. Structurally the same two-step pattern as the other two; what differs is how many schools have taken it up, which is far fewer.

Path 1: Turnitin LTI, the Most Common Integration

Turnitin LTI is the heaviest integration in the Canvas ecosystem. The August 2025 Turnitin update introduced a layered classifier that combines three signals: a perplexity baseline, a burstiness signature, and a lexical-substitution fingerprint trained against the public outputs of major paraphrasers. Vendor accuracy on raw GPT output runs 94 to 97 percent.

LTI prerequisite explainer. Turnitin LTI runs at the assignment level, not the course level. If you do not see a “Turnitin enabled” line in the submission agreement when you click submit, that assignment was not flagged for AI scanning.

File type coverage. Turnitin LTI scans .docx, .pdf, .txt, and .rtf reliably. Image-based PDFs are not OCR'd by the AI classifier and return no score. This is sometimes confused with “passing”: it is not passing, it is unscannable.

For a deeper Turnitin-specific reference outside of Canvas, see our Turnitin AI Checker guide. For pre-submission, the Turnitin walkthrough and the Turnitin AI feature deep-dive.

Path 2: Copyleaks LTI, the Enterprise Challenger

Copyleaks pushed a major detector update in February 2026 (the V9 release) that introduced two parallel modules: AI Insights (a section-by-section likelihood reading) and AI Logic (a Bot Detection sub-module targeting paraphrased prose).

Copyleaks V9 flags individual paragraphs, not just whole submissions. If your essay reads at 11% overall but one introduction paragraph reads at 84%, the rubric will surface the paragraph-level flag. For pre-submission, see our Copyleaks walkthrough.

Path 3: GPTZero Scaffold, the Manually Configured Option

GPTZero reaches Canvas through an LTI app built with K16 Solutions: an administrator enables it in the Canvas app store, then instructors activate it per course or per assignment. That is the same shape as Turnitin and Copyleaks, so the difference is adoption rather than architecture. South Piedmont Community College runs it at scale, and Instructure hosts the case study on its own site, which is about as clear a signal as a named deployment gets. What is true is that far fewer schools have bought it, so a single professor having it while the course down the hall does not is a common pattern. The classifier itself runs the v6 (January 2026) architecture, combining lexical predictability cones with a burstiness layer.

For background on what the v6 classifier flags, see our GPTZero walkthrough. A note: Originality AI does not have a Canvas LTI as of May 2026, but instructors sometimes paste student text into it for a manual second-opinion check. See our Originality AI reference for the three-variant breakdown.

Canvas Plagiarism Checker vs AI Detection, and What Canvas Cannot See

Canvas Quizzes are not routed to LTI AI detectors. Quiz responses are stored in the Canvas QuizSubmission object, which is a separate data model from the Assignment object that Turnitin LTI and Copyleaks LTI hook into. As of May 2026, no detector vendor has shipped a first-party Quizzes integration.

Canvas Studio video submissions have no AI detection. Studio video submissions do not pass through Turnitin LTI or Copyleaks LTI because video is not the input format those classifiers were trained for.

Discussion posts are usually not scanned at all. Turnitin documents discussion coverage as its own feature, called Discussion Boards, and it is enabled separately from assignment checking. Most institutions configure assignments and stop there, which leaves discussion replies sitting in Canvas as plain DiscussionEntry records that never pass through any detector. An instructor can still copy a reply into a checker by hand; nothing does it automatically.

The practical takeaway: a Canvas course can have AI detection on its essay assignments but no detection on its weekly quiz prompts, discussion replies, or video reflections.

Can Canvas Tell If You Screenshot or Copy and Paste?

This is a different worry from AI detection and it deserves a straight answer, because the guesswork circulating about it is worse than the facts. Screenshots first: Canvas cannot see them. Taking a screenshot is something the operating system does, and the operating system never tells the browser it happened, so no website can detect one. That is not a gap in Canvas, it is how browsers work. Proctoring products like Respondus LockDown Browser, Proctorio and Honorlock are a separate purchase that can block or flag screen capture, but they are not Canvas and Canvas does not include them.

Copy and paste splits by where you are typing. In a normal assignment submission or a discussion reply, the text box is an ordinary browser field with no clipboard monitoring, so pasting is invisible. Canvas Quizzes behave differently. The quiz log records how fast an answer appeared, so twelve hundred words materialising in three seconds shows up as an anomaly. Note what that actually is: an inference from typing speed, not a record of your clipboard. Canvas never sees what you copied or where it came from.

Tab switching splits the same way, and this is where most of the rumours live. While you are simply reading course pages, Canvas cannot see other tabs or other sites. The University of Michigan’s own IT communications call that rumour flatly false. Inside an active Canvas Quiz it is different: the quiz log records each time the tab loses focus, and for how long.

What matters more than any of that is how little those logs are worth as evidence, and Instructure says so itself: its own guidance is that quiz logs should not be used to validate academic integrity or identify cheating. Yale’s Poorvu Center puts the same point plainly, that Canvas was built to support teaching rather than to catch dishonesty. The reason is failure rate. In 2021 Dartmouth’s Geisel School of Medicine accused seventeen medical students of cheating on the strength of Canvas log data and later withdrew the accusations after outside experts took the analysis apart. In a separate case at James Madison University a student’s log recorded him leaving the exam page ninety-nine times when he had not left it once, because notifications and background apps steal focus and the log cannot tell the difference. If a log is being read as proof of something you did not do, those two cases are worth citing by name.

How to Pre-Check Before Submitting to Canvas

1. Identify which detector your Canvas course uses. Open the assignment in Canvas. Click “Submit Assignment.” Read the submission agreement. If it mentions “Turnitin,” Path 1. If “Copyleaks,” Path 2. If “GPTZero,” Path 3.
2. Paste your draft into the free detector. Open humanizemy.ai/detect in a new tab. The free tier allows 4 scans per day at 250 words per scan, and no account is needed.
3. Read the score against the integration threshold. Under 10% reads as within the normal range. 10–20% is the gray zone. 20–40% is the range where instructors typically open a review conversation. Above 40% the instructor will almost always raise the question.
4. If flagged, write more naturally and re-check. If you used an AI drafting tool earlier and need the final draft to read more naturally, our corpus-trained HumanizeMyAI humanizer runs on a free account: four rewrites of up to 250 words, granted once rather than daily, and no card is asked for. The Basic plan at $18/month lifts the per-run cap to 1,000 words.
5. Re-run /detect to confirm and submit. Confirm the score has dropped into the range you are targeting. Then submit to Canvas.

ESL Writers and False Positives in Canvas-Integrated Detectors

Liang et al. (2023), published in Patterns, evaluated seven commercial GPT detectors against authentic TOEFL essays. The headline result: detectors flagged 61.3% of authentic TOEFL essays as AI-generated. A much lower native-speaker rate also circulates, but it appears only in the preprint version of the study, so the DOI-backed 61.3% is the number this page relies on. For a deeper treatment, see our ESL detector bias analysis.

Four-university institutional precedent. Vanderbilt University Center for Teaching disabled Turnitin AI Detection in August 2023. Yale's Poorvu Center for Teaching and Learning published guidance against using AI detection scores as primary evidence. Waterloo's Office of the Associate Vice-President, Academic discontinued Turnitin AI detection for all users effective September 2025. Curtin University disabled Turnitin AI detection across all its campuses in January 2026. These four institutions form a defensible precedent chain that an ESL writer facing a flag can cite when requesting manual review. UC San Diego is a narrower case: its Extended Studies continuing-education division deactivated Turnitin's AI indicator on April 7, 2025, while the campus otherwise leaves detector decisions to individual instructors.

Measured Results: HumanizeMyAI Across All Three Integration Detectors

Taken on 31 August 2026, by pasting HumanizeMyAI output into each detector’s own interface. Two of the six answer in words rather than digits: Turnitin displays nothing below 20%, and the free Originality AI tier caps out at a 15%-or-less band, so neither one hands you a point estimate. For a broader view of which humanizer tools deliver detector-passing prose, our Best AI Humanizer reference covers the cross-detector matrix for eight tools.