Pangram Works With Canvas Through an Institutional License Set Up by Its Education Team

Both, and the license decides which. Pangram's knowledge hub names Canvas among six supported platforms. It says detection runs automatically on submissions and the results appear in the existing grading workflow, so instructors don't change how they work ¹. Those integrations, Pangram says, are only available via institutional license ⁹.

The knowledge hub's full list runs Google Classroom, Canvas, Schoology, Moodle, Blackboard and Brightspace, with more promised ¹. Canvas turns up again on the pricing page, under the Educational Institution tier, next to Brightspace, Moodle, Google Classroom and Google Docs ³.

Without a license there are two routes, both Pangram's own. The free dashboard takes pasted text or uploaded documents, up to 2,000 words a day on the free tier ³. The Chrome extension checks selected text from a right-click menu, and Pangram says it works anywhere on the internet, Canvas SpeedGrader included ⁹. Pangram describes neither as writing a score into Canvas.

The route matters later. A reading an instructor takes in a personal account isn't held under the institution's contract, and the school may have no copy of it until somebody forwards one. Whether it becomes an education record depends on what is kept and by whom. The integration has the opposite problem: it runs automatically on submissions across the course, so scores arrive for papers nobody had a question about.

Instructure's own post presents AI writing reports in Canvas as part of Turnitin's integration, built on the LTI 1.3 standard ⁶. It describes no detection built into Canvas itself. On that reading, a score in a Canvas grading screen comes from a tool the institution connected, under a contract somebody signed.

Pangram's education team does. Pangram says its team handles the technical connection and the rollout with the institution's LMS administrator ². It starts with the enterprise contact form ¹. There's no published price. LMS checking sits inside the Educational Institution license, and that tier asks for an enquiry where the individual and professional tiers print a monthly figure ³.

What the tier includes, in Pangram's own wording, is "Unlimited AI checks and plagiarism checks integrated directly into the LMS" ³. Unlimited is doing real work in that sentence. The published individual plan meters words instead: $20 a month buys 300,000 of them, and the professional tier is $65 a month for 1,500,000 ³. A department buying seats would work inside a monthly word allowance, for a volume that's hard to forecast in week 1 of a term.

That leaves two things for whoever owns the budget line. First, the number arrives after a conversation, so it can't be set against a rival rate card without asking both vendors for a quote; what a school actually pays for Pangram starts there.

Second, timing. An integration built jointly with an LMS administrator belongs at a term boundary, and a request from one department in week 6 is unlikely to move that date.

By Pangram's account, a score in SpeedGrader, though the page's own word for it is plagiarism. Pangram's integrations page says the Canvas integration "automatically runs AI detection in the background and provides plagiarism scores directly in Canvas SpeedGrader" ⁸. The knowledge hub adds that results appear in the existing grading workflow, so instructors don't have to change how they work or ask students to do anything new ¹. Neither page says in so many words where the AI reading appears.

That's where the cited pages stop, and the gaps are the questions that decide how a flag behaves in practice. Where does the AI reading appear? Can the student see the score in the submission view? Can anyone above the instructor? Is it still there after the grade is posted? Can it be limited to chosen courses or assignments? A vendor demo can show all of these, and a contract can fix them in writing.

What the detector reports is better documented. Pangram's chief executive told Nature that a work labeled 100 percent AI does not mean every word was machine-written ⁵. So a percentage on the grading screen describes the text. It says nothing about the student who handed it in.

For an administrator the trouble shows up weeks later, in a meeting, when two people read one number differently. A detector score alone cannot carry a finding, and a score that lands in the grading screen automatically is easy to read as though it can.

Pangram states that it is FERPA compliant and SOC 2 Type 2 certified, that student submissions are not used to train external models, and that institutions can configure retention so data is deleted after analysis or kept for audit and integrity review ⁴. Those are vendor statements. In the United States the obligations FERPA sets fall on the institution ⁷. What follows describes the federal regulations as they stood in 2025 and is not legal advice.

Under FERPA, an education record is one directly related to a student and maintained by the school, or by a party acting for it ¹⁰. A submission a school keeps through an LMS integration will meet that definition in most cases.

The usual route for an outside provider is the school official exception at 34 CFR 99.31, though consent is another ⁷. Pangram's pages don't name the route it expects. Among the exception's conditions, an outside provider has to do a job the school would otherwise do with its own employees, stay under the school's direct control over the use and maintenance of education records, and observe the limits on redisclosure.

Contract language is where a school shows those conditions are met. A badge on a marketing page shows nothing.

Two clauses earn their space in the agreement.

Retention first. Pangram offers a choice between deleting data after analysis and storing it for audit and integrity review ⁴. Make that choice on purpose. A parent or eligible student may ask to inspect education records, and the school has to comply within a reasonable period, no later than 45 days ¹¹.

Training second. The higher-education page says both that submissions are not used to train external models and that Pangram never trains its models on student papers ⁴. The pricing page's institutional tier says the company does not train on student data ³. Those are clear public statements, but the version a school can enforce for itself is the one in the signed agreement.

The FERPA questions a detection contract has to answer are the same whichever vendor wins the bid. A Canvas integration raises the stakes because it runs automatically on submissions, where an instructor checking by hand picks a handful.

A pilot answers what no vendor page can: how many flags land in one department's real coursework over a real term. Pick two courses with different writing loads and run the integration on both. Then count the readings that led to a conversation against the ones that led nowhere.

One term is long enough to capture three numbers: how many submissions were scored, how many scores an instructor acted on, and how many of those actions held up after a conversation with the student. Only the third says whether the purchase changed anything. Pangram's pages describe no dashboard that reports it.

At least one university is ahead. Nature reports that the University of Chicago has started using Pangram to vet students' coursework ⁵. The report doesn't say whether that runs through an LMS. A reference call with a school already using the tool is worth asking the vendor for alongside a demo. Nature also reports that the NeurIPS conference rejected 18 percent of submissions after screening them with Pangram ⁵. That's the tool in use at scale, though on conference papers and outside any LMS.

Set the escalation rule before the first flag lands. The questions worth putting to a detection vendor in writing include who sees a score, how long it's kept, and what the institution is expected to do on the day it's wrong. Pangram publishes answers to some of those and not to others. That's a reason to ask, and no reason to walk away.