Are universities using less AI than we think? Canvas data only sees part of the picture
Instructure's 29 September study covers 19,480,990 US higher-education users. Dedicated AI tools missed its top 100, but the measurement excludes use outside Canvas integrations.
The short version
- What happened
- LTI activity measures formal integrations, not all student or lecturer AI use.
- Who it affects
- Students need clear support and accessible alternatives regardless of which software they use.
- What remains unproven
- Canvas users are not a representative sample of every university worldwide.
Keep in mind
Canvas users are not a representative sample of every university worldwide.
Company-authored analysis excludes off-platform and native API activity.
Source analysis · 2 linked sources. Primary analysed source date: 29 September 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The population behind the ranking
Instructure announced its first higher-education EdTech Top 40 on 29 September 2026. The study uses de-identified Canvas Learning Tools Interoperability launches from 19,480,990 people at US higher-education institutions between 1 September 2025 and 30 April 2026. No dedicated AI tool reached the top 100. Google Gemini reached nearly 88,000 users across 211 institutions, according to the company.
The important boundary is the measurement channel. LTI launches reveal tools reached through formal learning-platform integrations. They do not capture all the consumer AI services that students or lecturers open separately. Native API integrations are also excluded. This is a vendor analysis of platform activity, described through a press release and a companion blog; it is not a random survey of all university students or a peer-reviewed test of educational benefit.[1]
Why the distinction matters for students
Our interpretation is that a low position in an integration ranking cannot settle the argument about how much AI students use. Consider a student who reads a course page in Canvas, then asks a separate chatbot to explain the material. The first activity may leave a platform record while the second falls outside this study's lens. Conversely, opening an integrated tool does not tell us whether the student understood the answer, checked it or learned anything durable.
These gaps have practical consequences for university guidance. A policy based only on approved integrations could overlook the help that students already seek elsewhere. A policy based only on broad claims that everyone uses AI could pressure students into tools they do not need. Institutions should explain which services are supported, what information may be shared and how students can obtain equivalent help without paying for an additional account. Those are service-design questions, not conclusions measured by this ranking.
What a useful campus evaluation would ask
Mary Styers, Instructure's evidence and learning strategy director, describes security, accessibility and institutional approval as important considerations in adoption. That provider perspective is relevant, but it also reflects a business whose products sit inside institutional infrastructure. The portal's assessment is that approval documentation and actual classroom outcomes need separate scrutiny. A completed accessibility document is a starting point for testing the tasks a student must perform, not a substitute for observing whether those tasks are usable.
A university pilot could define a single learning objective and compare an AI-supported activity with its established alternative. Assessment should include an unaided follow-up task, so immediate assistance is not mistaken for retained understanding. Staff time spent preparing prompts, correcting errors and supporting students should be included in the cost. Students using assistive technology should be able to report obstacles. These are suggested evaluation criteria; the new ranking does not supply those results.[2]
How the assessment could change
An adoption study would become more informative if it connected several clearly separated sources: integration logs, voluntary reports of off-platform use and course-level evidence about learning. Combining them would require a proportionate privacy design, with no assumption that a student's browsing history is necessary to understand teaching quality. Published definitions should say whether a user is a student, instructor or another role, and whether the metric counts unique people, launches or institutions.
The US Canvas population is substantial, but size does not remove platform selection effects. Results should not be presented as a global university adoption rate or a finding about schools using different systems. Our assessment would change if independently evaluated courses showed durable learning gains, lower staff burden or better access under clearly specified conditions. Until then, the report is useful evidence about one route into educational software, and a reminder to ask what a dataset cannot see.