Kelly Wen
Co-Founder, EdPilot
The obvious risks get discussed. The subtler ones — the ones that compound quietly over a semester — are harder to see and do more lasting damage to actual learning.
The conversation about AI risks in higher education tends to focus on a few scenarios: a student submitting AI-generated work as their own, a model hallucinating a source that the student cites, an assignment that could be completed by anyone with a ChatGPT subscription.
These are real risks and worth taking seriously. But they're the visible ones — the scenarios that produce an obvious problem the institution can point to. There's a category of risk that's subtler, less dramatic, and in some ways more corrosive to actual learning.
When a student uses a general-purpose AI to study, the AI draws from whatever it was trained on — not from their course materials. The information is often accurate in a general sense. It may still be wrong for this course.
A professor's framing of a concept isn't arbitrary. They chose particular examples because they connect to what comes next. They defined terms a specific way because the rest of the course builds on that definition. They emphasized certain aspects of a theory and deemphasized others because of where they're taking students.
A student who spends a semester studying with a general-purpose AI is getting a parallel education — one that sounds like what their professor is teaching but is subtly misaligned. They develop mental models that make sense internally but don't connect correctly to the course's framework. The confusion this produces is genuinely hard to diagnose, because the student doesn't feel confused. They feel like they understand things.
That's the miscalibration problem. It shows up in exams as wrong answers delivered confidently.
There's a second category of risk that gets less attention: what students share when they use public AI tools.
Students routinely paste assignment prompts, exam review materials, and course documents into public AI systems. In doing so, they're transmitting institutional content to platforms the institution has no agreement with, no visibility into, and no control over.
This matters for a few reasons. Course materials are intellectual property — lecture slides and assignment designs belong to the faculty member or institution. In some disciplines, students work with sensitive materials — case studies, datasets, documents — that may have privacy implications. And the training practices of public AI providers aren't always transparent; there are legitimate questions about how input data is used.
Most students don't think about any of this. They're using a free tool to study for an exam. The institution has no way to see it happening.
The most long-term risk is harder to quantify but may matter most.
AI tools are very good at producing the artifact of understanding — text that looks like someone has comprehended and synthesized something — without the underlying comprehension. A student who uses AI consistently as a study aid may gradually lose the habit of building their own understanding, because the friction that drives real learning keeps getting removed.
This isn't hypothetical. Learning requires productive struggle — working through confusion, making errors, revising thinking. AI that smooths all of that away efficiently doesn't build the cognitive muscle that education is supposed to develop.
The question isn't whether students should use AI. It's whether the AI they use is designed to build understanding or to produce the appearance of it.
These risks don't argue for prohibition. They argue for structure.
An institution that provides students with curriculum-grounded AI — AI that faculty have configured, that draws from course materials, that prioritizes understanding over output generation — addresses most of these risks directly. Miscalibration goes away when the AI is grounded in the actual course. Privacy concerns are manageable when the platform is institutional rather than public. The dependency problem is at least partially mitigated when the AI is designed to guide learning rather than replace it.
The risks of public AI use in higher education are real. They're also mostly avoidable, given the right infrastructure.