Student AI Use Is a Demand-Side Problem

Einstein is an AI bot described this way: “Einstein is an AI with a computer. He logs into Canvas every day, watches lectures, reads essays, writes papers, participates in discussions, and submits your homework—automatically.”

Yes, the product claims to be able to complete any work in a Canvas course entirely on its own. There may be exceptions—like uploading a video of yourself—but clearing a hurdle like that is ultimately pretty solvable.

What does this mean? I think it means both less and more than some people might be thinking.

I am heading to a speaking gig, making me short on time to develop a full-fledged, deeply thoughtful piece of writing, but let me throw out a series of thoughts to perhaps contextualize what I think is going on here.

AI agents that can go into an LMS and complete assignments is not new. Marc Watkins and I talked about them in a Q&A last November, referencing work by Anna Mills from even before that.

We don’t wholly know what this is. Companion.AI, which appears to be the hosting entity for Einstein, is the apparent project of Advait Paliwal, a 2024 graduate of Michigan State.
Whatever it is, it does not appear to be a major, venture-backed project. Of course, given the state of automated coding tools and the relatively easy integration of LLMs into a program, it’s easily believable that a single person or small group could spin up something like this that looks and acts plausibly.
I haven’t tried it because I don’t have access to a safe Canvas course shell, but someone should. We should know how capable these agents are.

Yes, writing about it, or posting obsessively about it on BlueSky (as I and others in my circle have been doing) is doing some measure of marketing work for the product, but this stuff—particularly if it works—is ultimately going to disseminate anyway.
There is no winning the detection arms race against AI automation technology. Not only will detection fail, but investing significant energy in detection is a misdirection of time and resources.
Analog assessment is only a partial defense against these agents, and choosing to retreat exclusively to these methods risks narrowing the type of learning that students experience.

The deep roots of the existence of agentic AI like Einstein are the transactional model of education that reduces school to credentialing. Einstein promises to do the “busywork so you don’t have to.” That students see some significant part of what happens in school as busywork is the problem that needs tackling.
It is obvious to those of use with more seasoning that learning nothing while getting a credential is going to cause problems in the long run, and for sure many students get this, but many do not.

For this reason, it is my view that we should be looking at the current challenges as a “demand-side” problem. It is now trivial for students to outsource their schoolwork to an AI model or agent. The chief innovation of Einstein—if it works—is eliminating the need to cut and paste LLM outputs. The only way to get students to learn is to make learning a more meaningful and attractive proposition than outsourcing the earning of a credential to a bot.

As I’ve written previously, there are approaches that allow for steady progress in addressing the challenges of the existence of this technology in ways that are consistent with local and individual contexts. There’s nothing about agentic AI doing coursework that changes this.

We know what learning looks like—pretty much the opposite of Einstein—but as a system, we haven’t been as focused on learning as we could have been. Each of these occurrences should be viewed as an occasion to consider what learning really looks like and how what we ask students to do in class supports that learning.

And to repeat, it’s possible that a recent graduate with a software development background has developed something that works in the way the product claims, but as of this time, it has yet to be proven. Lots of generative AI applications that appear on the scene and grab lots of attention are more novelty than substantive. We need to know if this is substantive.

Even if this doesn’t work, something like it probably will someday, so let’s go ahead and prepare for that eventuality by decreasing the salience and relevance of this technology.