Agentic AI Comes for Teaching

Will agentic AI let teachers spend more time teaching, or will it remove even more of the human touch from the academy? This is one of higher education’s most important questions following the release of the now-defunct Einstein and Canvas’s own agentic AI teaching agent.

Canvas, an online-learning-management system operated by Instructure Holdings, is used by 50 percent of all students in North America. This market share means that any changes the firm makes reverberate throughout higher education. Canvas’s new AI agent, IgniteAI, lets professors delegate tasks such as rubric generation, grading, and the creation of course-specific Canvas sites. The hope is that this development will free up the time currently wasted by professors on the minutiae of deadlines, web design, and administrative tasks so that teachers can spend their days teaching.

If AI automates rubrics and basic grading, then teachers will have less of a role. However, that hope is yet to be realized. Ed-tech’s history has been the story of countless hoped-for revolutions, from distance learning for the masses to so-called Massive Open Online Courses (MOOCs). Each development sought to change education as we know it, yet, while they often helped make higher education more accessible and improved lives, they were not truly game-changing. The reason these revolutions failed to change everything can help us understand the promise and perils of agentic AI in higher education.

Agentic AI has immense potential to either improve the business of teaching or continue its sad transition. Many educational technologies have been confronted with a deadly mixture of hype and reality. Their creators and evangelists proposed that the new technology would make education perfect and easy, ignoring the fact that education necessarily requires people to struggle and work hard over the course of years. Teaching and learning aren’t always easy, nor are they always fun. Both are best understood as meaningful yet difficult. That distinction is important, suggesting that we must be careful about what we let AI replace.

Take rubric design, for example. Designing formal guidelines for every assignment is a long and tedious process, but rubrics done right aren’t an arbitrary way of assigning points based on the satisfaction of silly requirements. Rather, they, and by extension grades, measure how well students meet key learning benchmarks. Similarly, grading is not mindless box-checking; it is how teachers measure students’ performance and oversee their academic development using their own professorial expertise. Both of these things are difficult and tedious, yet they are core to the instruction process.

If AI automates rubrics and basic grading, then teachers will have both less of a role in shaping the benchmarks students follow and a weaker grasp of whether students ultimately meet those standards. Even if an AI bot is trained on materials that give it a sense of these goals, that is no replacement for the human touch. The fact that there is likely a tacit, uncapturable element to grading and rubric design may actually make such training impossible, as the information we need may not be quantifiable in bits and bytes.

As a result, this proposed automation may create an overly optimized educational system that meets the metrics set by administrators and accreditors while effecting little actual teaching. On the surface, everything is graded, all assignments are submitted, and every milestone is met. Yet, delve a little deeper, and it is clear that very little education has taken place. This isn’t pure speculation. Over the last two decades, the endless pursuit of higher test scores initially drove up exam results but left college students unable to read books and ultimately failed to keep scores high.

This isn’t to say that agentic AI is an educational dead end. There are real, beneficial use cases for this technology. For example, professors today waste much of their time complying with various administration-imposed requirements that have little bearing on students’ actual educations. While reducing these requirements would be ideal, public-choice theory shows how unlikely that is. Having AI address them is a good second-best option, one that really might allow teachers to spend more time on the actual business of teaching.

Agentic AI has immense potential to either improve the business of teaching or continue its sad transition from an instrument of growth to an instrument of hollow credentials. What ultimately happens will come down to how educators choose to use AI and how companies develop it. Canvas is already releasing the agent. When educators start using it, I hope they think about how to use it for good.

Zev van Zanten is an economics and mathematics student at Duke University.