At a recent academic retreat I attended, the air was thick with what I can only call educational gaslighting. A panel of graduate and undergraduate students looked a room full of professors in the eye and claimed they only used AI to verify their work because they valued learning too much to take shortcuts. Minutes later, when the answers were blind, those same students estimated that over 80% of their peers were using the technology for nearly everything.
As an engineering professor at the University of Michigan, I believe we need to move past the fear and hype. The future job market will not be dominated by autonomous AI, but by experts who have mastered their field so thoroughly that they can use it to multiply their output exponentially.
But how do we help students become experts if they don’t show up?
This question precedes the LLM onslaught. Since the pandemic, traditional lecture attendance has cratered, but active learning has been shown to significantly improve both turnout and long-term retention.
By evolving my courses to embrace this data, I’ve seen attendance surge, even in the freeze of a Michigan winter.
Flipping the lecture cycle
Classically, engineering courses default to hours of lectures where students are expected to take notes, with problem sets and exams bolted on. Many students treat a lecture as passive entertainment. And, often, the material is so technical that it is disconnected from real-world use, leading to even less engagement and retention.
To break this cycle, I’ve flipped my classroom. Each week, I assign a 2-hour recorded video lecture, along with a related article. The assignments are made in Perusall, an AI-enabled tool that treats the video and article a bit like a social network. Students are graded based on their active engagement with the material, such as how much of the lecture they view, what questions and comments they leave in the system, and so on. I can monitor which students leave comments, answer peer questions, and engage with the material before they ever set foot in my classroom. And if they try to cut and paste comments in multiple locations, the system flags them. It does not yet flag comments that seem AI-generated, but I expect that will be coming soon.
With everyone primed to dig in, only one-third of my students’ time with me is devoted to classic lecturing. I offer a one-hour live lecture and invite industry guests to share stories of computer vision in the wild. Then my students spend the rest of our in-person time participating in small breakout sessions, a large group discussion, and an in-person quiz. Not only do they grade their own quizzes, but they only get credit for an answer if one of them argues the logic behind it.
The result? My students show up to class because the value is no longer in the information I provide—it’s in the friction and growth of live exchange.
This fall, I’m taking this a step further. We won’t just read technical papers; we will debate them. Anyone can be called to the front of the room to spontaneously argue one side of a research argument, which means every student must come prepared.
By moving the passive learning to the home and continuously pushing students to test their knowledge, I’ve reclaimed the classroom to create what AI cannot replicate: spontaneous, high-stakes human interaction.
Using AI as a supercharged tutor
As we try to understand how AI can help and hinder learning, the most dangerous misconception is that it is a labor-saving device for the mind. In reality, AI is an expertise-amplifier that can turn weeks of manual programming into a few hours of focused work. But for a novice, relying on AI before mastering the fundamentals creates a technical debt that leads to a lack of depth.
As someone at the forefront of AI research and creation, I don’t coach my students to avoid it, but rather I use it as a sophisticated, one-on-one tutor that facilitates active learning and helps them grow their expertise. This means moving beyond passive consumption and toward a rigorous, iterative process of trial, error, and refinement.
Some best practices I share with my students include:
Mastery-First Workflow: Solve problems manually first. Then use AI to check your work and identify where your logic diverges from the model.
AI as a Problem-Generator: One of the most effective ways to learn is through constant testing. Use AI to generate new practice problems and engage in active learning.
Brain Dump Standard: Never ask AI to write from scratch. Instead, provide a brain dump of ideas and structure. After the AI helps organize your expertise, personally refine it through meticulous review or even rewrite, if necessary.
Redefining the honor code in the age of AI
I am not an AI police officer. I cannot—and should not—spend my academic career hunting for digital shortcuts in my students’ work. I can only set the boundaries and allow them to choose how they show up.
Amid the promise of AI to supercharge the work of experts, we must treat this technology with the same proactive mastery we apply to any other essential tool of modern life.
By shifting the focus to high-stakes, spontaneous human interaction and leveraging AI for active learning rather than trusting it to do the work, educators can ensure that the knowledge lives within the student, not just the model.
You must be logged in to post a comment Login