‘When Life Gives You LLMs’ – a teaching recipe for limiting the impacts of AI

The way I teach my classes has changed more in the past three years than in my entire prior career. This change was forced upon me, but it has ultimately been for the better. As the saying goes, “When life gives you lemons…”

What life gave us beginning in late 2022, of course, was not lemons but LLMs. Here I offer my lemonade recipe. I’m still refining it. Perhaps I can help you develop your own. Perhaps you’ll tell me how to improve mine.

Let me start with the reasoning that led me to my current pedagogical approach.

During 2023-2024 I arrived at three basic premises. First, I could no longer trust that essays written outside class represented students’ own abilities. Second, I did not want to give up assigning such essays. Third, I did not want to police students’ essays for AI content.

Here’s the conclusion I reached: I now have two streams of assignments in my classes. One stream comprises work done in class, where I prohibit AI use. The other comprises work done outside class, where I allow use of AI. But here’s the catch: the two sets of assignments aren’t averaged for a final grade. Instead, the class grade is determined by whichever stream grade is lower. For example, a C in the in-class (IC) stream and a B+ in the out-of-class (OC) stream results in a class grade of C.

You might think students would resent having their class grade set by the lower of the two stream grades rather than by their average. However, I’ve found that most students accept the structure described above once they understand it. Let me explain further.

In case it’s not obvious, the reason for the dual structure is to make it hard for students to achieve a high grade by using AI in their OC assignments. They can indeed do very well in the OC stream by using AI, as such use is not directly penalized. But their class grade is still limited by their performance in the IC stream. If that IC performance doesn’t match their AI-assisted OC work, then that AI use will all be for naught.

In order to implement this structure, I had to dramatically ramp up the amount of in-class work I assigned. Lecturing dominated my teaching prior to 2023; in-class assignments were occasional. That proportion has now flipped. I lecture only occasionally. Most time in class is spent on hand-written assignments, with devices banned (barring accommodations, of course; but no student has felt the need to make such a request so far). The assignments frequently involve small-group collaboration.

All in all, this has been a positive change. I knew of the evidence that lecturing is not very effective. The arrival of LLMs gave me a reason to act on it.

Most of the effort for me in making this change has been in developing the IC assignments, which occur in almost every class. They often require students to read a text closely and answer questions about it. Often this takes the entire period. Sometimes I interweave an assignment with lecturing. For example, I might walk through an argument from an essay they’ve read and have them evaluate it as I progress.

I believe that close reading is the most important skill that is threatened by AI. Though writing is a close second, a student’s thinking can’t even get off the ground unless they wrestle with a text themselves. If I assign tasks like these in class, I can (for now) know that the reading and writing are done by the student.

I do not think that all AI use is harmful. But it is harmful for students who are still developing these skills of close reading and focused writing. If a college education does nothing else, it should inculcate the ability to read and think for oneself, rather than having to rely on pablum regurgitated by a machine. A person who never learns to read complex texts for themselves, but relies on AI summaries, is a person who is easily manipulated.

Despite this, I also believe it is pointless to try to stop students using AI. There are benefits to avoiding watching for, investigating, and prosecuting AI use. I find these forensic tasks personally soul-crushing and pointlessly adversarial. Thankfully, I don’t have to worry about a student’s AI use unfairly inflating their grade: my dual structure limits the extent to which heedless AI use—handing off an entire task, or most of it, to AI—can boost a student’s final grade. Here are two more ways I harness this structure to prevent such heedless use.

One strategy is to link OC assignments to IC assignments. If a student hands off the former to AI, their performance on the latter will suffer. OC assignments are preparation for IC assignments. Importantly, I’m using specifications grading: assignments receive either an ‘S’ for ‘satisfactory’ or an ‘N’ for ‘not yet’ (they’re revisable), where an S is full credit and an N is none. There’s no ‘low road’ to a passing grade via collecting minimal points. If a student performs poorly on an assignment, they simply earn no credit. And since credit on OC assignments is worthless without correlative credit on IC assignments, students are motivated to do the OC assignments themselves in order to be prepared for the IC assignments.

A second strategy to discourage heedless AI use that I deploy in Gen Ed classes is to make the final essay optional (it functions as a boost for students who want an A) but require students to meet with me to discuss their draft if they submit one. I tell them, truthfully, that the goal of this meeting isn’t to unmask AI use. Instead, it’s a chance for them to talk through their ideas. If they come in with nothing or very little to say—the likely result of heedless AI use—then I’ll deny them permission to submit a final essay. Some students still use AI in their draft. But instead of me grilling them for a confession, we can openly discuss how they used AI, how this may have weakened the draft, and how they might do better—whether that is on their own or by using the AI in a smarter way.

Typically, a third to a half of my students in these Gen Ed classes end up foregoing the essay assignment. Does it bother me that many students end up not writing an essay? Yes. But those students are still getting practice with the elements of essay writing. As I’ve said, the IC assignments focus on skills such as close reading and argument analysis. I also require a prospectus for the essay (before they decide whether to write one). So even if a student never writes an essay in my class, I’m still training them in some of the relevant skills. This is a bargain I’m willing to strike in exchange for not having to act as a police officer patrolling my students’ submissions for AI skullduggery.

That, then, is my lemonade. It’s all hostage to evolving technology, of course. My big fear is that AI-enabled glasses will become common, which could make it much harder to keep AI out of my classroom. But we can only take these developments as they come. For now, the strategies I’ve outlined are working for me. What’s your recipe?