Critique the Model, Then Talk to Each Other
Featured faculty: Felix Muzny
Associate Teaching Professor & Director of Teaching Assistants
Khoury College of Computer Sciences
TLDR: Felix Muzny turns AI output into a classroom debate: students critique what the model got wrong, then learn more from each other than from the model itself.
Felix begins every semester with a simple premise: don’t make a law nobody is going to follow. On day one, students in his natural language processing courses collectively write their own AI collaboration policy which includes when to use it, when not to, whether and how to cite it, and what should happen if the policy is broken. He calls it the jaywalking principle. Community buy-in, he argues, is the precondition for everything that follows.
What follows is a curriculum built around treating AI as a subject of rigorous scrutiny rather than a shortcut.
What he’s doing: Felix teaches natural language processing to both undergraduate and graduate students. These courses are about the technology his students are also using. His signature approach: when students are assigned a reflective prompt, he runs the same prompt through an LLM and brings both outputs into the classroom. Students are then asked to critique the model’s response by locating where it is wrong, where it flatters rather than reasons, where it performs agreement rather than analysis. He facilitates a parallel activity applied to research paper reading. Instead of discouraging AI-generated summaries, Felix generates one himself, then asks students to find what it missed such as misread figures, omitted nuances, and misleading framings. In both cases, Felix is deliberate about doing this work in class, in conversation. Getting students talking to each other, he notes, is itself a learning goal.
What’s working: Because Felix’s courses are explicitly about how LLMs function, critique is the point. Students engage with the AI’s output not to consume it but to interrogate it, and the in-class format generates the kind of lateral discussion that surfaces observations individual students are unlikely to achieve on their own.
| Adapting Across Contexts: The core move here is simple: run your own assignment prompt through an LLM, bring the output to class, and ask students to find what it missed. The activity works in any discipline where students are developing evaluative judgment — history, nursing, law, engineering — as long as the critique task is anchored in the standards of that field, not just a generic hunt for errors. What counts as a meaningful failure looks different everywhere: a misread figure in a research summary, an omitted precedent in a legal brief, an inappropriate framing in a clinical note. The in-class format matters too; students notice different things, and the conversation that follows is where the learning actually happens. |
What’s not working: The critique task has become harder as the models have improved. Early on, errors were easily identifiable and students could quickly disagree. Now the failures are subtler: a plausible but incomplete framing, a technically accurate summary that misses what mattered. Model critique also works better in person, but it draws on class time even though the amount of math Felix needs to teach hasn’t decreased.. At the same time, a portion of students has grown more trusting of AI output precisely as it has grown more fluent. Felix describes this as a double difficulty: the job is genuinely harder, and the inclination to do it carefully has, for some students, decreased.
What’s next: One response to that double difficulty is to make the stakes of critique more explicit. Last fall, Felix piloted and co-taught CS 1720 with Meica Magnani, a philosophy and CS colleague. In this course, students explore how being a fully literate user of generative AI requires both understanding how systems work technically-speaking and how to contextualize them within the framework of existing in society and being human. Aimed at non-majors and split between technical and philosophical content, it pushes the critique further: one activity asks student groups to teach the model something false and get it to assert the claim as fact. Groups diverge in strategy: some try to introduce wholly invented information; others coach the model to declare university rankings as objective truth. The variation reveals a shared vulnerability. Having demonstrated its value as a special topics course, CS 1720 has been assigned a permanent course number and will run as a standing course beginning fall 2026.
Have you tried something new with AI? How did it go? Send an email to [email protected] to let us know!