Perturbed Assignments and Subgoal-Labeled AI Literacy Instruction in Undergraduate Discrete Mathematics

Generative AI chatbots answer standard coursework accurately enough that a submitted solution no longer shows whether a student reasoned through a problem or delegated it. This ongoing classroom study in undergraduate discrete mathematics (N = 8) observes student-chatbot interaction twice and places instruction between the two observations. Two assignments were built with adversarial assignment perturbation, an instructor-led, iterative process that modifies standard questions until a target GenAI chatbot scores below a rubric threshold across three attempts. Between the assignments, one class session taught three subgoal-labeled procedures for diagnosing a hallucinated proof, strengthening a prompt, and collaborating with a chatbot to improve learning gains. Chat logs, pre- and post-AI literacy surveys, and two cued office hours sessions were collected; the logs are being coded with the DRIVE framework for further analysis. In office hours sessions to discuss their AI use with the professor, both consenting students described attempting problems before asking the chatbot to confirm their work and recalled the taught procedures in general terms. The study contributes a second classroom implementation of the perturbation method with a public tool, subgoal-labeled instruction for appraising chatbot-generated proofs, and insights into how undergraduate students use generative AI chatbots for learning to solve proofs.