by Orkun İrsoy, PhD Candidate in Electrical & Computer Engineering at CMU
Over the past few years, I’ve served as a teaching assistant for more than fifteen different courses, and one challenge has been prevalent: ensuring fairness and consistency in how students are evaluated. It is not rare that professors enjoy keeping the same TA across semesters to avoid having long discussions to transfer the grading standards and educational outcomes that are established over the years. This problem becomes particularly visible in courses with projects or other qualitative assessments, where there is no single correct answer and judgments can vary widely. Each TA leaves behind valuable know-how about what to look for and how to evaluate work, but that knowledge is rarely documented in a structured way. As a result, every new TA essentially starts from scratch.
Last year, when I was a first-time TA for Networks in the Real World, I encountered this problem head-on. The course required evaluating both project presentations and written reports, I was working with three other graders who mentored multiple project groups, and while I had access to evaluations and feedback from previous years, they were scattered, anecdotal, and unstructured. I decided to use ChatGPT as a way of organizing this inherited experience. By feeding in past comments and asking the model to identify patterns, I was able to generate a draft rubric that captured recurring themes such as clarity, rigor, creativity, and quality of delivery. After refining the rubric with the professor, we shared it with all evaluators. As a result, grading became more consistent, students knew exactly what was expected, and I felt more confident that evaluations were fairer across the board.
A similar challenge came up again this year, this time around project selection. Traditionally, students were expected to propose their own project topics at the beginning of the semester while they were provided a list of examples. The problem was that many had little sense of what different kinds of projects would require. Some chose ideas that were far too ambitious, others ended up with topics that were too narrow, and the resulting variation made it difficult both for students to succeed and for TAs to evaluate their work. With the professor, we wanted to provide more structure by offering predefined categories, or “archetypes,” that outlined typical project paths which would still allow independence over the topic itself. A project on spreading processes, for instance, usually involves very different steps than one focused on graph learning, and these steps were actually informally defined by the past TA’s who conducted several project meetings over the semester with the individual project groups providing personalized feedback. Once again, I turned to ChatGPT. I collected final projects of past projects spanning a few years and asked the model to distill them into categories and typical workflows given the resulting reports. The draft it produced wasn’t perfect, but it gave me a starting point to refine with the professor’s input. The end result was a set of clear archetypes that students could choose from, providing both structure and flexibility.
Looking back, what intrigues me is not that AI replaced my work as a TA, but that it helped preserve and translate the tacit know-how of those who came before me. By turning scattered insights into rubrics and project archetypes, it made expectations clearer for students, reduced grading drift for professors, and freed me to focus more on mentoring. More generally, this points to an important role AI can play in education: translating the informal, often invisible knowledge accumulated by instructors into explicit, reusable structures. In this way, AI becomes less a replacement for human judgment and more a bridge between individual experience and collective learning.