Run oral final exams with a voice agent and a panel of AI graders

NYU Stern professor Panos Ipeirotis, with co-teacher Konstantinos Rizakos, saw polished take home work from students who could not explain it when cold called. So they ran the final as an oral exam with an ElevenLabs voice agent that asked about each student's own project. Three models from Claude, Gemini and OpenAI graded every transcript, compared notes and revised. Thirty six students sat exams over nine days, averaging 25 minutes, for $15 in usage plus a $99 monthly subscription. Most final grades from the models landed within a point of each other. Problems included stacked questions and a voice students found harsh. Only 13% preferred the format and 83% found it more stressful, yet 70% said it tested real understanding.