NYU professor fights AI cheating with AI-powered oral exams that cost 42 cents per student

AI General
1

NYU Professor Innovates Against AI Cheating Using AI-Powered Oral Exams at 42 Cents per Student

In an era where artificial intelligence tools like ChatGPT enable widespread academic dishonesty, educators are seeking robust countermeasures. A professor at New York University (NYU) has pioneered an ingenious solution: AI-powered oral examinations that verify student knowledge through live verbal interaction. This approach, costing just 42 cents per student, leverages readily available AI technologies to administer personalized, proctored oral exams remotely, effectively neutralizing common cheating methods.

The Challenge of AI-Driven Cheating

The proliferation of large language models (LLMs) has transformed academic assessment. Students can now generate sophisticated written responses instantaneously, rendering traditional exams vulnerable. Detection tools exist, but they are imperfect, often flagging legitimate work or missing disguised AI output. Written proctored exams introduce logistical hurdles, such as scheduling and supervision costs.

Enter oral exams, a time-honored method that demands real-time articulation of concepts. Historically labor-intensive for instructors, they scale poorly for large classes. NYU Professor Joseph O’Rourke, a computer science educator, recognized this gap and engineered an AI-assisted system to make oral exams feasible, scalable, and economical.

How the AI-Powered Oral Exam System Works

O’Rourke’s implementation centers on a streamlined workflow using off-the-shelf AI services. Students join a scheduled Zoom session individually. Upon connection, an AI avatar—powered by HeyGen—greets the student and initiates the exam. The avatar poses questions drawn from a curated question bank tailored to the course syllabus.

Key components include:

  • Question Generation and Personalization: Questions are pre-prepared by the professor but dynamically selected and voiced by the AI. This ensures coverage of core topics like algorithms, data structures, or theoretical computer science, depending on the course. The AI maintains a conversational flow, following up on responses to probe deeper understanding.

  • Speech Synthesis and Interaction: HeyGen’s AI avatar delivers questions with natural intonation, lip-syncing for realism. Students respond verbally, and their answers are transcribed in real-time using OpenAI’s Whisper API.

  • Proctoring and Evaluation: Zoom’s recording captures video and audio for review. The professor later assesses transcripts and recordings, grading based on clarity, accuracy, and depth. No automated grading occurs, preserving human judgment.

  • Technical Integration: A simple script orchestrates the process. It triggers the AI avatar at the session start, streams questions, and logs transcripts. The entire setup requires minimal coding expertise.

The system’s cost breakdown is impressively lean: HeyGen charges approximately 20 cents per minute for avatar sessions, with exams averaging two minutes. Whisper transcription adds about 10 cents, and Zoom hosting is negligible at scale. Miscellaneous API calls push the total to 42 cents per student—far below human proctoring rates.

Implementation Details and Scalability

O’Rourke piloted this in his “Computational Geometry” course, enrolling 30 students. Scheduling was handled via Calendly, staggering sessions over several days to avoid bottlenecks. Each exam lasted 5-10 minutes, including setup.

Technical prerequisites are straightforward:

  1. API Keys: Obtain keys for HeyGen, OpenAI (Whisper), and Zoom.

  2. Question Bank: A Google Sheet or JSON file with 50-100 questions, categorized by difficulty and topic.

  3. Automation Script: Python code using libraries like pyautogui for Zoom control, openai for transcription, and HeyGen’s SDK for avatar deployment.

  4. Backup Measures: Fallback to professor-led exams if AI glitches occur.

Scalability shines here. For a class of 100 students, total cost is $42, versus thousands for in-person proctoring. O’Rourke notes the system handles variability in student accents and technical issues gracefully, thanks to Whisper’s multilingual capabilities.

Student and Instructor Feedback

Students report a positive experience. The AI avatar reduces anxiety compared to human examiners, feeling like a neutral conversational partner. One student remarked, “It forced me to explain concepts in my own words, which written tests don’t.” Pass rates aligned with expectations, with failures attributable to genuine knowledge gaps.

For instructors, time savings are substantial. O’Rourke spends post-exam hours reviewing rather than conducting live sessions. Grading focuses on qualitative insights, such as reasoning chains, undetectable by LLMs in real-time speech.

Limitations exist: The system assumes reliable internet and webcam access, potentially disadvantaging some students. It also requires upfront preparation of the question bank. O’Rourke mitigates equity issues with optional in-person alternatives.

Broader Implications for Education

This model exemplifies “AI fighting AI,” turning cheating tools against themselves. Oral exams inherently resist LLMs, as generating coherent speech responses demands integrated knowledge application. Unlike written plagiarism, verbal slips reveal imposters.

O’Rourke shares his setup openly on GitHub, inviting adaptations. Early adopters in other NYU courses report success, suggesting wider adoption. As AI cheating evolves, hybrid assessments like this—blending technology with pedagogy—offer a sustainable path forward.

Cost-effectiveness democratizes advanced assessment. Community colleges or underfunded institutions could implement it without budgets strained by proprietary proctoring software. Ethical considerations, such as data privacy, are addressed via Zoom’s compliance and local transcription storage.

In summary, O’Rourke’s innovation demonstrates how accessible AI can restore integrity to evaluations. At 42 cents per student, it redefines feasibility, empowering educators to prioritize learning over enforcement.

Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.

What are your thoughts on this? I’d love to hear about your own experiences in the comments below.