What happens when a coding exam becomes a spoken defense with an AI partner?

Prof. Ts. Dr. Nurfadhlina Mohd Sharef teaches data mining at Universiti Putra Malaysia. She also serves as the university's Chief Digital Officer. In October 2026 she published a description of a new assessment method on the official Google blog. The post explains how she uses Gemini inside Google Colab to turn a traditional programming exam into a conversation about reasoning.

The classroom routine before the change

Students used to sit for timed coding tests. They wrote scripts from memory, debugged syntax errors, and submitted files for automatic grading. The score reflected whether the code ran and produced the expected output. That approach rewards memorization of library calls and familiarity with boilerplate. It does not reveal whether a student understands why one modelling choice fits the data better than another.

The old way measured speed and recall.

How the AI tutor fits into the workflow

Gemini runs inside the Colab notebook each student opens. The model suggests code snippets, offers alternative algorithms, and explains trade‑offs when asked. Students decide which suggestion to keep, which to modify, and which to discard. They must question the output, not copy it. The professor describes this as a form of human–AI teaming where the machine handles syntax and the human handles judgment.

The interview replaces the script

Instead of collecting notebooks, the professor conducts verbal interviews. Each student defends the modelling decisions made during the session. They troubleshoot live, point out where Gemini's recommendation was wrong, and explain why they chose a different path. The conversation lasts long enough to expose the student's reasoning process. A correct answer arrived at by luck scores lower than a flawed attempt that the student can diagnose and correct.

Scaling the evaluation

Behind the scenes the professor uses Gemini Canvas and Gemini Notebook. These tools help her review the interview transcripts and the notebook histories at scale. They surface patterns in how students interact with the AI, flag common misconceptions, and reduce the manual load of listening to every recording. The document does not specify how many students participate or how much time the tools save.

Where the redesign stops

The professor states a core belief: learning is not being replaced by AI, it is being redesigned to build lasting human capability rather than tech dependency. The post does not report student outcomes, pass rates, or comparative data against the old exam format. It also does not describe how the method would transfer to larger lecture sections or to courses without a coding component. The mechanism works for her data mining module; whether it generalizes remains an open question.