Professor Alexandros Labrinidis is using AI in teaching, research, and more. He used it the day a classroom platform crashed five minutes before a lesson. And he uses it to spark class discussions, particularly when AI's mistakes create a live teaching moment that deepens the conversation. And he has also seen a rise in students who ace every assignment but can't pass an exam. Labrinidis, who teaches Intro to Data Science at Pitt, has settled into a consistent philosophy: AI as active collaborator, with himself firmly in the driver's seat.
Getting the Whole Class to Actually Try
During SQL lessons, Labrinidis used to pose a problem, tell students to work with their neighbors, then reconvene. In theory, great. In practice: "Maybe 10% of the class really thought about it. 80% didn't."
His fix was a custom ChatGPT-based chatbot that collected student responses in real time, compared them to a preloaded correct answer, and organized submissions by proximity to that answer. Labrinidis debriefed from the summary in real time, and participation jumped. Better yet, the AI occasionally mislabeled a correct answer as wrong. Rather than a glitch, it became a teaching moment: AI has real limits, and recognizing that requires genuine expertise.
He used a similar move when his SQL sandbox platform went down mid-semester. A quick AI prompt — here's the database I was using, can you recreate it? — had him back up and running before students noticed anything was wrong.
A Feedback Loop with a Catch
Labrinidis also used Claude to analyze end-of-class exit tickets across three consecutive lectures. Students reported what clicked, what confused them, and the main topics. The first summary was impressive: clear patterns, concrete suggestions. "I was blown away."
But he has noticed that the same pain points keep resurfacing, even after he feels that he has addressed them directly. "They say, 'we don't understand X, Y, Z,' I do three examples about those topics, and the next exit ticket says 'we still don't understand X,Y, Z,.'" His conclusion: without students studying independently between sessions, the feedback loop captured confusion that hadn't had a chance to be resolved. This is likely not a failure of the tool but a limitation of the conditions around it, which he might also be able to tackle with Claude as a thought partner during the lectures.
The 85% Rule
Whether drafting grant language, brainstorming research angles, or generating exam questions, Labrinidis applies the same standard: "It gets it right about 85% of the time, but unless you read the whole thing, you won't know what that 15% is." He always reviews the final output himself and keeps full agency over what gets used.
What's Still Unresolved
When assignment scores diverged sharply from exam performance, Labrinidis responded by raising the exam weight to 64% of the course grade. It addressed the problem of students copying AI outputs without learning, but it didn't solve the harder question of how to teach students to use AI as a thinking tool rather than bypassing thinking entirely. That's his next frontier.
What Faculty Need Most
Labrinidis sees HAIL's AI Teaching Field Notes as exactly the right kind of resource: a shared, honest record of what works and what doesn't. "Semi-positive results, fully positive, somewhat negative — we tried this, it didn't work, don't try it again." Alongside that, he flags practical friction: access to compute credits and institutional AI accounts matters more than it might seem. Reducing those barriers is where an institution can make a real impact.
Alexandros Labrinidis is a Professor in the Department of Computer Science at the University of Pittsburgh. Kendra Oliver conducted this conversation as part of HAIL's AI Teaching Field Notes series. This write-up was produced with the assistance of Claude.