Students Using ChatGPT Ace the Assignment But Forget How to Code

A team of researchers ran a controlled experiment with 55 undergraduate computer science students doing introductory C programming tasks. One group could use ChatGPT-4.5; the other had to rely on ordinary web searches with no AI. The results split sharply in two directions. On the work itself, the ChatGPT group clearly won: they scored 89% on coding tasks compared to 69% for the search group.

The catch showed up later. When researchers tested what students actually remembered, the AI-assisted group recalled only 41% of the material right after finishing, versus 53% for the others. Two days later, that gap remained — 39% versus 52%. In other words, the AI group finished faster and scored better without really absorbing the material. They also reported less mental strain as tasks went on, suggesting the AI was doing more of the heavy thinking.

Perhaps the most striking finding was about ownership. Students using ChatGPT said only 45% of the submitted code felt like their own work, compared to 81% for the web-search group. That matters beyond grades: if you don't feel like you wrote something, you're less likely to defend it, improve it, or remember it. The researchers also tried measuring stress and focus through pupil response and heart rate, but data problems made those results unreliable.

The takeaway isn't that AI tools are bad for learning. It's that high scores can hide shallow understanding. The authors argue schools should design assignments where students explain their code, recall key ideas from memory, and contribute something genuinely their own. For anyone using AI at work or school, the lesson is simple: if AI hands you the answer, quiz yourself afterward. Otherwise you may look competent today and feel lost next week.