AI helps grades, not learning: study

This study adds to a growing body of evidence that AI can improve task performance without improving underlying skills. A prior analysis of over 500,000 US college grades found that top grades increased most in writing and programming courses with heavy homework after ChatGPT's launch. Controlled experiments showed that when AI was removed, participants performed worse than a control group, especially those who used GPT for direct answers. A 30-month study of 26,000 students in China found homework improved but exam scores dropped, with entrance exam results 18 to 24 percent lower over the long term.

The Bocconi experiment highlights a tension in education: current grading rewards conventional, polished answers, not originality or deep reasoning. The authors argue for changing grading criteria to explicitly reward originality and reasoning. However, as the article notes, this would likely require systemic changes. The study also has limitations: it involved only freshmen at one university, randomization was at the class-section level, and some measures were evaluated by AI models, including from OpenAI, which was involved in the research. Whether the GPT advantage reflects actual learning or just better output remains an open question.