ChatGPT and Critical-Thinking Training Improved Student Work in Different Ways, OpenAI-Bocconi Experiment Finds

A randomized experiment involving more than 1,000 first-year undergraduate students at Bocconi University offers a useful signal for schools trying to separate AI-assisted performance from independent reasoning skills. The study, conducted by Bocconi researchers in collaboration with OpenAI Economic Research, assigned class periods to four conditions: ChatGPT access, causal-reasoning training, both interventions, or neither.

OpenAI reports that students with ChatGPT access scored nearly a point higher on a five-point grading rubric and produced work that was clearer, more idea-rich and more similar to expert recommendations. The causal-reasoning exercise produced a different effect: it did not raise the conventional rubric score, but students generated a wider range of ideas and more explicitly examined why proposals might work or fail. Students receiving both interventions showed benefits across the broadest set of measures.

The practical implication is not that AI replaces critical thinking. It is almost the opposite: as generative AI makes polished output easier, conventional grading can become a weaker proxy for what a learner actually understands. Assessment design may need to reward originality, reasoning, assumption-checking and comparison of alternatives alongside presentation quality.

The result should also be interpreted at the right scope. This was a specific business-case assignment using GPT-4o with first-year Bocconi students; it does not prove identical effects across every subject, age group or model. Still, the randomized design makes the experiment more informative than simple surveys of AI use. For education leaders, the strongest takeaway is that AI access and explicit reasoning instruction appear complementary rather than mutually exclusive, and future curricula may need to teach both.