Study finds ChatGPT and critical-thinking training improve different student skills

OpenAI reports that a randomized experiment involving more than 1,000 Bocconi University undergraduates found ChatGPT and critical-thinking instruction improved student work in different, complementary ways. The findings, published by OpenAI News on August 27, 2026, offer a useful signal for educators and builders designing AI-assisted learning tools.

AI raised quality; reasoning widened ideas

Students worked on a marketing case for the university’s merchandise store and were assigned to one of four groups: GPT-4o access, causal-reasoning training, both, or neither. Human graders scored submissions on a five-point rubric, while researchers separately analyzed idea variety, causal explanations, and similarity to work from three experts.

Students with ChatGPT access scored nearly one point higher on the five-point scale. Their submissions contained more ideas, clearer logic, and recommendations that more closely resembled expert work. The experiment did not simply measure students outsourcing the assignment: participants still had to prompt the model, assess its responses, and decide what to use.

The reasoning exercise produced a less visible but important benefit. It helped students explain why recommendations might succeed or fail and generated ideas that were more varied and distinct from their peers’ work. However, those students did not score higher on the conventional rubric, which focused on two standard marketing goals.

Implications for AI tools and education

The combined group showed the broadest gains, pairing ChatGPT’s improvements in polish and coherence with stronger reasoning and originality. For AI product teams, the result argues against treating answer quality as the only measure of learning. Tools could also support idea exploration, assumption-checking, causal explanations, and multiple possible approaches.

For schools, the study suggests polished final answers reveal less about understanding when AI is available. Assignments and evaluation systems may need to reward originality and reasoning—not just clarity and expert-like presentation.

Source: OpenAI News