Bocconi-OpenAI study finds ChatGPT lifts scores but not original thinking
A Bocconi University and OpenAI study of 1,053 students found ChatGPT access raised task scores 0.86 points, while reasoning training grew idea diversity.
Photo by Yan Krukau on Pexels
A randomized controlled trial by Bocconi University and OpenAI Economic Research, published 27 August 2026, found that giving students access to ChatGPT (GPT-4o) raised their task scores but did not make their ideas more original. A separate training arm that taught students causal reasoning had the opposite pattern: it did not raise scores, but it broadened the diversity of ideas students produced. The findings, drawn from a ChatGPT critical thinking study of 1,053 first-year Bocconi students, point to a gap between what generative AI improves in student work and what structured reasoning training improves.
What the study tested
Researchers, led by Bocconi economist Alfonso Gambardella with co-authors affiliated with Duke University, UC Berkeley, SDA Bocconi and the ION Management Science Lab, randomly assigned 1,053 first-year Bocconi economics, finance and management students by class period to one of four groups: access to GPT-4o only, a causal-reasoning training module only, both combined, or neither. Each group completed the same task: writing marketing recommendations to increase awareness and use of Bocconi’s campus merchandise store. Human raters scored submissions on a 1-to-5 rubric, and the researchers ran additional automated text analysis on the recommendations.
Scores up, originality unchanged
Students with ChatGPT access scored 0.86 points higher on average than the control group’s mean of 2.09, a jump to roughly 2.95 on the 5-point rubric. The study found ChatGPT-assisted answers moved closer to what expert marketers would recommend, with clearer logic and more developed ideas. It did not find an increase in the originality or diversity of those ideas: ChatGPT users tended to converge on similar, expert-like recommendations rather than producing a wider range of distinct approaches.
Reasoning training moved the opposite lever
The causal-reasoning training arm, which taught students to identify mechanisms and test them against alternative explanations, did not raise rubric scores. It did increase the diversity and distinctiveness of the ideas students proposed relative to their peers, and improved their use of mechanism-based and falsification reasoning. Students who received both the training and ChatGPT access showed elements of both effects, according to the study.
| Group | Task score (1-5 rubric) | Effect on idea diversity |
|---|---|---|
| Control (no AI, no training) | 2.09 (baseline) | Baseline |
| ChatGPT (GPT-4o) access only | 2.95 (+0.86 vs. control) | No increase in originality |
| Causal-reasoning training only | No statistically significant change reported | Increased diversity and distinctiveness |
| Training combined with ChatGPT access | Score increase reported | Diversity increase reported |
Source: OpenAI Economic Research and Bocconi University, “Training novices to think, or giving them LLMs? Evidence from an RCT,” 27 August 2026.
Why the gap matters for assessment
The researchers frame the result as an assessment-design problem rather than a case against either intervention. Because ChatGPT access raised scores on the study’s rubric without increasing originality, and reasoning training raised originality without moving the rubric score, standard grading criteria may reward the wrong skill depending on what a course is meant to teach. The finding lands alongside a separate white paper from the Computing Research Association’s Education Committee, published the same week, that urges computing programs to move beyond product-only grading toward methods such as oral exams and process-based assessment as generative AI becomes able to produce polished work on its own.
Background
OpenAI’s Economic Research team has published a series of studies since 2025 examining how large language models affect classroom and workplace tasks, typically in partnership with universities running randomized trials rather than observational surveys. Bocconi University, based in Milan, has run several AI-in-education pilots through its SDA Bocconi School of Management and has published prior research on how students use generative AI tools in coursework. This study is among the first to isolate the separate effects of AI access and reasoning instruction within the same experimental design, and its authors say the same method could be applied to other task types where colleges are still deciding how to grade AI-assisted work.
Sources: OpenAI Economic Research and Bocconi University (working paper); OpenAI; Bocconi University; EdTech Innovation Hub
Featured image: photo by Yan Krukau on Pexels (free Pexels license).
Become a Sponsor
Our website is the heart of the mission of WINSS – it’s where we share updates, publish research, highlight community impact, and connect with supporters around the world. To keep this essential platform running, updated, and accessible, we rely on the generosity of you, who believe in our work.
We offer the option to sponsor monthly, or just once choosing the amount of your choice. If you run a company, please contact us via info@winssolutions.org.
I specialize in sustainability education, curriculum co-creation, and early-stage project strategy. At WINSS, I craft articles on sustainability, transformative AI, and related topics. When I’m not writing, you’ll find me chasing the perfect sushi roll, exploring cities around the globe, or unwinding with my dog Puffy — the world’s most loyal sidekick.
