ChatGPT makes homework polished, but human thinking keeps it original

What happens when you unleash AI on a cohort of fresh college students?

A randomized study conducted by Bocconi University researchers alongside OpenAI Economic Research tracked over 1,000 first-year undergraduates tackling a real marketing challenge: pitching a business strategy for the university merchandise shop.

The students were divided into four cohorts: one got access to ChatGPT powered by GPT-4o, another received an AI-free causal reasoning training module, a third received both, and the control group got neither.

Submissions were graded by human evaluators on a standard five-point rubric, while automated text analysis tracked originality, logical structure, and resemblance to solutions written by industry experts.

Polish from the bot, spark from the brain

The split in results was stark.

Students using ChatGPT saw their scores jump by nearly a full point on the five-point scale. Their papers offered clearer logic, richer idea density, and closely mirrored the work of seasoned professionals. Essentially, the AI acted as an instant leveler, lifting raw novices to an expert-looking baseline.

Yet AI alone did not make their ideas any more novel. That breakthrough came entirely from the causal reasoning training.

This training—a game- and feedback-based exercise teaching students how to link causes with effects and probe where an idea might break down—produced a much wider, more diverse spread of ideas. Interestingly, traditional rubrics completely failed to reward this originality. Because the grading scale only measured standard marketing targets, unique ideas went unnoticed in the final score.

The group that received both tools delivered the broadest gains, pairing the sharp, expert-like structure of ChatGPT with the distinct reasoning and diversity of the thinking exercise.

Why it matters

Debates around AI in education often trap people into a false choice: teach kids to think, or teach them to prompt.

This study proves they are distinct, complementary levers. ChatGPT handles the heavy lifting of structure, synthesis, and polish, but causal reasoning provides the directional compass and prevents generic, homogenized thinking. Novices can churn out boardroom-ready memos in minutes, but the underlying strategy remains flat unless they know how to stress-test assumptions themselves.

More importantly, the research exposes a glaring flaw in modern evaluation. Traditional grading rubrics are engineered to reward tidy, structured, and conventional answers—the exact traits large language models excel at faking.

We think schools and corporate managers are going to hit a wall if they keep grading final outputs. When polish becomes a cheap commodity, the only meaningful metric left is whether someone asked the right questions and arrived at an insight the machine could not have stumbled onto alone.