As generative AI becomes a routine learning tool, analysis suggests a new educational task is emerging: not just using it a lot, but using it effectively.

Education materials provided by OpenAI on Sept. 11 show that researchers at Italy's Bocconi University, working with OpenAI Economic Research, ran a randomised experiment on 1,053 first-year undergraduate students majoring in economics, business administration and finance to compare how ChatGPT and causal reasoning training affect learning.

Students were randomly assigned to four groups by class session: a group using ChatGPT Edu (GPT-4o), a group receiving causal reasoning training, a group receiving both, and a group receiving neither.

Students who used ChatGPT scored about 0.86 points higher than the control group on assignments graded on a five-point scale. They also presented more ideas, and showed higher logical consistency and greater similarity to expert recommendations.

Causal reasoning training showed a different effect. It did not raise overall assignment scores, but helped students produce more varied and distinct ideas. It also improved their ability to explain why their proposed approaches might work or fail.

The group that received both ChatGPT and causal reasoning training showed both effects together. This suggests that while AI use can improve the completeness of outputs, separate thinking training can supplement diversity and originality of ideas and reasoning ability.

AI use among university students is already spreading rapidly. An analysis by the Korea Information Society Development Institute (KISDI) of Korea Media Panel data showed the share of university and graduate students in South Korea using generative AI rose to 68.0 percent last year from 35.9 percent in 2024, up 32.1 percentage points in a year.

OpenAI said more than 200 million people aged 18 to 24 worldwide use ChatGPT every week. Uses are also expanding, including rechecking concepts not understood in lectures, creating expected questions and quizzes from lecture materials, and identifying weak areas to review before exams.

Heavy AI use does not necessarily mean using AI well. OpenAI explained that even advanced users in the university-age group use ChatGPT's capabilities at only about 1 to 10 percent of the level of power users.

A similar gap appeared in British universities. In a survey released in March by the Higher Education Policy Institute (HEPI), 94 percent of university students said they use generative AI for assessed work such as assignments or exams. While 68 percent saw AI capability as a key future skill, only 48 percent said faculty members support the development of AI skills needed for future careers.

As a result, universities' AI education is shifting away from simply teaching how to use it toward teaching when to ask AI for help, how much work to delegate, and how to verify AI-generated results.

Seoul National University provided ChatGPT Edu to all about 47,000 members this year, including students, faculty and staff. OpenAI last month also released a student education plug-in that can be used in institution-managed environments such as ChatGPT Edu, allowing users to create customised study guides, quizzes, flashcards and interactive visual materials based on lecture materials and documents selected by the student.

OpenAI said students at educational institutions that adopted ChatGPT Edu showed more advanced usage patterns in analysis, calculation and learning than free users, and tended over time to move closer to the usage style of power users.

An OpenAI official said, "AI's role is not to give students faster answers or to study on their behalf, but to help them set their own learning goals, solve difficult problems and use AI in a deeper way."