Students who drafted first and then asked AI for feedback earned the top grades in a Finnish study

The denominator is small. Of 891 students whose take-home exams were studied, 159 reported using AI ^[7], about 17.8% ^[16]. That leaves 732 non-users ^[15] set against three AI groups carved out of the 159. Even an even split would put about 53 students in each ^[17].

The exams came from 2023 and 2024, and students could use AI as long as they disclosed it ^[6]. Disclosure was the only way the researchers knew who used what. If any students used AI without saying so, they would sit among the non-users, and that group would include some AI users.

Within those limits, the ordering in the release is consistent. The three approaches line up by how much work the student did before the AI was involved. Delegators sit at the bottom ^[4] and idea-generators in the middle ^[5]. Students who drafted first and then asked for feedback or missing perspectives sit at the top, ahead of non-users too ^[1]. The release hedges both ends. Delegators' grades "may even be lower" than non-users' ^[4], while reflective use "may lead to significantly better performance" than the other approaches ^[2]. Phys.org's account does not report group sizes, grade gaps, the statistical tests, or whether prior achievement was accounted for.

The release also says AI use may improve performance on average ^[13]. Any such average includes the higher grades of the reflective users and the possibly lower grades of the delegators.

The study cannot show which way the effect runs. Reflective use begins with students producing content on their own ^[1], so that group is made up, by definition, of students who could write a draft unaided. Stronger students may simply choose that method. Separating the method from the student would take a trial that assigns approaches, or at least an adjustment for earlier grades. I'd expect the top-of-table result to hold up under a better design more readily than the delegation penalty, which the release itself states with a "may" ^[4]. A take-home grade also scores work produced with the tool at hand, so it cannot measure what a student can do without the tool.

According to Outi Hakola, the paper's first author ^[14], public debate often assumes students use AI in only one way, and earlier research mostly tested how well AI can complete assignments ^[9]. "This is an overly simplistic view of a complex issue," Hakola said ^[10]. The associate professor's conclusion is about teaching. "This is why it is important to teach students how to use AI effectively and purposefully, without using it as a substitute for their own work or thinking, so that it becomes a useful part of their learning and competence development," Hakola said ^[8].

The tools also changed while the data were being collected. "During the course of this study, the number of AI applications increased and people's awareness of them grew," Hakola said ^[11]. Hakola suggested looking again in about five years at how AI has become part of students' learning ^[12].