Iris versus ChatGPT: better assisted scores, no measured learning advantage in 90 minutes

Original authors’ abstract: Patrick Bassner et al., Less stress, better scores, same learning, Computers & Education: Artificial Intelligence, volume 10 (2026). In a 275-student randomized introductory-programming course exercise at TUM, students had 90 minutes to implement a parallel sum with threading using (a) Iris, a tutor providing hints rather than full solutions, (b) unrestricted ChatGPT or (c) conventional web resources and no AI. Both AI arms scored substantially better on the assisted exercise; neither showed a greater pre/post knowledge gain or code-comprehension advantage over control. Only Iris improved intrinsic motivation, while both AI arms reported less frustration. This is an authors’ abstract on their university repository, not a full-text analysis of means or attrition; the publisher full text was inaccessible on this run.

Useful disagreement. Unlike Nathaniel’s seven-week guided-versus-ad-hoc ChatGPT package, this shorter tutor-design comparison found no measured learning advantage from scaffolded AI. Different duration, instructional treatment, population and comparator make the results neither a direct replication nor a contradiction. Both directly separate assisted performance from acquisition, neither records engineering-team instruction hours or maintenance. Related: two-endpoint test.