What’s the Eberly Center reading and thinking about this month?

The Research and Scholarship Digest, published the first Monday of each month, consists of short summaries of recently peer-reviewed studies on teaching and learning topics. This digest offers a view into what we are reading and thinking about at the Eberly Center that:

• adds to our understanding of how students learn
• is potentially generalizable across teaching contexts in higher education
• provokes reflection on implications for our teaching and educational development practices.

We hope the readers of this digest will find it a useful resource for staying in-tune with the rapidly expanding education research literature.

Chen, S. C., Schunn, C. D., Cao, J., Banihashem, S. K., Gašević, D., & Fan, Y. (2027). GenAI-supported internal feedback intervention for writing improvement: The roles of evaluative judgment and student-AI interaction in a cluster-randomized experimental study. Computers & Education, 105767. https://doi.org/10.1016/j.compedu.2026.105767

The authors of this paper examined the effects of GenAI-supported internal feedback on writing performance in a 14-week cluster-randomized experiment involving 368 undergraduates. The experimental group compared their own writing with GenAI generated writing samples, used GenAI to evaluate their own writing, critically assess its feedback, and revise their drafts; while the control group received conventional teacher feedback. The experimental group achieved significantly higher scores on an independent, AI-free writing post-test. Within the experimental group, evaluative judgment and prompt complexity, rather than interaction frequency, were more strongly associated with writing improvement. However, despite improvements in AI interaction skills, students’ evaluative judgment did not significantly develop. One thing to note is that the intervention combined GenAI feedback with structured evaluative activities, making it impossible to isolate the effects of AI from those of instructional scaffolding.

Di Stefano, G., Ambrosetti, A., & Irrera, L. (2026). Does generative AI support enhance application without improving immediate knowledge acquisition? An experimental study of learning complex theoretical models. Scholarship of Teaching and Learning in Psychology. https://doi.org/10.1037/stl0000497

In an experimental study (total n = 90), undergraduate psychology students were recruited to participate and randomly assigned in a stratified manner to a treatment (AI) or control (no AI) condition. Students in the treatment condition used a custom AI tutor to learn about a complex framework in occupational psychology (job demands-resources model), while students in the control condition used traditional digital learning materials. All students were assessed afterward on both theoretical and applied levels. Results showed that students in the AI condition scored significantly higher on the elaborative/applied level assessments with a large effect size (d = 0.81), but there was no difference in general knowledge acquisition. Furthermore, participants completed AI perceptions measures that captured students’ perceived usefulness of AI and perceived threat of AI to their learning, neither of which moderated the results (i.e., students attitudes toward AI did not affect whether it helped their learning). The authors discuss different interpretations of these results and emphasize the importance of well-designed AI-enhanced learning environments.

Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance. arXiv preprint arXiv:2604.04721. https://doi.org/10.48550/arXiv.2604.04721

In three experiments (total n = 1,222), researchers tested participants’ ability to persist and solve problems when AI assistance either was or was not available. Participants were recruited from a crowdsourcing marketplace and asked to complete a pretest, learning, and test phase. In experiments 1 and 2, participants were asked to solve mathematical fraction problems. They could skip problems if they chose to. For participants in the AI condition, an AI assistant was available during the learning phase that was programmed to give the participant the answer if wanted it. In the control condition, there was no AI assistant. For the testing phase, AI was removed altogether, and researchers measured how successful each group was at solving problems, and how much they persisted (i.e., did not skip problems). In the third experiment, a similar protocol was used with reading comprehension tasks instead of fraction problems. Overall, results showed that participants in the AI condition who prompted the AI assistant to just give them the answers had a lower solve rate and a higher skip rate than participants in the control condition. 

Paul, C., Webb, D. J., & Chessey, M. K. (2026). Missing data on physics exams: Demographic patterns, course-level predictions, and implications for equity. Physical Review Physics Education Research, 22(2), 020149. https://doi.org/10.1103/b4q5-dkn5

Researchers analyzed nearly 10 years worth of data from over 15,000 students in introductory physics courses to undercover demographic patterns in students who left exam and quiz questions blank. They found that students from historically marginalized groups disproportionately left more questions blank compared to their peers, and that this was not correlated with a lack of understanding or overall performance in the course. Through their analysis, they concluded that uncollected (blank) test questions correlated more with course structure and advocated for a course deficit model (as opposed to a student deficit model) when considering equity gaps. They argue that blank responses do not demonstrate a lack of student understanding, rather an inability for the instrument to collect meaningful data, which should prompt instructors to consider alternative grading strategies that allow all students to best demonstrate their knowledge.