A 90-minute GenAI literacy course improved knowledge, prompting, source checking and self-efficacy across 65 university sections

Connell Pensky and colleagues test a practical question for universities: can a short, reusable online course improve students' capacity to work with generative AI across disciplines? Their 2026 article in Computers & Education evaluates four asynchronous modules designed at Carnegie Mellon University. The approximately 90-minute sequence combined explanations, examples, practice and immediate feedback. It addressed how generative-AI systems work, why appropriate use depends on the person and task, ethical concerns, strategies for responsible educational use and students' confidence in applying those strategies.

The study involved 1,368 undergraduate and graduate students in 65 sections of 53 courses taught by 46 instructors. Course sections, rather than individual students, were assigned to complete the modules between a pre-test and post-test or to a wait-list control condition. That design reduced the chance that students in one section would receive different versions of the intervention, while the analysis accounted for students being nested within courses. Measures included knowledge questions, self-efficacy ratings and authentic tasks for prompt engineering and output evaluation. A randomly selected subset of 174 students had authentic-task responses scored independently without raters knowing the condition or testing time.

Students who completed the modules improved more than control students in knowledge of how large language models work, prompt-engineering skill and self-efficacy. The published abstract also reports gains in fact- and source-checking. Effects were not uniform across every outcome: the modules did not improve the skill of critically evaluating potential bias in generated output, and the university's study report found no added improvement in responsible-use knowledge or overall output analysis. This contrast is important. A concise module can build a useful foundation, but knowing model basics and writing a better prompt are not the same as judging fairness, omissions, evidence or downstream harm.

The authors report that improvements were equitable across the examined categories of birth sex, race and ethnicity, student level, first-generation status and academic discipline. That finding means the analysis did not detect differential benefits across those groups in this setting; it does not prove that one module will remove every access or participation gap. The study was conducted at one university with volunteer instructors, and the post-test followed soon after the intervention. It therefore does not establish long-term retention, transfer into real coursework, changes in academic performance or responsible behavior when tools and incentives differ. Generative-AI products also change quickly, so examples and assessments require maintenance.

For Hong Kong universities, the intervention offers a feasible starting architecture: a shared foundational module, embedded practice, immediate feedback and authentic tasks that require students to create prompts and inspect outputs. Institutions should add discipline-specific cases, multilingual examples and repeated source, bias and uncertainty checks across a semester. Evaluation should include delayed assessments, course artifacts and observed application, not completion rates alone. The result is encouraging precisely because it is bounded: short training can improve several competencies at scale, while critical judgment needs sustained, contextual practice.