Information literacy through open pedagogy: What's the role of AI tools?

Synthesis: Clinton-Lisell (2026) asked whether a semester-long open pedagogy project improves college students' information literacy, and whether adding generative AI tools changes the gain. Two sections of an introductory child development course at a US public university worked with the instructor to update citations in a 420-page openly licensed textbook. One section used library databases only (n = 55); the other also used Scite.ai and SciSpace (n = 58), for 113 students whose assignments were analyzed. Perceived source-finding skills rose in both sections, from 3.76 to 4.24 on a five-point scale, F(1, 84) = 47.30, p < 0.001, partial η2 = 0.36, with no time by section interaction, F(1, 84) = 0.04, p = 0.85. The AI section was likelier to have a recent article approved on first submission and to finish both assignments, but scored lower. Students were not randomly assigned, so none of this is causal.

Key Findings

1. Perceived source skills increased across the semester in both sections, F(1, 84) = 47.30, p < 0.001, partial η2 = 0.36, with the pooled mean moving from 3.76 (SD = 0.55) to 4.24 (SD = 0.52) on a five-point scale.
2. AI enhancement added nothing to that growth: the interaction was null, F(1, 84) = 0.04, p = 0.85, partial η2 = 0.00, and mean change was 0.48 (SD = 0.59) in the traditional section against 0.50 (SD = 0.71) in the AI-enhanced section.
3. The sections did not differ overall on perceived source skills, F(1, 84) = 2.99, p = 0.09, partial η2 = 0.03, a difference the paper reports as unreliable.
4. For the recent article, the distribution of submission attempts differed, χ2(2) = 6.56, p = 0.04, Cramer's V = 0.25: 81.0% of the AI-enhanced section had an article approved on the first submission against 67.3% of the traditional section, and 1.7% submitted none against 14.5%.
5. For the historical article there was no such difference, χ2(2) = 1.40, p = 0.50, Cramer's V = 0.14, with 75.9% and 69.1% approved on the first attempt.
6. Summary table scores were high in both sections, 174.73 (SD = 15.40) and 173.86 (SD = 8.87) out of 180, but the distributions differed, Mann-Whitney U = 618.00, p < 0.001, r = 0.44, with the traditional section ranking higher at a mean rank of 65.27 against 39.54.
7. Completion favored the AI-enhanced section: 96.6% finished both summary tables against 81.8%, χ2(1) = 6.46, p = 0.01, Cramer's V = 0.24.

Two sections, one textbook, no random assignment

The project was embedded in a face-to-face introductory Child Development course in the College of Education and Human Development at the University of North Dakota, across the fall semesters of 2024 and 2025. Students read the open textbook Child Growth and Development, a 420-page California Community Colleges text under a CC BY license, and annotated all 15 chapters in Perusall, flagging citations older than five years. At the end of term each student was assigned a chapter and had to find a scholarly article that could update an outdated source, plus a historical article by or about an underrepresented scholar. A librarian taught one session in each section; the traditional section saw library databases, and the AI-enhanced section also saw Scite.ai for locating articles and SciSpace for understanding them. Completed tables have since been used to revise the textbook.

The comparison is between intact course sections, and the paper says plainly why: students cannot be randomly assigned to college courses. That choice buys realism and costs internal validity, because section membership may carry differences in scheduling, prior library instruction, or student preparation that the analysis does not model. The perceived-skills measure was ten items from the Information Literacy Self-Efficacy Scale, scored 1 to 5, with McDonald's omega of 0.82 at the start of term and 0.90 at the end. Assignment scores came from a rubric, and the total possible across the two tables was 180 points. Of 123 enrolled students, 86 completed both surveys and both tables; 113 consented to having their assignments analyzed.

Growth that both sections shared

The clearest result is the one that has nothing to do with AI. Students in both sections rated their own source-finding skill higher at the end of the semester than at the start, and the effect was large by the paper's own partial η2 of 0.36. The open-ended responses point the same way, and the mean scores on the two summary tables sat near the top of the scale. Students who did the project generally did it well.

What the design cannot do is attribute that growth to the project. There was no section that skipped the open pedagogy assignment, so the paper has no control group for the open pedagogy claim itself, only for the AI enhancement. A semester of ordinary library instruction, practice with databases, and maturation would produce some of the same movement, and the author acknowledges the gap rather than papering over it. The self-report measure compounds the problem: students who have just spent a semester on a task may rate their competence on that task more generously, and no objective information literacy test was administered.

Completing more, scoring lower

The two assignment findings pull in opposite directions and the paper is honest that it cannot reconcile them. Students in the AI-enhanced section were more likely to get a recent article approved on the first try, and far more likely to submit both tables at all: the 14.5% who submitted no second article in the traditional section compares with 1.7% in the AI-enhanced section. Yet their tables scored lower in distribution, with a mean rank of 39.54 against 65.27, driven by a concentration of perfect scores in the traditional section.

Two explanations are offered, and both are labeled conjecture. The first is selection within the assignment: students who would ordinarily have stalled and submitted nothing did submit with AI help, and their work was weak because they were struggling. The second is disengagement, the possibility that easy access to a tool reduced the productive struggle the task was designed to require. A related mechanism the paper raises is fading. A scaffold is supposed to be withdrawn, and students had independent, continuous access to these tools, so no fading was possible. Neither explanation is testable with the measures collected, and the ceiling effect makes the score comparison fragile: with averages near 97%, there was little room for the AI section to score higher.

What students said they learned

Eighty-seven students answered an open-ended item about the skills the project developed, and inductive content analysis produced seven themes. Finding scholarly sources was the most common at 46.0% of responses, followed by reading journal articles at 34.5% and recognition that research findings need to be recent at 24.1%. Critical thinking appeared in 17.2% of responses, writing and citing in 8.0%, and 2.3% said they learned no new skills. The pattern matched the project's stated objectives, and the ACRL frames the course had targeted were visible in what students described.

One difference between sections is large enough to notice: 45.9% of AI-enhanced responses mentioned reading journal articles, against 20.5% of traditional responses. The author suggests SciSpace may have helped students get through dense articles, a plausible reading that the design cannot confirm. No statistical tests were run on the themes, on the stated ground that repeated testing across seven themes in a small sample invites Type I error, so the section comparison is descriptive only.

What this means for practice

Instructors can take the completion result at face value and the quality result as a warning. Adding generative tools to an open pedagogy assignment appears to keep weaker or more reluctant students in the work long enough to submit something, which matters for a project whose product is a revised textbook that future students will read. It does not appear to raise the ceiling, and in this course it lowered the average rank of finished work.

The practical move is to decide which outcome the assignment is for. If the aim is a completed, usable revision, AI scaffolding in a collaboration coach role is defensible. If the aim is information literacy, the scaffold needs a withdrawal plan: a stage where tools are removed, or a task component that requires the student to justify a source without one. The author's own recommendation is cautious, limited, and monitored use, with students told about the risks.

Limitations

• Sections were intact and not randomly assigned, so differences in scores and submission patterns may reflect who enrolled in each section rather than the tools; the paper makes no causal claim and the design cannot support one.
• Only 86 of 123 enrolled students completed both surveys and both tables, and analyses were restricted to completers. Students who did not submit rarely completed the self-reports, which excludes from the comparison the group the AI tools may have most affected.
• Information literacy was measured by self-report, ten self-efficacy items, plus instructor-scored tables at ceiling. No objective test of searching or evaluation was used, and the near-97% averages leave little variance to explain.
• The sample is one introductory course at one university in the United States, overwhelmingly young White women among those who reported demographics, with two specific AI tools. Generalization to other disciplines, tools, or populations is not supported.

Citation

Clinton-Lisell, V. (2026). Information literacy through open pedagogy: What's the role of AI tools?. Sustainability, 18, 8684.