On this page

Synthesis: Pal, Jones and Hoover (2026) report a semester-long case study of an undergraduate leadership ethics elective at Texas Tech University in which GenAI was built into the course rather than fenced off at its edges: a flexible use policy that required students to record when and how they used a tool, a first-week debate on the ethics of AI in education, a two-stage case sequence that ran once without technology and once with ChatGPT, and a confidence survey given before and after. Roughly 80 students across two sections took the course, and 38 matched survey pairs were analyzed. Students did not conclude that AI use was unethical; they concluded that unrestricted use was, and they drew the line at a tool doing the work rather than supporting it. On 14 of the 16 ethical-decision-making items, self-rated confidence rose significantly, with the two exceptions being the introspective items about naming and explaining one's own values — a pattern that locates the gain in procedural skill rather than moral self-knowledge, and that the design cannot attribute to the AI component alone.

Key Findings

  • Students policed delegation, not use. Most admitted using GenAI for formatting, brainstorming, explaining confusing concepts and quizzing themselves before exams, and treated those uses as legitimate. What they described as crossing a line was putting an assignment prompt into a tool and submitting the response, which several framed as cheating themselves out of the learning they came to university for.
  • Responsibility was distributed across the room, not delegated to the tool. In the ChatGPT case, one team member wanted the model to produce the annotations comparing the team's analysis with the model's; a teammate objected that this defeated the assignment's purpose and could affect her grade. The team negotiated a standard with the instructor, then annotated the model's annotations with their own commentary.
  • The policy had to be co-created to be understood. The instructor split the class in week one and assigned half to argue for and half against GenAI use in education before discussing the course policy. Documentation was the weak point: students routinely cited the tool without explaining how they had used it, and the instructor had to restate the requirement through the semester.
  • Inconsistent rules across courses pushed students to one of two extremes. The university left AI policy to individual instructors, so students met everything from outright bans to unrestricted acceptance. Some used tools "within reason" while unsure of the parameters; one said she avoids AI entirely so her work can never be challenged as not her own.
  • Worked examples of bias were more persuasive than warnings. A team workshop on hiring had ChatGPT generate candidate profiles and rank them. Students saw bias in the ranking, and most concluded that the model should not make the decision alone. Another team's scenarios covered monitoring employees, criminal justice and grading student work.
  • The model was a useful second reader with a blind spot. Several teams had missed stakeholders in their case analyses. ChatGPT surfaced some of them but also glossed over moral components the students thought central, which is what led students to argue for human oversight and collaboration rather than delegation.
  • Confidence rose on procedure, not on self-knowledge. Self-rated ability to recognize the ethical dilemma moved from 7.24 to 8.39 and to make a sound and justifiable decision from 7.24 to 8.45, both at p < .001 on a 0–10 scale. Using ethical theories and terminology to explain a decision rose from 5.39 to 7.08. Identifying one's own key virtues, values and rights (p = .063) and explaining how personal values affect a decision (p = .124) did not move significantly.
  • The course framed this as AI literacy for work, not only for school. Three of the 16 final team workshops chose AI in the workplace as their topic and five cited AI in their projects, most often to build discussion scenarios for classmates.

How the study was done

The case is bounded by one course in one semester: an upper-level elective in an undergraduate management degree, taught by one instructor across two sections in spring 2025, with approximately 80 students enrolled. Students completed two team cases. The first required an ethical-decision-making process with no technology. In the second, teams drew a case prompt from ChatGPT, worked the same process themselves without technology, then walked ChatGPT through the process and annotated the differences between its response and theirs, closing with a reflection on what felt right and what a more ethical outcome would have been. The course ran on a whole person learning frame with four components — intellectual, emotional, behavioral and spiritual actualization — drawn from Hoover's work on experiential learning, and the authors argue that the AI assignments were designed to reach beyond cognitive performance.

The quantitative strand is a researcher-created confidence instrument of 16 items covering the ethical-decision-making process the course teaches, rated 0 (cannot do at all) to 10 (highly certain can do) following Bandura's guide for constructing self-efficacy scales. It was administered in week one, before the first non-AI case, and again immediately after the AI-assisted case. Matching relied on student-created identifiers, the survey was optional, and unmatched responses were dropped, leaving 38 paired samples. A sign test was run in SPSS 30 for each item. Qualitative data were the instructor's reflective observations of sessions and discussions, team case submissions, workshop materials and the syllabus; validity rested on triangulation, reflexivity about the instructor's dual role, and thick description.

What this means for practice

  • If you write an AI policy, specify documentation, not just permission. The failure mode in this course was not students hiding their use but students citing a tool without saying what they did with it. A policy that asks for both — when and how — gives you something to respond to and gives students a concrete standard to meet.
  • A course-level rule does not travel. Where an institution leaves policy to instructors, students carry a patchwork. Saying explicitly how your rule differs from the course next door, and why, costs one discussion and removes a source of quiet non-compliance.
  • Design the comparison as the work. The learning here came from students testing their own reasoning against the model's and annotating the gap. An assignment where the model performs the comparison produces an integrity dispute instead.
  • Sequence without the tool first. Running the same case unaided and then with ChatGPT gave students a baseline to compare against; it also made the model's contributions and omissions visible in a way a single AI-assisted attempt would not.
  • Read confidence gains carefully. Gains clustered on the analytical steps the course drilled and stopped at the two items that ask students to know their own values. If you use a similar survey, treat movement on procedural items as a check on instruction rather than as evidence that the AI component worked.

Limitations

  • The paper states its own limits: one course, one instructor, one university, so transferability is limited, and other elements of the course design — cases, debate, team workshops, reflection — plausibly affected confidence as much as the AI integration did.
  • The instructor of record was also a researcher, and the authors acknowledge that her values and perceptions may have shaped collection and analysis; because she was a participant-observer, students' voices are paraphrased rather than directly quoted.
  • Only 38 of roughly 80 students produced matched survey pairs on an optional instrument, and the responses are self-reports of capability rather than observations of ethical reasoning.
  • There is no comparison group, so the paper cannot separate the AI-assisted case from the ordinary effects of a semester of practice, maturation or repeated testing.
  • The two items that did not reach significance started from the highest pretest means (7.42 and 8.08), which is consistent with a ceiling effect as much as with a genuine boundary on what the course changed; the paper does not discuss this.
  • The classroom evidence is reported as themes from instructor observation, without a coding scheme, member checking or inter-rater procedure, so the qualitative findings rest on one instructor's reading of her own class.

Citation

Pal, P. R., Jones, S. J., & Hoover, J. D. (2026). Ethically Aware Learning in an AI-Enabled Classroom. Journal of Teaching and Learning with Technology, 14(Special Issue), 146–163.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.