Three-Year Study Reveals AI Ethics Skills Grow Unevenly in Students

Artificial intelligence has swept into university classrooms faster than almost any technology in living memory, and nowhere is that more visible than in computer science programs, where generative assistants, code-completion systems, and automated debugging tools have become part of everyday coursework. But a pressing question has lingered behind the convenience: do students actually develop ethical judgment as they gain experience with these tools, or does awareness of bias, privacy, and authorship simply fail to keep pace with usage? A new longitudinal study from Nigeria offers one of the most detailed answers yet, and its findings are more surprising than educators might expect.

Researchers led by Musa Adekunle Ayanwale of the University of Pretoria tracked 350 undergraduate computer science students across two Nigerian public universities over three years, measuring their ethical engagement with artificial intelligence at annual intervals. The study, published in Discover Artificial Intelligence, is notable for its methodology as much as its conclusions. Rather than treating ethical competence as a single skill, the team separated it into three distinct dimensions: AI Ethics Awareness, which covers knowledge of ethical principles and attitudes toward them; Responsible AI Use, which captures students’ self-perceived alignment with fairness, privacy, transparency, and accountability expectations; and AI-Driven Reflection on Decision-Making, which measures how deeply students critically evaluate AI outputs and reconsider their own reasoning when working with these systems.

The results showed something striking. All three dimensions rose substantially over the three annual measurement waves. Responsible AI Use climbed from an average of 3.16 to 3.97 on a six-point scale, AI Ethics Awareness rose from 2.99 to 3.75, and AI-Driven Reflection increased from 2.89 to 3.60. When the researchers standardized these changes, the two-year effect sizes were large, ranging from 1.26 to 1.60. In plain terms, successive cohorts of students reported dramatically higher ethical awareness, more reflective engagement, and greater perceived responsibility in their AI use as they progressed through their degrees.

But here is where the story takes an unexpected turn. When the team examined whether these three dimensions grew together—whether students who started with high ethical awareness developed responsible practices faster, or whether growth in reflection tracked growth in reported responsibility—the answer was essentially no. The correlations among the growth factors were weak, mixed, and sometimes negative. Initial ethical awareness was slightly negatively associated with subsequent growth in both responsible use and reflection. Growth in responsible use was negatively correlated with growth in awareness. Higher starting points did not predict faster development, and there was no evidence of a compensatory pattern in which students who began with lower awareness caught up more quickly.

This decoupling matters because much of the existing literature has implicitly assumed a tidy developmental sequence: students learn about AI ethics, they reflect on it, and then they act responsibly. The study’s findings challenge that assumption at its foundation. Drawing on Transformative Learning Theory as an interpretive lens, the researchers had hypothesized that critical reflection on disorienting encounters with AI—biased outputs, uncertain authorship, conflicts between efficiency and integrity—might drive perspective transformation and behavioral change. The growth in awareness and reflection is compatible with that theory, but the weak alignment among the three trajectories means awareness and reflection cannot guarantee responsible practice. The theory, the authors conclude, works best as a heuristic lens rather than a deterministic roadmap from knowledge to action.

The researchers used latent growth curve modeling, a statistical framework from structural equation modeling that estimates both initial levels and rates of linear change across repeated measurements. Time scores were fixed at zero, one, and two for the three annual waves, allowing the team to model an intercept representing Year 1 status and a slope representing average annual change. Before any growth modeling, they confirmed the measurement structure with confirmatory factor analysis, achieving strong fit indices including a comparative fit index of 0.961 and a root mean square error of approximation of 0.048. They also established full longitudinal measurement invariance across the three waves, meaning changes in scores over time reflected genuine developmental differences rather than shifts in how students interpreted the survey items.

The study’s design deserves careful attention. Because participant attrition from graduation, transfers, and withdrawals is inevitable in multi-year educational research, the team used a rotating-cohort design, replacing departing students with demographically matched participants from the same programs. This preserved sample size and cohort composition but came at a cost: the resulting growth parameters describe aggregate trends across successive cohort compositions rather than uninterrupted individual trajectories. The authors are explicit about this limitation. The findings show that successive cohorts reported higher ethical engagement, not that every student improved, and the observational design cannot establish that AI exposure or any particular teaching practice caused the increases.

There is also an important caveat about what the responsible-use measure actually captures. Because it relies on self-report, it reflects students’ ethical self-appraisal rather than independently verified behavior. Students may genuinely have become more responsible, but their answers could also be shaped by social desirability, self-enhancement, or shifting interpretations of what counts as responsible AI use in a landscape where institutional rules remain ambiguous. The researchers took procedural precautions—anonymous administration, research assistants independent of course instructors, attention-check items—but they acknowledge that common-method bias cannot be fully excluded.

The socio-technical framing of the study helps explain why the three dimensions developed asynchronously. Students’ decisions about AI are embedded in assessment rules, institutional policies, disciplinary expectations, peer norms, and technological affordances. A student may fully understand algorithmic bias yet remain uncertain how to apply that understanding when acceptable AI assistance is poorly defined, or may critically evaluate AI outputs while still adopting expedient practices that prevailing assessment conditions reward. In the Nigerian context, the researchers note, these dynamics are compounded by uneven infrastructure, imported technological standards, and limited institutional governance—conditions that decolonial AI scholarship warns can reproduce external values unless local knowledge and communities shape how AI is evaluated and used.

The practical implications are concrete. The authors argue that universities should not rely on ethics-awareness initiatives alone. Instead, they recommend explicit AI-use policies that specify permitted, restricted, and prohibited forms of assistance for every assessment; authentic ethical tasks requiring students to disclose AI assistance, verify generated information, and identify bias and privacy risks; assessment rubrics that evaluate transparency, fairness, human judgment, and accountability; and behavioral evidence that goes beyond self-report. Institutional governance, they suggest, should involve students, lecturers, librarians, data protection specialists, and academic integrity committees, and should address locally relevant concerns such as language representation, data ownership, and technological dependence rather than treating AI ethics as a generic plagiarism problem. The deeper message of the study is that ethical engagement with AI is multidimensional, asynchronous, and shaped by institutional conditions—and that closing the gap between knowing what is right and doing it requires more than awareness. It requires educational environments deliberately designed to make responsible practice possible, visible, and rewarded.

Subject of Research: Longitudinal development of students' ethical engagement with artificial intelligence in higher education

Article Title: Cohort-level growth modeling of ethical engagement with artificial intelligence in education

Article References: Ayanwale, M. A., Omeh, C. B., Mnguni, L. E., & Oyeniran, F. M. (2026). Cohort-level growth modeling of ethical engagement with artificial intelligence in education. Discover Artificial Intelligence, 6(1), Article 1439. https://doi.org/10.1007/s44163-026-02298-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02298-6

Keywords: AI ethics, higher education, computer science education, latent growth curve modeling, responsible AI use, transformative learning theory, longitudinal study, Nigeria, generative AI, academic integrity, socio-technical systems, moral development