Does knowing AI ethics mean students use it responsibly?
A 350-case study in two Nigerian universities found rising self-reported awareness, reflection and responsible use across three annual waves. But changing participants and the absence of observed behaviour prevent a claim that individual students became safer AI users.
The short version
- What happened
- The analytic dataset contained 350 complete cases across three annual measurement occasions at two public universities in southeastern Nigeria; departing participants were replaced with demographically matched students.
- Who it affects
- Teachers can distinguish knowledge, reflection and demonstrated conduct when assessing AI use.
- What remains unproven
- A rotating cohort does not establish individual change; replacement and period effects remain entangled.
Keep in mind
A rotating cohort does not establish individual change; replacement and period effects remain entangled.
All principal outcomes were self-reported, with no observed compliance or practical safety test.
There was no standardised intervention, randomised comparator or causal measure of AI exposure.
Only two universities and computer science programmes were studied; data are available by request.
Source analysis · 1 linked source. Primary analysed source date: 12 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: awareness and conduct still need separate evidence
Students in two Nigerian public universities reported higher AI ethics awareness, more reflection and more responsible use over three annual survey occasions. The three measures did not develop as one tightly coordinated process. The study therefore supports treating knowledge, judgement and perceived practice as separate educational targets. It cannot show that knowing ethical principles made students use AI responsibly, because responsible use was itself measured by students' descriptions of their conduct.
That distinction matters for teachers. A student can correctly describe privacy, bias or authorship and still struggle to apply those ideas under assessment pressure. Conversely, cautious practical behaviour may develop without a high score on an awareness questionnaire. This research helps frame that problem, but it does not measure how often either pattern occurred in actual coursework or identify a teaching method that reliably closes the gap.[1]
What the 350-case denominator represents
The participants came from undergraduate computer science programmes at two public universities in southeastern Nigeria. The setting included introductory programming, object-oriented programming and software-development courses using projects, debugging, peer review and problem-solving. AI tools were optional learning supports. Ethical issues were discussed periodically, but no standardised ethics curriculum or controlled training programme was introduced.
Surveys were administered in scheduled classes at approximately 12-month intervals, during the second semester of each academic year. Research assistants independent of the instructors coordinated anonymous Google Forms questionnaires lasting around 20–25 minutes. The paper describes 350 complete analytic cases for nine composite scores: three constructs at three waves. This should not be described as 350 identified students followed continuously for three years. Students who graduated, transferred, withdrew or took leave were replaced using gender, age category, academic level and programme matching.[1]
The finding is an aggregate change, with an important design constraint
On six-point response scales, mean self-perceived responsible use rose from 3.16 to 3.97, awareness from 2.99 to 3.75, and reflection from 2.89 to 3.60. These are questionnaire scores, not percentages of compliant work or successful ethical decisions. The authors used confirmatory factor analysis to assess the proposed constructs and linear latent growth models to describe change and relationships among initial levels and slopes.
The rotating cohort creates a substantive qualification. Demographic matching can maintain a similar-looking sample, but it does not establish that replacement participants have the same experiences, prior knowledge or response tendencies as those leaving. A complete matrix of scores also does not restore individual continuity. Changes may combine maturation, different people, broader cultural shifts and increasing familiarity with AI. The paper acknowledges that its growth parameters describe successive cohort compositions rather than isolating within-person development.[1]
What the survey measures leave untested
Responsible-use items asked about checking output, respecting privacy, maintaining judgement and taking responsibility for submitted work. Such questions capture ethical self-appraisal. The study did not inspect code submissions, verify disclosures, test whether students detected fabricated claims, or observe whether they protected confidential information in a practical task. Anonymous administration and independent assistants reduce evaluation pressure, but cannot eliminate socially desirable answers or a common tendency to rate oneself positively across several scales.
The authors report predominantly weak and mixed relationships among growth processes. That limits a simple story in which awareness rises first, reflection follows and responsible conduct automatically results. However, weak survey relationships do not prove that ethics teaching is ineffective. There was no randomised comparison, defined intervention or measured exposure capable of answering that causal question. Contextual interviews were not transcribed and systematically analysed, so they cannot supply a separate explanatory qualitative result.[1]
A practical response for teaching and assessment
For educators, the useful implication is to assess what students do alongside what they say they know. A programming task can require a brief AI-use disclosure, identification of one unreliable suggestion, verification against an original source and an explanation of why the final decision remains the student's. These are possible applications of the study's distinctions, not interventions proven effective by this research. Teachers should test whether they improve observable judgement before making larger claims.
Assessment briefs also need specific permission rules. Students cannot demonstrate compliance consistently when one course permits generated code and another leaves the boundary unclear. A rubric can separately consider verification, privacy, attribution and independent reasoning. The aim should be usable feedback rather than scoring an individual's morality from a questionnaire. This is particularly relevant for teachers preparing future clinicians or public-service staff, whose decisions may affect people beyond the classroom, although those professions were not sampled here.[1]
Local context, disclosures and the evidence that comes next
The Nigerian setting adds evidence from a region often missing in AI education debates. Infrastructure, language coverage, access costs and dependence on externally developed tools can shape what responsible use is feasible. The distinction between awareness and enactment may travel across countries, but the numerical growth estimates should not be exported to another discipline or national education system without testing that the measures mean the same thing there.
The researchers report no specific grant and no competing interests. Ethics approval came from the University of Nigeria's Faculty of Education committee, with consent at each wave. They disclose NotebookLM assistance in refining a conceptual figure and say the underlying data are available on reasonable request. A stronger next study would preserve participant linkage, report attrition and replacement counts transparently, test a defined curriculum against a comparator and include observed behaviour. Improvement in those outcomes would strengthen the case for a teaching approach; rising self-report alone would leave it unresolved.[1]