Ethical Generative AI Integration in English for Academic Purposes within Higher Education: A Mixed-Methods Study

Synthesis: Alharbi et al. (2026) respond to the fact that roughly 60% of higher education institutions report faculty using Generative AI tools in teaching while fewer than 15% have implemented formal ethical guidelines. Using a convergent mixed-methods design guided by Mezirow's Transformative Learning Theory, the authors designed, implemented, and evaluated an eight-week GenAI-integrated professional development program at a multicultural Global South institution, engaging 97 higher education educators in quantitative testing and 15 in qualitative interviews. Quantitative results showed significant gains in ethical awareness (d = 0.93) and digital andragogical competence with high inter-rater reliability; qualitative findings confirmed persistent dilemmas such as plagiarism and student overreliance on AI, and surfaced an unexpected barrier — reconciling institutional policy gaps with personal ethical values. The study addresses gaps in Global South evidence and field-tested professional development models, offering recommendations on AI policy literacy, peer-led communities of practice, and shared AI resources.

Key Findings

• Large gains in ethical awareness and reasoning: ethical awareness scores (Scenario Rubric) rose from a pre-test mean of 3.21 to 4.05 post-program, t(96) = 15.12, p < .001, with a large effect size (Cohen's d = 0.93); Artifact Rubric scores rose from 2.94/5 to 4.21/5 with high inter-rater reliability (κ = 0.87).
• Plagiarism and student overreliance persist as core dilemmas: educators cited difficulty navigating the absence of institutional Generative AI policies (15 mentions) and uncertainty over acceptable AI use (12 mentions) as primary challenges, confirming that Academic Integrity concerns shift but do not disappear.
• An unexpected barrier emerged — reconciling policy gaps with personal values: beyond technical dilemmas, participants struggled to align institutional policy gaps with their own ethical values, a finding that complicates simple compliance-based approaches to Ethics.
• Policy familiarity jumped dramatically: affirmative responses on knowledge of institutional AI policy rose from 27% to 79% (p < .001, McNemar's test), and higher policy awareness was associated with stronger Alignment with Policy scores in artifacts (r = .36, p < .05).
• Engagement amplifies gains: higher engagement (an index of attendance, task completion, and reflective participation) correlated with higher post-test scores and rubric performance (r = .41, p < .01 for plagiarism/authorship scenarios), and highly engaged participants positioned themselves as "change agents" in their departments.
• Educators proposed institutional strategies: participants recommended establishing AI ethics committees (12 mentions), embedding AI policy literacy into professional development (11 mentions), and creating centralized repositories of vetted Generative AI tools (10 mentions), plus ongoing rather than one-off training.

Study Design & Method

This is an empirical mixed-methods study using a convergent design within a pragmatist paradigm, integrating quantitative evidence of change with qualitative accounts of experience. The quantitative strand used stratified random sampling to recruit 97 higher education educators in an English for Academic Purposes (EAP) unit at a multicultural Global South institution (staff from the Gulf, South Asia, Central Asia, and Africa); the qualitative strand used purposeful maximum variation sampling to select 15 participants for semi-structured interviews. Original instruments included an adapted 24-item Likert survey (Cronbach's α = 0.81–0.87), a Scenario Rubric, an Artifact Rubric (inter-rater reliability κ = 0.85), and an engagement index. The eight-week professional development program was explicitly aligned with Mezirow's Transformative Learning Theory, sequenced from disorienting dilemmas (weeks 1–2) through reflective discourse (3–4), communicative learning and peer assessment (5–6), to perspective transformation and reintegration (7–8). Qualitative data were analyzed via Braun and Clarke's (2006) six-phase thematic analysis (intercoder κ ≈ 0.85), with integration occurring during the interpretation phase through triangulation. Rigor was ensured via member checking, thick description, audit trails, and reflexive journaling; the single-institutional setting limits statistical generalizability but supports analytical generalizability to comparable Global South contexts.

What this means for practice

• Instructors. Embed structured ethical reasoning inside GenAI professional development rather than treating technical and ethical skills as separate tracks — the eight-week program raised ethical awareness from a mean of 3.21 to 4.05 (Cohen's d = 0.93) and artifact quality from 2.94/5 to 4.21/5.
• Instructors. Use case-based analysis, reflective dialogue, and artifact development to convert awareness into classroom application, following the program's sequence from disorienting dilemmas through reflective discourse and peer assessment to perspective transformation.
• Instructors. Teach AI policy literacy explicitly, especially where formal governance is underdeveloped: affirmative knowledge of institutional AI policy rose from 27% to 79%, and higher policy awareness was associated with stronger policy-aligned artifacts (r = .36).
• Faculty developers. Design for participation rather than one-off training: engagement correlated with post-test scores (r = .41 for plagiarism and authorship scenarios), so build structured participation mechanisms instead of relying on attendance.
• Administrators. Establish AI ethics committees and centralized repositories of vetted Generative AI tools, embed policy literacy across all training, and treat highly engaged participants as catalysts for wider cultural change.

Limitations

• The program ran at a single multicultural Global South institution with 97 educators in the quantitative strand and 15 interview participants, so the authors claim analytical generalizability rather than statistical generalizability.
• Self-selection may have favored educators already predisposed toward innovation, and the absence of longitudinal follow-up means sustained classroom transformation and student-level impact are unmeasured.
• Outcomes rest on self-reported instruments — an adapted 24-item Likert survey (α = 0.81–0.87) and rubric-scored artifacts — rather than on observed changes in teaching or learning.

Citation

Alharbi, W., Hamid, S., Abbas, S., & Salieva, Z. (2026). Ethical Generative AI integration in English for Academic Purposes within higher education: A mixed-methods study. Journal of University Teaching and Learning Practice, 23(5). (CC BY-ND 4.0.)