Ethical Generative AI Use in Higher Education: An Action Research Study Using the Think Aloud Method

Synthesis: Karaki and Baroudi (2026) ran a mixed-methods action-research study with 44 first-year entrepreneurship students at a UAE university, using the Think Aloud method — verbalizing reasoning while interacting with generative AI — as a pedagogical intervention to support metacognitive awareness and ethical engagement. Think Aloud made students' prompt construction, verification practices, and ethical decision-making more visible, and self-reported metacognitive awareness was the highest of three surveyed constructs (M = 3.79). But critical evaluation of AI outputs often remained at a surface level, particularly among students needing additional support: students spotted generic or misaligned responses ("too general," "missing details") yet continued relying on them heavily. The authors' conclusion — Think Aloud is not a stand-alone solution and must be combined with explicit instruction in AI literacy, prompt construction, and verification strategies — extends the process-evidence literature from what records capture to what vocalization alone changes. No control group, no pre-test: the single post-intervention survey measures perceptions, not measured change.

Summary

As generative AI becomes embedded in higher education, it presents challenges of ethical use, academic integrity, and over-reliance on AI-generated outputs. The Think Aloud method — originally a psychological research tool for protocol analysis (Ericsson & Simon, 1993), since adapted as an instructional strategy — offers a way to make students' reasoning visible as they interact with AI: articulating how they construct prompts, evaluate responses, recognize limitations, and decide whether to accept, modify, or reject AI content. The study asked what role Think Aloud plays as a pedagogical intervention in supporting students' metacognitive awareness and ethical engagement with generative AI in a first-year entrepreneurship course.

The exploratory mixed-methods action-research study ran in a design-thinking-grounded entrepreneurship course over one semester, in two implementation cycles (the second refined after the first). Students applied Think Aloud during prototype and test activities, attending to prompt construction, evaluation of AI-generated content, and ethical decision-making. Data came from reflective journals, six classroom observation sessions (three per section), and a post-intervention survey completed by 38 of 44 students. Qualitative data were analyzed thematically (sensitized by Rest's Four-Component Model); survey responses used descriptive statistics plus an exploratory correlation between perceived metacognitive awareness and self-reported ethical AI use. An observation protocol coded verbalization patterns, prompt construction, verification, ethical decision-making, and self-correction.

Key Findings

• Think Aloud made reasoning processes visible. Qualitative data showed Think Aloud supported students' emerging awareness of prompt construction, verification practices, and ethical decision-making — verbalization helped students articulate how they construct prompts, evaluate responses, and decide whether to accept, modify, or reject AI output.
• Critical evaluation stayed surface-level for many students. Despite progress, "students' evaluation of AI-generated content often remained at a surface level" — students identified basic weaknesses ("too general," "missing details," "not fully aligned with the course requirements," "inaccurate or inconsistent") but observation notes indicated continued heavy dependence on AI. The pattern was strongest among students requiring additional support.
• Survey perceptions were moderately positive across three constructs. Self-reported metacognitive awareness was highest (M = 3.79, SD = 1.19), followed by critical evaluation of AI-generated content (M = 3.73, SD = 1.12) and ethical AI use (M = 3.68, SD = 1.09) — but the single post-intervention administration means these are perceived development, not measured pre-post change.
• The awareness–ethics correlation is methodologically entangled. The exploratory correlation between perceived metacognitive awareness and self-reported ethical AI use was interpreted cautiously because both constructs were measured through the same post-intervention self-report survey — common-method variance limits what the association can show.
• Over-reliance has two faces. The authors frame student AI dependence as not automatically misuse: over-reliance risks cognitive offloading and reduced independent thinking, but students who lack confidence, language fluency, or academic experience may use AI because they need support — the pedagogical challenge is distinguishing risk from need.
• Process-based ethics over compliance. The paper's framing positions Think Aloud within a shift "from reactive compliance toward process-based ethics" — an internal, metacognitive framework in which moral agency remains with the learner throughout the task, against the detection-and-punishment models dominant in academic-integrity policy.

What this means for practice

• Instructors. Think Aloud is a usable classroom intervention for making AI-related reasoning visible — have students verbalize prompt construction, response evaluation, and accept/modify/reject decisions during AI-supported tasks — but do not expect vocalization alone to produce deep evaluation: the study's central caution is that it must be combined with explicit teaching of AI literacy, prompt construction, verification strategies, critical judgment, and repeated guided reflection.
• Faculty developers. First-year courses are the named opportunity: responsible AI habits form early, and students needing additional support are exactly the ones for whom surface-level evaluation persists — targeted scaffolding, not general exhortation.
• Curriculum designers. Process-based ethics reframes the goal: design tasks where moral agency stays with the learner throughout (not a compliance checkbox at the end), and where accept/modify/reject decisions are rehearsed, not assumed.
• Instructors supporting multilingual or less-prepared students. The over-reliance-as-need framing argues against treating dependence as automatic misconduct — diagnose why a student leans on AI before correcting the behavior.
• Researchers. The common-method caveat (awareness and ethics measured in one survey) is the study's own honest limit — and the single post-test design means "perceived development" is the strongest claim the survey data support.

Limitations

• No control group, no random assignment, no pre-test/post-test design: the results "do not imply causation and cannot be generalized beyond the study context" — the authors state this explicitly.
• The post-intervention survey was administered once, so quantitative findings are students' perceived development rather than measured change.
• Small sample (44 students, 38 surveyed) in a single course at a single institution (a UAE federal university), in one discipline (entrepreneurship).
• Common-method variance: perceived metacognitive awareness and self-reported ethical AI use were measured in the same survey, so their correlation cannot be cleanly interpreted.
• The focal course context (design-thinking entrepreneurship with prototype/test activities) shapes where Think Aloud was applicable — transfer to lecture-based or assessment-only courses is untested.
• The intervention's effect was uneven by student: the surface-level evaluation finding is itself the study's key limitation as an intervention — it worked unevenly, and most weakly where support needs were greatest.

Citation

Karaki, S., & Baroudi, S. (2026). Ethical generative AI use in higher education: An action research study using the Think Aloud method. Journal of Computer Assisted Learning, 42(6), e70352. https://doi.org/10.1002/jcal.70352