Research Article
Exploring prompt literacy in allied health professions education: an explanatory sequential mixed-methods study of student and educator perspectives
Synthesis: At one public university in Ghana, GenAI is already part of how allied health students study, and almost nobody taught them how to use it. Of 417 surveyed undergraduates, 76.3% (n = 318) reported no prior training in responsible AI use, and 40.8% said the guidance they got from academic staff was inadequate. Ago, Kwao and Quranchie integrated that survey, 205 coded open-ended responses, and interviews with 18 students and eight academic staff into four explanatory categories. Students who could prompt used the exchange as task design, setting the level, format and sequence of an explanation or a self-test. Students who could not prompt got fluent answers with little basis for auditing them, which is the study's central concept: a prompt-verification paradox, in which AI is most attractive exactly where the learner's prior knowledge is thinnest and their capacity to catch an error is lowest. Prompt engineering is the smaller half of what participants needed.
Key Findings
- Most students had never been taught responsible AI use. 76.3% of the 417 surveyed undergraduates (n = 318) reported no prior training, and guidance from academic staff was rated inadequate by 40.8% (51 strongly disagreed, 119 disagreed) against 28.1% who rated it appropriate (82 agreed, 35 strongly agreed), with 130 (31.2%) neutral.
- Routine use did not mean the ability to direct the system. Limited AI literacy and competence accounted for 26 of the open-ended codes, 17 of them describing low technical know-how that reduced prompting efficiency and output quality. One fourth-year radiography student: "Sometimes I wish I could give AI a very good prompt to get some good answers, but then I'm not good with that. I'm not able to use it properly."
- Prompting worked as dialogue when it worked at all. Students specified the level of explanation, the response format, the number and focus of practice questions, and when answers should be revealed, turning lecture slides into a self-test. One third-year medical laboratory science student asked for the answers to come at the end so the exercise stayed a test.
- The errors students reported are the ones verification exists to catch. Inaccurate or misleading responses drew 39 codes, information overload and inconsistent outputs 26, vague or overly complex responses 10, hallucinations 8, and fabricated or inaccurate references 6.
- The prompt-verification paradox surfaced in one student's account of their own risk. "In a case where I do not really have a lot of knowledge about the material I'm giving to AI, the possibility of me taking every single thing AI gives me is high, and that's dangerous because sometimes AI is wrong" (fourth-year, occupational therapy).
- Students asked for curriculum rather than prohibition. AI literacy and capacity building accounted for 185 codes (76.76%) of the suggested support strategies: classroom teaching and ethical AI education 55, workshops and professional development 60, curriculum integration 42, structured training programs 28. Subscription costs (9 codes) and internet access (3 codes) were named as material constraints on use.
Prompting where training had not reached
The setting is a public university in Ghana offering four-year undergraduate programs in Radiography, Physiotherapy, Occupational Therapy, Dietetics, Respiratory Therapy and Medical Laboratory Sciences, with the third and fourth years as the main clinical years. Students there are not reluctant users; they select different tools for brainstorming, writing, explanation and research, and treat it as ordinary academic infrastructure. Frequency of use and capability diverge: some students never learned to shape an instruction.
The first explanatory category, access without consistent prompting capability, is where that divergence sits. Students described difficulty with "the way to go in giving AI the correct prompt" and with "generating prompts that will help with better pooling of resources." Prompting was also inseparable from tool choice: one second-year physiotherapy student named the combined problem as "knowing the right AI tool and the appropriate time to use together with the right prompt."
The guidance gap had a source in the staff interviews: academic staff had entered AI-enabled teaching at very different rates. One dietetics lecturer described routine use for lecture material and slide development; a radiography lecturer said, "I have not encountered its use myself in teaching or research or anything." Students therefore could not assume consistent guidance, and prompting capability was uneven on both sides of the classroom.
Prompting as task design, and the limit of a good prompt
The second category, prompting as purposeful dialogue for personalized learning, describes that productive version. When students could direct the system, they adapted it: simplifying language, breaking down a difficult concept, or progressively raising the level of an explanation. A third-year respiratory therapy student described asking for explanations pitched at a child's or a high schooler's level and moving up as understanding grew, then noted the property that separates GenAI from a textbook: "It's quite fun because it's also interactive, so you can ask questions along the way." Another student used the same mechanism for retrieval practice, naming the content area, the number and type of questions, where to concentrate, and when to release the answers.
That is metacognitive work expressed as a prompt, and it is where prompt engineering stops short of prompt literacy: optimizing an input is not the same as framing a task, interpreting a response and deciding what to do next. The same students ran into its limit.
The third category, prompt quality, prior knowledge and verification as interdependent safeguards, is that limit. Better direction improved relevance without guaranteeing truth. Participants reported wrong answers, inconsistent outputs, non-existent references, and responses that argued convincingly for their own mistakes. Verification therefore mattered, but the capacity to verify depended on disciplinary knowledge, which produces the paradox: the tool is most useful for unfamiliar material and least checkable there. Some students verified against lecture slides, textbooks and scholarly articles. Others checked one AI tool with another, a practice that can reproduce an error rather than surface it.
A curricular and institutional responsibility
The fourth category, prompt literacy as a shared curricular and institutional responsibility, moves the problem off the student. Participants did not think rules would do the work, and their suggested strategies were overwhelmingly educational rather than disciplinary: teach how and when to use AI, how to get useful output without compromising authenticity, and how to evaluate and acknowledge assistance. One medical laboratory science student asked the university to stop framing AI as a rule violation and "incorporate it into our curriculum so that students are able to know how and when to use AI in academics." Training was requested for lecturers as well as students, and educators were positioned as people who should model prompting, disclosure, verification and the boundaries of use.
Read through the Capability-Opportunity-Motivation-Behavior model that frames the paper, students had motivation in abundance and opportunity only partly. Paid subscriptions and connectivity were uneven, and thin academic support increased reliance on the tool. The discussion adds a distinction worth carrying into course design: prompting that follows engagement with the material can augment cognition, while prompting that produces a finished answer before the student has engaged can substitute for the work the course exists to build — the upskilling inhibition the authors cite from the deskilling literature. The educational question is not whether a prompt was used but what intellectual work remained with the student.
How the study was done
This is a focused analysis within a larger explanatory sequential mixed-methods study of responsible AI use, reported against GRAMMS. The quantitative phase surveyed undergraduates by convenience sampling: 446 responses, 417 complete and retained. The present analysis uses one closed item on prior responsible-AI training and one open-ended question on challenges and academic integrity. Of 261 students who answered the open question, 53 reported no challenges and three did not address it, leaving 205 responses for conventional content analysis in Microsoft Excel; the broader analysis generated 200 codes. Descriptive statistics ran in IBM SPSS v32.0.
The qualitative phase drew on 86 students who volunteered contact details. Third- and fourth-year students were purposively selected for information-rich accounts and maximum variation across the six disciplines, yielding 18 student interviews; 11 academic staff were approached directly and eight were interviewed. Interviews ran on Microsoft Teams, recorded with consent and transcribed verbatim, and were analyzed in NVivo 15 using Braun and Clarke's six-phase reflexive thematic analysis. Integration was narrative: the first author compared convergence, complementarity and divergence across sources into a joint display, and the four categories are its output. The questionnaire was piloted with 39 recent graduates, and data collection ran from 20 May to 15 July 2026. Two authors were insiders — a radiographer trained at the setting and a final-year radiography student — and the paper says how that was handled.
What this means for practice
- Instructors. Teach verification against disciplinary sources — lecture material, textbooks, articles — and treat checking one chatbot with another as a known failure mode, not a strategy. Ask students to justify whether an output should be accepted, modified or rejected.
- Curriculum designers. Build prompt literacy progressively into clinically authentic tasks rather than a standalone module: students should frame the task, critique the response, verify the claim, and explain what reasoning stayed theirs. Assess that reasoning, not only the final product.
- Faculty developers. Train staff before assuming students receive consistent guidance; in this sample, lecturer familiarity ranged from routine course design to no personal use, and students read the inconsistency accurately.
- Administrators. Treat subscriptions and connectivity as part of the literacy environment. Nine codes named subscription cost and three named internet access, and unequal access to paid tools is an equity problem the curriculum cannot fix on its own.
Limitations
- Prompt literacy was not measured. No instrument assessed prompting performance, verification accuracy, cognitive substitution or learning outcomes; the construct was developed interpretively from a broader responsible-AI study, so the categories describe accounts rather than competence.
- The interviews drew on students who volunteered from the survey and were restricted to third and fourth years, so the accounts may not represent second-year students, who were 45.3% of the survey sample.
- One institution, convenience sampling and uneven disciplinary representation limit transferability; radiography supplied 44.6% of survey respondents and 10 of the 18 student interviews.
- The prompt-verification paradox is proposed from self-report and untested; the authors call for authentic prompting tasks and evidence on whether instruction improves verification.
- The design is cross-sectional, so nothing here establishes that training would change prompting or verification behavior.
Citation
Ago, J. L., Kwao, R. K., & Quranchie, D. (2026). Exploring prompt literacy in allied health professions education: an explanatory sequential mixed-methods study of student and educator perspectives. BMC Medical Education.