Applications, attitudes and ethical considerations of Generative Artificial Intelligence (Gen AI) in nursing education: a scoping review

Synthesis: Hardie and colleagues map how generative AI is used, received, and governed in nursing education, drawing on 103 articles published between January 2014 and July 2025. The corpus is split almost evenly between evidence and argument: 59 empirical studies and 44 discussion, opinion, or conference papers. Four application clusters organize the findings, namely content creation, Simulation, personalized learning and tutoring, and Assessment. Attitudes are broadly positive but conditional on trust, training, and task fit, and the ethical concerns that recur across the corpus are Privacy, transparency, bias, academic integrity, and accountability. The review is a map rather than a test of effect, and it says so: nearly half of what it counts is commentary, and no included study was appraised for quality.

Key Findings

1. Searches of five databases returned 1,251 records, and a gray-literature search across six platforms returned 142 more. After duplicate removal and screening, 103 articles were included, spanning 31 countries, with the United States contributing the largest share at 42.
2. Of the 103 included items, 59 are empirical studies and 44 are discussion, opinion, or conference papers. The empirical designs range from randomized controlled trials and quasi-experiments to cross-sectional surveys and small qualitative studies.
3. Gen AI is used for four things: content creation and enhancement, simulation, personalized learning and tutoring, and assessment. Case-based learning appears within the simulation theme in individual studies but is not one of the review's four named categories.
4. Performance findings are mostly favorable but uneven. Several studies report gains in knowledge, case management, confidence, and critical reflection, while others find no difference in knowledge, clinical reasoning, or satisfaction.
5. Accuracy is adequate for some tasks and unreliable for others. ChatGPT-4 answered 88.67% of NCLEX-RN questions correctly in one cited study, but ChatGPT failed to produce a pass-grade response on written assessment questions in a nursing informatics program while passing three of four data science assignments.
6. Students and educators report mixed attitudes. Adoption tracks perceived usefulness and ease of use, and higher AI literacy is associated with greater acceptance, but plagiarism concerns, shallow outputs, and the need for verification temper enthusiasm.
7. Ethical concern is close to universal in the commentary. All but three of the 44 discussion papers raise ethical issues, most often privacy and data security, transparency, bias and equity, academic integrity, and human oversight.

Four uses, unevenly evidenced

Content creation covers a wide range of tasks. Studies report DALL-E3 generating images for teaching materials, ChatGPT summarizing lecture slides and research papers, chatbots drafting patient education leaflets, and models producing NCLEX-style questions and care plan exercises.

Simulation is a heavily studied use. Large language models generate realistic case studies and interactive patient-nurse dialogues, and AI-powered virtual reality simulations supply patient interactions that adapt to a learner's responses. Debriefing is a smaller but distinct strand: one study used a ChatGPT-based chatbot to support individual debriefing after simulation and reported better engagement and critical reflection than conventional debriefing, though clinical understanding was similar across groups. Case creation sits inside this cluster: Higashitsuji and colleagues asked faculty and students to compare manually written cases with ChatGPT-generated cases in a case-based learning exercise and found creation time fell from 106 to 71 minutes, with no significant difference in perceived burden or discussion quality.

Personalized learning and tutoring cover AI teaching assistant systems, conversational chatbots that answer clinical questions, and tools that tailor revision or feedback to individual students. Assessment is the most contested category. It includes automated question generation, automated essay scoring, and AI-supported writing assessment, and it is where the integrity questions concentrate.

What the performance studies can and cannot show

Several studies report gains. Akutay and colleagues randomized 188 third-year nursing students and found higher case management and knowledge scores in the group taught with AI-generated visual narratives, with no difference in satisfaction. Benfatah and colleagues studied 40 students and reported that AI-assisted debriefing improved confidence, engagement, and critical reflection more than conventional debriefing, although asthma management knowledge was similar. Gonzalez-Garcia and colleagues surveyed 98 students and found self-reported grade improvements associated with ChatGPT use.

Null and negative results matter just as much. Han and colleagues compared 61 students receiving video plus chatbot instruction with students receiving video alone and found no significant differences in knowledge, clinical reasoning, confidence, feedback, or satisfaction, though the chatbot group showed more interest and self-directed learning. Reeder and Lee found that ChatGPT could not produce a passing response on written assessment questions in a nursing informatics program, while it passed three of four assignments from a data science program. Kapadia and colleagues found GPT-3.5 performed well in health history scenarios but portrayed doctors poorly in bedside conversation simulations.

The review's own summary is that the evidence is promising but limited. Samples are small, most designs are quasi-experimental or cross-sectional, outcomes are often self-reported perceptions rather than measured competence, and few randomized trials exist. The review identified no robust longitudinal studies and no current postgraduate research projects on the topic. It also notes that the empirical literature concentrates heavily on ChatGPT, leaving Claude, Gemini, DeepSeek, and Copilot largely unexamined.

Attitudes split along a trust line

The dominant attitude finding is conditional acceptance rather than enthusiasm or resistance. In a survey of 176 nursing students, nurses, and educators in northern Italy, between 50% and 85% of respondents viewed AI positively for personalized training, cost savings, and inclusivity, while remaining skeptical about replacing staff or educators. Alenazi found that perceived usefulness and ease of use significantly predicted willingness to adopt. Hashish and Alnajjar and Sumengen and colleagues found that greater AI literacy correlated with greater acceptance and readiness.

Trust is the hinge. In a qualitative study of 19 students, Hawk and colleagues reported a shift from initial skepticism to cautious optimism, with students describing chatbot answers as helpful but shallow and stressing the need to validate outputs and watch for fabricated sources. Lebo and Brown found that students valued AI patient cases for practicing dialogues and medication review, but some struggled when they asked complex questions. Among educators, Bouriami and colleagues found general acceptance of ChatGPT alongside persistent plagiarism concerns, and, tellingly, that educators who did not regard plagiarism as a serious risk were more likely to use the tool. Reeder and Lee's finding that output quality varies by subject underscores the same point: the attitude follows the task.

Ethics as design constraints, not add-ons

Privacy and data security come first in the corpus. The recurring instruction is to avoid entering confidential student or patient data into prompts and to make data collection and use transparent. Bias is the second cluster. Several authors warn that models trained on incomplete or skewed data can produce educational inequities, and Lim specifically warns of linguistic bias, noting that students with English as an additional language may be more likely to be suspected of AI use. Access is part of the same problem: paid tiers produce different outputs from free ones, and inequitable access can reproduce existing advantage.

Academic integrity dominates the assessment discussion. Text-based take-home responses are the most exposed format, and the recommended response is not detection alone but redesign: competency-based clinical skills, performance tasks, problem-based learning, and more oral assessment through presentations, debates, and mini-vivas. Disclosure is the common thread, including the five-level Gen AI Assessment Scale, which runs from no AI use to full collaboration and specifies citation and disclosure requirements at each level.

Accountability and human oversight close the set. Hallucinations and fabricated references require verification, and several authors argue that overreliance risks weakening the empathy and interpersonal skills central to nursing. The policy response described in the review is institutional: clear guidelines, staff and student training, transparent rather than purely punitive rules, and AI literacy embedded in the curriculum.

What this means for practice

• Instructors. Treat model output as a draft that requires verification, and design activities in which students compare, check, and correct AI-generated clinical content rather than simply consume it.
• Assessment designers. Shift weight from take-home text toward performance, oral, and in-person components, and publish explicit disclosure rules with the level of permitted AI use named for each task.
• Administrators. Adopt a policy that specifies what student and patient data may never enter a prompt, who is accountable for AI-supported decisions, and how incidents are reviewed, rather than relying on misconduct penalties alone.
• Faculty developers. Build training around task fit and verification, since adoption tracks perceived usefulness and ease of use, and target the subject areas where output quality is weakest.
• Researchers. Prioritize randomized and longitudinal designs, measure competence rather than satisfaction, and test tools other than ChatGPT before drawing general conclusions about generative AI in nursing education.

Limitations

• The review is a scoping review and did not appraise the quality of included studies, so it maps the field rather than grading its evidence.
• Forty-four of the 103 included items are discussion, opinion, or conference papers, so nearly half the corpus reports proposals and perceptions rather than measured outcomes.
• The empirical studies rely heavily on surveys and self-report, with small samples and few randomized trials, which limits causal inference and generalizability.
• Only English-language publications were included, and studies of Gen AI outside nursing education were excluded, so relevant work may be missing and publication bias is possible.
• The search closed on July 1, 2025, in a fast-moving field, and the empirical literature is concentrated on ChatGPT, so newer tools and recent studies are underrepresented.

Citation

Hardie, P., Darley, A., Derwin, R., Eustace-Cook, J., Kearns, S., Mc Brien, B., Siddiquee, A., Zheng, D., & Mooney, M. (2026). Applications, attitudes and ethical considerations of Generative Artificial Intelligence (Gen AI) in nursing education: a scoping review. BMC Nursing, 25, 148.