We conducted a web-based descriptive cross-sectional study in the form of an online survey targeting health professional students at UGA (France). The aim was to document actual use, perceived benefits and risks, as well as training needs related to GenAI in the context of clinical placements. Data collection took place between July 17 and September 30, 2025. This survey was carried out as part of a national initiative, funded by the French National Research Agency, to support the development and dissemination of digital health and AI in health education in French university health programs (France 2030 program, ANR n°23-CMAS-0035).

This study has received ethical approval from the “Ethics Committee for the Integrity and Ethics of Research in Health Professions Education” of the “Société Internationale Francophone d’Education Médicale” (Notice n°1905‐2025, issued July 17, 2025). Entry into the survey was conditional upon acceptance of an electronic informed consent form, accompanied by an information sheet detailing the study objectives, data confidentiality, voluntary nature of participation, right to withdraw, and potential risks. The survey was fully anonymous, and no incentives were offered for participation. Participation was voluntary, and informed consent was obtained electronically from all participants.

Following approval by the deans of the Faculties of Medicine and Pharmacy at UGA, an email invitation via the official academic mailing list was distributed to potential participants enrolled in medical or allied health training programs. Overall, approximately 4553 eligible students (n=2196 in medicine, n=283 in pharmacy, n=1904 in nursing, n=120 in midwifery, and n=50 in physiotherapy) were invited (see Multimedia Appendix 1 for details). In addition, a flyer was posted on student information boards at the entrance of lecture halls and classrooms throughout the recruitment period, ensuring a reinforced exposure to the survey. A total of 388 students completed the survey, corresponding to an overall response rate of 8.5%, ranging from 3.1% in medicine to 26% in physiotherapy. Interested students accessed the survey via a LimeSurvey link.

In the French health professions education system, clinical placements are mandatory components of curricula. Rotations typically last 2 to 10 weeks. Students progressively assume responsibilities according to their year of study, under the supervision of licensed health professionals from the relevant discipline, with day-to-day oversight and validation of key care and documentation tasks. Early-year students mainly observe and perform basic tasks, while senior students (in the final years of training) take a more active role in discipline-specific activities, including patient care tasks, documentation, information-seeking, care coordination, and patient communication or education.

We collected data through an anonymous online survey administered with LimeSurvey. Students interested in the study were invited to access the anonymous link, which contained 61 items: 18 single-choice questions, 38 single-choice questions on a 4-point Likert scale (1=strongly disagree to 4=strongly agree), 1 multiple-choice question, and 4 open-ended questions. The full survey questionnaire is available in Multimedia Appendix 1.

The items were developed based on 4 published studies: Blease et al [21] on generative GenAI in primary care, Lobet et al [23] on the spontaneous use of ChatGPT by university students, Karaca et al [24] on the Medical Artificial Intelligence Readiness Scale for Medical Students, and Boillat et al [25] on physician and student preparedness for AI. Items were then adapted by consensus to the clinical context.

We organized the instrument into 4 domains:

We conducted this web-based survey in accordance with the CHERRIES (Checklist for Reporting Results of Internet E-Surveys) checklist (Checklist 1) [26]. The questionnaire was hosted on LimeSurvey and included a single skip logic based on declared GenAI use versus nonuse; all other sections were shown to all participants. A total of 38 items were mandatory; missing responses were otherwise allowed, and respondents could review or change answers before submission. Only fully submitted questionnaires were analyzed, and no technical measures (cookies, IP checks, or registration) were used to prevent multiple entries.

Content validity was assessed through expert review by 2 clinicians and health-professions education specialists (SK and MM), who evaluated item relevance, clarity, and coverage of the 4 intended domains. Minor wording changes were made to improve clarity and contextual alignment with clinical placements (4 items required rewording for better understanding; no items were removed or added). A pilot test with 10 students (n=4 students in medicine, n=3 students in pharmacy, and n=3 students in nursing) confirmed intelligibility. Feedback from experts and students was incorporated to produce the definitive version. The average completion time was 15 minutes.

To capture students’ self-perceived maturity toward GenAI, considered as their perceived readiness to understand, evaluate, and appropriately use GenAI in educational and professional contexts, we constructed a composite GenAI maturity index from 9 Likert-type items [24]. This approach follows previous work on AI readiness and familiarity in health professions, where multiple items are aggregated into a single readiness or familiarity score [24,25]. One item assessed general digital comfort (“I feel comfortable with digital technologies in general”), and 8 items targeted GenAI-specific knowledge and self-efficacy (basic GenAI concepts, distinction between GenAI and other AI, ability to judge the reliability of GenAI outputs, awareness of appropriate use contexts and current limitations, and perceived ability to use GenAI in one’s education and future practice (full wording in Multimedia Appendix 1).

All 9 items used a 4-point agreement scale (1=strongly disagree, 4=strongly agree). To respect the ordinal nature of the 4-point Likert responses, we computed, for each respondent, the median of the 9 item scores and rounded it down to the nearest integer. This yielded a 4-level ordinal GenAI maturity score: 1 (minimal), 2 (limited), 3 (moderate), and 4 (high maturity). We required at least 5 nonmissing responses out of 9 (≥50%) to compute the index, consistent with commonly used “half-rule” approaches for scoring multi-item scales in the presence of item-level missingness [27].

In this sample, the 9 items showed good internal consistency (Cronbach α=0.869). Internal consistency was similar when excluding the general digital-comfort item (α=0.870 for items 2‐9), suggesting that digital comfort functions as a foundational component of the same underlying construct. Inspired by previous AI-readiness instruments that combine several facets into a single readiness score [24,25], we therefore use this composite GenAI maturity index pragmatically as an ordinal stratification of self-perceived GenAI maturity.

Open-ended responses were analyzed using a qualitative descriptive approach with descriptive content analysis. Responses were segmented into 213 meaning units (eg, items separated by semicolons or distinct ideas) and analyzed by question (reasons for nonuse, additional tasks, suggested uses, suggested risks, and training or support needs). A hybrid coding strategy was used, combining a first deductive framework aligned with these domains and inductive identification of recurrent subthemes. Two researchers (SK and MM) independently coded all segments using the agreed codebook. Initial agreement was 83.1% (177/213); disagreements (36/213, 16.9%) were resolved through discussion until consensus, and the final coding was used for thematic reporting and choice of illustrative quotations. Quotations were anonymized and translated from French to English. Qualitative findings are reported as recurring themes with illustrative excerpts to complement the quantitative results rather than as a stand-alone qualitative study.

Data were exported to Excel and analyzed in Python (version 3.11) using pandas (2.2.2), NumPy (1.26.4), SciPy (1.13.1), statsmodels (0.14.2), and matplotlib (3.9.2). Quantitative variables were described using means or medians with IQRs, while qualitative variables were presented as counts and percentages.

We performed overall group comparisons using Pearson chi-square test to assess the heterogeneity of distributions across academic disciplines and maturity levels. When expected cell counts were small, Fisher exact test was applied. For pairwise comparisons between a given discipline and all others, odds ratios with 95% CIs were calculated.

To assess gradients across maturity levels, 4-point Likert responses were dichotomized into 2 categories (disagree=1‐2; agree=3‐4) to facilitate interpretation and to enable proportion-based trend testing across ordered groups [28-30]. Linear trends were then assessed using the Cochran-Armitage trend test. When multiple simultaneous comparisons were conducted, the family-wise error rate was controlled using the Holm method. Statistical significance was set at P less than .05. We acknowledge that dichotomizing ordinal responses entails information loss compared with ordinal models [28].

The survey included 388 health professional students (Table 1), predominantly women (n=308, 79.4%), with most respondents aged 20‐24 years (n=198, 51.0%). The sample was largely composed of nursing students (n=217, 55.9%), with additional representation from medicine, pharmacy, midwifery, and physiotherapy programs (Table 1). Years of study covered the full curriculum, with a higher proportion of early-year students (y 1‐2) and smaller proportions in later years (Table 1).

Among the 388 respondents, 380 (97.9%) provided sufficient data to compute the GenAI maturity index. Most students reported limited to moderate maturity (levels 2‐3: 329/380, 86.6%), while relatively few reported minimal or high maturity (levels 1 and 4: 51/380, 13.4%; Table 1). Item-level details are provided in Multimedia Appendix 1.

Across 7 items assessing training and support needs, the median response was “Agree” for every statement (Figure 3). The strongest endorsement concerned a best-practice clinical guide (292/373, 78.3% agree or strongly agree) and ethics or regulatory awareness training (294/378, 77.7%). Important levels of agreement were also observed for profession-specific coaching and training in human-AI collaboration (Figure 3). Overall disagreement remained limited (strongly disagree ≤11% across items). Full response distributions, item-level medians, and IQRs are reported in Multimedia Appendix 1.

Free-text comments on additional needs (24 meaning units) provided further context. A large share reported no additional needs or interest (8/24, 33%). Among expressed needs, the most frequent theme concerned practical training and ongoing support to integrate GenAI appropriately into clinical learning and practice (6/24, 25%). The full thematic breakdown, including less frequent themes and illustrative quotations, is provided in Multimedia Appendix 1.

Seven items assessed students’ perceptions of organizational readiness and governance for GenAI use during clinical placements (Figure 4). Overall, respondents perceived limited institutional communication and safeguards: only a small minority recalled supervisors providing information on legal responsibilities or local conditions of use (Figure 4). Governance was also perceived as insufficient, with 36 of 359 (10%) agreeing that GenAI use is sufficiently regulated. Operational readiness received the poorest ratings, as only 18 of 372 (4.8%) agreed that staff are trained for clinical use and 56 of 343 (16.3%) agreed that tool biases are accounted for. In contrast, transparency toward patients appeared as the single area of strong consensus: 248 of 331 (74.9%) agreed that patients should be informed when AI is used in their care (Figure 4). Full response distributions, item-level medians, and IQRs are reported in Multimedia Appendix 1.

In this cross-sectional survey of 388 French students in health professions, 52.6% (204/388) reported using GenAI during clinical placements. Adoption varied across disciplines, with lower uptake in midwifery, and increased markedly with higher self-perceived GenAI maturity. Reported uses primarily involved information retrieval and documentation- or writing-related tasks, while patient-facing activities were less frequently reported.

Notably, close to a quarter of GenAI users reported at least 1 instance of entering patient-identifying information, and close to half reported processing real medical content perceived as anonymized. Students perceived substantial benefits for documentation support but also expressed prominent concerns about dependency, erosion of clinical skills, and breaches of confidentiality, alongside strong demand for ethics or regulation training and best-practice guidance. Overall, these findings show that GenAI is already embedded in students’ placement practices, but with uneven adoption, limited institutional framing, and persistent confidentiality risks.

In this survey, students appeared to mobilize GenAI during placements primarily as a cognitive and organizational aid (eg, information retrieval and documentation support) rather than as a tool embedded in direct patient-facing work. This “low-stakes versus clinically situated” pattern plausibly reflects implicit boundaries around what feels safe, acceptable, and professionally legitimate to delegate to a third-party model in real clinical environments and aligns with students’ expressed need for clear best-practice guidance.

Perceived readiness offers an additional lens to interpret why some students integrate GenAI into placement work while others do not. Our maturity index should be understood as a pragmatic stratification of how prepared students feel to evaluate and use GenAI critically in clinical learning contexts, rather than as an objective measure of competence. The strong adoption gradient across readiness levels is consistent with AI-readiness frameworks in health professions education [24,25]. Emerging evidence also suggests that structured GenAI-supported simulations can improve targeted competencies in medical students [31,32]. Educationally, this matters because uneven readiness may translate into uneven exposure to GenAI practices during placements, potentially reinforcing a hidden curriculum in which norms of safe and legitimate use are acquired informally and shape professional identity formation at a formative stage [33,34].

This raises a broader tension between efficiency and reflection in workplace learning. While GenAI may support information management and autonomous learning [35], medical documentation and reasoning processes are also critical spaces for cultivating clinical rigor, accountability, and reflective judgment. Without explicit pedagogical framing, early delegation of reasoning-related tasks may narrow opportunities for reflexivity and increase vulnerability to automation bias and overreliance. These mechanisms were previously described in clinical decision support contexts, where they can compromise patient safety [36]. Conversely, when embedded in structured educational approaches (eg, simulation with feedback and validated rubrics), GenAI can be leveraged to enrich reflection and support deliberate practice, as illustrated by initiatives such as MedSimAI [32]. Collectively, these findings support the need to move from informal, individual experimentation toward supervised and pedagogically grounded uses of GenAI in clinical learning environments [34,36].

Students’ reasons for not using GenAI during placements suggested that nonadoption often reflected deliberate reservations rather than simple unfamiliarity. Reported concerns included limited perceived clinical relevance, fear of diminished reflexivity, confidentiality risks, and broader socio-technical critiques, such as environmental impact. This aligns with Tortella et al [6], who observed that reluctance among physiotherapy students was largely driven by concerns about inaccuracy and error, and perceived limits for complex tasks, rather than a lack of awareness. From a teaching perspective, such reservations should be treated as a pedagogical resource: they create opportunities to make expectations explicit and to cultivate critical judgment, irrespective of whether students ultimately adopt GenAI tools.

Perceived benefits were large but not uniform. Students strongly endorsed GenAI for documentation relief and improved access to information, positioning it primarily as a cognitive and organizational facilitator. In contrast, endorsement was more cautious for higher-stakes functions such as diagnostic support, personalized planning, and efficiency gains, reflecting conditional trust rather than full reliance. This pattern is consistent with Janumpally et al [37], who emphasized that GenAI may support learning and selected reasoning tasks but remains constrained by output reliability and cannot substitute for human judgment, particularly in domains directly implicating professional responsibility, such as patient data collection and prognostication.

Perceived risks were dominated by concerns about dependency, long-term erosion of clinical skills, and confidentiality breaches, alongside worries about bias, uncertain legal accountability, and a more subtle displacement of professional responsibility and human interaction. These concerns resonate with the broader literature on automation bias and overreliance in clinical decision support, including discussions of “assistive AI” as the lowest level of autonomy that may still erode vigilance if not framed and supervised [36]. Importantly, students did not primarily express fears of being replaced. Rather, they emphasized risks of weakened judgment, diluted accountability, and reduced quality of the care relationship. These concerns suggest that GenAI integration in workplace learning should be structured, reflective, and explicitly governed rather than left to informal experimentation [6,36,37].

From a patient-safety and governance perspective, the disclosure of sensitive information calls for specific attention. In our sample, 23.5% (48/204) reported at least 1 instance of entering patient-identifying information, and 47.1% (96/204) reported processing real medical content perceived as anonymized. Even when students perceive data as anonymized, free-text prompts may contain contextual details that increase reidentification risk, and the use of nonapproved third-party tools may involve data processing outside institutional control (eg, storage and transfer conditions) [38,39]. In European contexts, this directly intersects with General Data Protection Regulation obligations and institutional requirements for handling health data [40].

These findings imply a need for explicit supervision and coordinated governance during placements. Placement supervisors, faculty, and host clinical sites (eg, hospital departments and private practices) should clearly define acceptable versus unacceptable prompting, provide concrete deidentification rules and examples, and ensure access to approved, compliant tools (or clear prohibitions where such tools are not available). Because clinical training occurs across multiple host sites, governance should be coordinated between the academic institution and placement providers so that expectations and safeguards are consistent across settings. At the institutional level, local charters and “allowed tool” policies aligned with data-protection governance should go with training, so that confidentiality norms are not left to informal experimentation. In France, this approach is consistent with recent institutional guidance: the French National Academy of Medicine discusses ethical issues and recommendations for the use of GenAI in health [41], and the French National Authority for Health has published a practical guide to support appropriate and informed use of GenAI by health professionals [42].

Students’ expectations for training extend beyond technical mastery toward the full triad of knowledge, skills, and professional attitudes: knowledge of ethical and regulatory frameworks; practical competencies developed through workshops and profession-specific coaching; and professional dispositions cultivated through mentorship, reflexivity, and attention to sustainability. This aligns with calls to design health AI curricula that move past technical literacy to incorporate critical thinking, ethics, and patient-centered judgment. Komasawa and Yokohira [43] argue that the rise of GenAI requires rethinking health professionalism by combining technological fluency with enduring humanistic values such as empathy, integrity, and accountability.

Our findings point to an institutional readiness gap during placements: limited shared benchmarks, uneven supervisory guidance, and insufficient attention to bias, regulation, and data governance. They also support a stepwise pathway to move from informal, student-driven use toward safe and educationally meaningful integration across disciplines. First, a transversal common core should address ethics, regulation, professional standards, and model limitations or uncertainty, alongside concrete rules for health data handling in GenAI contexts (what must never be entered, how to deidentify, and which tools are approved). Second, competency development requires practice beyond theory: simulation- or case-based activities with GenAI, explicit reflection prompts, and structured feedback can help students learn when GenAI supports reasoning versus when it undermines it (eg, automation bias and overreliance) and can contribute to professional identity formation. Third, training for both academic staff and host-site supervisors is essential so that placement supervisors can set shared expectations, mentor reflective use, and intervene when unsafe practices occur. Fourth, institutional governance should provide clear local policies (eg, charters, approved tool lists, patient transparency where relevant, and escalation procedures for data incidents), ensuring alignment between pedagogy and professional standards. This enabling framework does not mandate adoption; it operationalizes conditions under which GenAI use remains compatible with reflexivity, professional accountability, and patient-centered care, and can help transform spontaneous use into integrated, supervised competencies in clinical learning settings, as illustrated by structured simulation initiatives such as MedSimAI [32].

This study has several limitations. First, it was conducted within a single French university (UGA), which limits international generalizability. The survey was embedded in a national initiative funded by the French National Research Agency to support the rollout of digital health and AI in health education across French university training programs. While this may strengthen the transferability of educational implications, it does not eliminate the single-site constraint. Notably, the main trends observed here are consistent with the international survey by Busch et al [44] (4596 students from 192 faculties across 48 countries), which similarly reported widespread student adoption of generative tools alongside a global deficit in formal AI education. Second, GenAI use and data-disclosure behaviors were measured through self-report and may be influenced by recall and social desirability biases. Although responses were collected anonymously and without incentives, self-reporting remains susceptible to these biases. Nevertheless, documenting students’ lived practices and perceptions was a core aim, particularly in contexts where formal pedagogical frameworks are still emerging; as such, perceptions may be viewed as outcomes in their own right, consistent with Busch et al [44]. Third, we did not collect data from clinical supervisors or educational leaders and therefore cannot characterize institutional policies or supervisory practices with certainty. This likely reflects a transitional situation in which both learners and supervisors are adapting to rapidly evolving GenAI tools and local guidance, as suggested by recent work by Blanco et al [45] and McCoy et al [46]. Fourth, the cross-sectional design provides a snapshot at a specific time point; given the rapid evolution of GenAI tools and institutional policies, the stability of these findings over time cannot be assumed. Fifth, interprofessional heterogeneity is an additional limitation: some health professions were underrepresented, limiting robust profession-specific comparisons. This reflects our primary aim of mapping placement-related practices rather than comparing disciplines; nonetheless, profession-specific differences in AI literacy and attitudes have been reported elsewhere and warrant dedicated analyses [47,48]. Finally, our GenAI maturity score should be considered an exploratory composite index of self-perceived readiness. Although internal consistency was good (Cronbach α=0.869), we did not undertake a full psychometric validation; further work should confirm and refine this construct.

Future research should move beyond single institutions to include multicenter and international comparisons, enabling the identification of contextual variations and common challenges. Longitudinal designs will also be needed to capture how perceptions, competencies, and professional identity evolve as GenAI becomes more embedded in health professions education. Finally, triangulating student perspectives with those of supervisors, educators, and institutional leaders will be crucial to developing evidence-based strategies that responsibly integrate GenAI into curricula while respecting profession-specific needs.

In conclusion, GenAI is already being used by French students in health professions during clinical placements, primarily as a cognitive and organizational aid. At the same time, students report concerns about dependence, skill erosion, and confidentiality, and express a strong demand for structured guidance. These findings support the need for explicit, ethically grounded, and supervised frameworks that align curricula, clinical supervision, and host-site governance so that GenAI use fosters reflexivity, autonomy, and professional accountability rather than unsupervised experimentation. More broadly, this echoes developments already observed across higher education and health care governance, including emerging institutional policies [13], the timely need for clear ethical and regulatory frameworks [49], and methodological approaches for developing evidence-based interdisciplinary guidelines [50].