We conducted a national web-based cross-sectional survey of health care professionals and students in France. The survey was disseminated through PulseLife, a digital platform and professional community used by health care workers in France. Recruitment was conducted via email between December 4, 2024, and March 5, 2025.

No financial incentive was offered. The exact number of invited individuals, page view rates, participation rates, and completion rates was not retained by the platform in a form that allowed for retrospective reconstruction; accordingly, a formal response rate could not be calculated. This study should therefore be interpreted as a convenience-sample survey rather than a probability-based national survey.

Eligible participants were health care professionals directly involved in care delivery and health profession students training in France. The survey was accessed through authenticated individual PulseLife accounts. Each account could submit the questionnaire only once, and multiple submissions from the same account were not allowed.

The survey instrument included demographic and professional characteristic items together with 12 substantive closed-ended questions and was administered in French. Questions were not technically mandatory, and respondents could skip items. The archived platform configuration did not retain sufficiently detailed information to reconstruct whether back navigation to earlier items was enabled. Incomplete questionnaires were retained in the analytic dataset; all percentages were therefore calculated using the number of respondents to each item as the denominator unless otherwise specified. For multi-selection items, totals could exceed 100%.

The primary objective was to assess use, concerns, and perspectives regarding AI technologies among French health care professionals and students. Secondary objectives were to explore associations between respondent characteristics and AI use or confidence and identify training needs with direct relevance to medical and health professions education.

The questionnaire covered respondent characteristics and AI use, confidence, perceived benefits, perceived concerns, and training interest (Multimedia Appendix 1). In this survey, “AI-based solutions” was used as an umbrella term encompassing, for example, decision support systems, imaging analysis tools, administrative automation, and generative AI–enabled information tools. The questionnaire was developed through iterative multidisciplinary review by clinicians involved in AI-related clinical practice and digital health implementation. The development process aimed to ensure face validity and practical relevance for health care professionals and students in France. Before dissemination, the questionnaire was pilot-tested with 20 health care professionals and students to assess clarity, wording, and feasibility. This pilot stage was intended as a face validity and usability check rather than formal stand-alone validation.

Because the instrument was created for this study, we report its psychometric properties as exploratory support rather than definitive external validation. Internal consistency across domains was acceptable (Cronbach α=0.76-0.84). Sampling adequacy was supported by a Kaiser-Meyer-Olkin index of 0.83, and the Bartlett test of sphericity was significant (P<.001). Exploratory factor analysis supported a 3-factor structure (concerns, perceived benefits, and assisted tasks), a Confirmatory factor analysis of the 14-item, 3-factor model indicated acceptable fit (χ²/df=2.1; df=74; comparative fit index=0.94; Tucker-Lewis index=0.92; root mean square error of approximation=0.05; standardized root mean square residual=0.06). Known group validity was supported by higher AI confidence among physicians, younger respondents, and participants with prior AI training. More detailed factor analytic parameters, including extraction method, rotation strategy, factor retention criteria, item retention thresholds, and confirmatory factor analysis estimator specifications, were not retained in a form that allowed for full retrospective reporting. Similarly, domain scores were interpreted at the construct level, but a formal scoring algorithm for external reuse was not predefined because the instrument was designed for this exploratory survey. Because exploratory and confirmatory analyses were conducted on the same overall dataset, these findings should be interpreted cautiously and are further discussed as a limitation. Additional psychometric details are provided in Multimedia Appendix 1.

Descriptive statistics are reported as counts and percentages. Associations between categorical variables were explored using Pearson chi-square tests. No formal correction for multiple testing was applied because the inferential analyses were exploratory. This approach increases the risk of type I error, and reported associations should therefore be interpreted as hypothesis generating. Effect sizes such as the Cramér V were not available from the archived analytical outputs in a form that allowed for reliable retrospective reporting across all comparisons. For this reason, we retained a conservative presentation of the inferential analyses and limited their interpretation accordingly. Analyses were conducted using R (version 4.3.1; R Foundation for Statistical Computing). No weighting was applied because sampling probabilities and suitable calibration margins were unavailable.

This study was conducted in accordance with the Declaration of Helsinki and the French legal framework applicable to health research and personal data protection. The protocol was reviewed and approved by the local ethics committee of Hôpital Européen Georges Pompidou, Paris, France; however, no formal approval number was assigned. Participation was voluntary, and completion of the online questionnaire was considered to constitute informed consent. The dataset used was deidentified before analysis. No directly identifying information was collected in the study dataset used by the academic investigators. Participants received no financial compensation.

This national survey found a marked mismatch between the rapid diffusion of AI discourse in health care and the limited formal preparation reported by French health professionals and students. Formal AI-specific training was uncommon, whereas interest in receiving such training was high. Confidence in AI for diagnosis and patient management support remained limited, and current use in professional activity was heterogeneous rather than routine.

When AI was used, it was more often used for supportive or informational purposes than for core diagnostic or therapeutic decision-making. Respondents expected AI to improve diagnosis, save time, reduce errors, and support follow-up, but they also expressed concerns about algorithmic bias, transparency, confidentiality, and the patient–health care professional relationship. These findings suggest that respondents were not opposed to AI in principle but, rather, were conditionally receptive provided that AI remained trustworthy, interpretable, and compatible with safe, human-centered care.

Earlier studies cited above have already shown that interest in AI frequently exceeds formal preparation across different health care and educational settings [3-6]. The more recent literature further reinforces this pattern and provides additional points of comparison with our French cross-professional sample. Rjoop et al [11] surveyed 394 medical students and pathology trainees in Jordan and found that 66.2% disagreed that their medical schools had educated them about AI, whereas 46.2% expressed interest in learning about AI in medicine; the same study also showed substantial agreement with pathology-specific AI applications, illustrating how perceptions may be shaped by specialty exposure. Syeda et al [12] surveyed 939 undergraduate health care students in Pakistan and reported that only 11.8% had received formal AI training, whereas 78.8% supported AI integration into curricula and 82.2% endorsed AI training as part of medical education. A large international survey of 4596 medical, dental, and veterinary students from 192 faculties in 48 countries similarly found limited AI knowledge, lack of AI courses, and strong desire for more AI teaching [13]. These results are closely aligned with the French pattern observed in our survey, particularly the coexistence of limited formal training and strong educational demand. French data are more limited but point in the same direction. Perrier et al [14] conducted a nationwide survey of 165 young French pediatricians and found favorable attitudes toward AI, but only 5% had received specific AI training, whereas 87% considered AI training necessary. Kotzki et al [15] recently surveyed 388 French health professional students during clinical placements and found that more than half reported using generative AI, mainly for information and documentation support rather than patient-facing activities. Earlier qualitative work by Laï et al [16] based on 40 French stakeholders also suggested that French health professionals were not simply resistant to AI but focused on safe, clinically useful, and responsible implementation. Compared with these French studies, our survey adds a broader cross-professional perspective by including both practicing health professionals and students and assessing use, confidence, perceived benefits, concerns, and training needs within the same national dataset. The broader health service literature also supports the need for this transversal perspective. A 2025 systematic review of 72 empirical studies of health care professionals’ perspectives on AI identified multiple facilitating and hindering factors across individual, interpersonal, institutional, community, and policy levels [17]. The review also noted that many studies focused on radiology or general medicine, limiting the ability to generalize to other fields [17]. Our study therefore complements rather than duplicates the existing literature: it provides a broader national view that is less likely to be driven by the technological culture or implementation history of a single specialty. Specialty-specific European studies further illustrate this issue: the 2024 European Network for the Assessment of Imaging in Medicine and European Society of Medical Imaging Informatics survey found that nearly half of the radiology respondents were already using AI tools, and radiotherapy implementation work has evaluated commercial AI-based auto-contouring systems in clinical practice [18,19].

The clearest educational signal emerging from our survey was the marked gap between the low proportion of respondents who had received formal AI-specific training (90/1366, 6.6%) and the high proportion who wished to receive such training (920/1175, 78.3%). This finding is consistent with the broader international mismatch between the rapid diffusion of AI and insufficient educational preparation. Prior AI-specific training was rare in our study, but interest in training was high. This training gap is highly relevant to medical and health professions education because confidence was greater among those who reported prior AI training. These findings support the integration of structured AI-related competencies across undergraduate, postgraduate, and continuing professional education [7,20,21].

Recent curriculum work suggests that AI training should not be limited to isolated technical exposure. Tolentino et al [7] identified only 2 explicit curriculum frameworks among 21 included papers in their scoping review. Moldt et al [9] further showed that stakeholders expect AI education to include technical knowledge, ethical reasoning, implementation awareness, and clinically meaningful use. Moldt et al [9] recently proposed a postgraduate family medicine AI framework, reinforcing the need for longitudinal, competency-based approaches to AI training [10]. In practical terms, our findings suggest that training should address not only how to use AI tools but also why and when they should be used, how to calibrate trust, and how to recognize situations in which AI may be misleading or clinically unsafe [22,23].

The concerns reported by respondents reinforce the importance of framing AI education within broader ethical, legal, and organizational contexts. Concerns about bias, transparency, confidentiality, and the patient–health care professional relationship were prominent in our sample. These concerns are concordant with wider literature showing that bias, explainability, privacy, and implementation barriers are central determinants of trust in health AI [17,24-29].

The organizational implications are also important. Respondents expected AI to support efficiency and safety, yet some also perceived it as a potential threat to professional roles or skills. This tension is understandable because the introduction of new technologies may initially disrupt workflows, increase cognitive load, and create additional work during transition phases, particularly when systems are poorly integrated [26,30,31]. At the same time, AI has been described as having the potential to alleviate part of the administrative and cognitive burden that contributes to professional exhaustion and burnout, thereby helping preserve time for complex decision-making and human interaction [30,32,33]. Our data therefore support a balanced interpretation: clinicians and students expect AI to be helpful but only if it remains interpretable, accountable, and compatible with human-centered care.

This study has several strengths. With 1625 respondents, including 1212 (74.6%) practicing health professionals and 413 (25.4%) students, it provides a large French dataset for assessing AI use, concerns, confidence, and training needs across both groups. To our knowledge, it is among the broadest French cross-professional surveys to date on this topic. The questionnaire underwent pilot-testing and exploratory psychometric assessment. A key strength of the present study is its broad and cross-professional sampling frame. Unlike studies focused on a single AI-exposed specialty or one educational subgroup, our survey was designed to capture a wider range of professional and educational backgrounds. This is important because perceptions of AI may be shaped by prior exposure; specialty-specific workflows; and the technological culture of fields such as radiology, radiotherapy, pathology, or other highly digitized disciplines [11,17-19]. By including both practicing health professionals and students, the present study provides a broader national perspective on AI use, concerns, confidence, and training needs in France. This makes the findings particularly relevant for general AI literacy, curriculum design, postgraduate training, and continuing professional development.

Several limitations deserve emphasis. First, this study relied on a convenience sample recruited through a digital professional platform, which may overrepresent respondents already interested in technology. As a result, the findings should not be interpreted as nationally representative. Second, because the exact number of invited platform users was unavailable, a formal response rate could not be calculated. No weighting strategy could be implemented for the same reason. Third, item-level missingness varied across questions; we addressed this by reporting item-specific denominators, but the resulting estimates remain subject to nonresponse bias. Fourth, no formal correction for multiple comparisons was applied because the inferential analyses were exploratory; this increases the risk of type I error and reinforces the need to interpret the reported associations cautiously. Fifth, the psychometric analyses should be interpreted cautiously because exploratory and confirmatory factor analyses were performed on the same overall dataset. Finally, the inferential analyses were unadjusted and exploratory; this study was not designed to support causal interpretation.

In this national French convenience sample, formal AI training was uncommon despite strong interest in receiving it, and confidence in AI for diagnostic support remained limited. These findings support the relevance of structured AI education across undergraduate, postgraduate, and continuing professional training. They also highlight the importance of integrating ethical, legal, and practical implementation issues into such teaching.

Given the descriptive design and sampling approach, the results should be interpreted as exploratory rather than nationally representative. Future work should include longitudinal and mixed methods studies, particularly studies evaluating how training interventions influence AI literacy, trust calibration, and real-world adoption in clinical environments.