To be included, participants had to be medical students aged 18 years or older, enrolled in a French medical school, and in their fourth to sixth academic years of study. These students are in the second cycle of medical studies, which corresponds to the clerkship year. Their training is divided between university teaching and clinical rotations in university hospitals, peripheral hospitals, and general practice offices.

Exclusion criteria included refusal to participate, incomplete questionnaire responses, and duplicate email addresses to prevent redundancy.

The recruitment process involved several steps to obtain a representative sample of medical students in France.

First, an email containing an informational letter was sent to all medical schools in France. Then, dissemination was conducted through the social networks of the student association.

With approximately 25,000 medical students in 35 faculties hosting clinical students, our goal was to achieve a 10% participation rate, or 2500 students.

To motivate students to respond to the questionnaire, we offered students who completed the questionnaire the opportunity to participate in a contest to win one of five €20 (US $22.86) Amazon or Fnac gift cards.

We used a published survey to assess the perceptions and attitudes of medical students regarding AI [21].

The questionnaire consisted of 3 sections (Multimedia Appendix 1): the first section asked about the participants' personal characteristics. The second section was an SAQ where participants were asked to describe what AI is. Finally, the last section consisted of several questions with Likert scale responses ranging from 1 to 10 about their attitude toward AI, as well as an SAQ about their feelings toward AI.

The original survey was designed for all Canadian health students. We moved the first section to the last to better meet the needs and interests of medical students (Multimedia Appendix 1). To adapt the questionnaire to our target population, we modified the institution question to ask about their medical school in France, their academic year, their self-assessment of their class ranking, and their career aspirations.

The survey was tested by 6 medical students to verify comprehension, identify errors, and assess completion time. The medical students, organized into 3 pairs, were informed of the study’s objectives and were encouraged to verbalize any confusion, identify errors, or request clarifications at any point during the process. During the pilot test, the first pair of participants highlighted 2 issues: a duplicated question, which was subsequently removed, and a multiple-choice question that was initially designed to allow only 1 response despite requiring multiple selections, which was corrected to permit multiple answers. After these revisions were made, the remaining 2 pairs completed the survey without encountering any further errors or comprehension difficulties. Importantly, none of the 6 participants reported any confusion or requested additional explanations, indicating that the questions were clear and well understood. The average time required to complete the survey was observed to be 5 minutes, confirming that the questionnaire was both concise and comprehensive for the intended audience.

Continuous and score variables were described using medians and IQRs, whereas categorical variables were summarized as counts and percentages. Questions allowing multiple responses were analyzed as separate binary variables for each response option; therefore, percentages may exceed 100%.

The primary grouping variable was the final AI definition category, obtained after adjudicating the free-text definitions according to the Organisation for Economic Co-operation and Development (OECD)–based coding framework: An AI system is a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment [34]. Interrater reliability between the 2 independent coders was assessed using raw agreement and weighted Cohen κ. Because AI definition categories were ordinal (unknown, incorrect, partially correct, and correct), weighted κ coefficients were calculated using both linear and quadratic weights. Bootstrap resampling was used to estimate 95% CIs for weighted κ.

Comparisons across final AI definition categories were performed using the Kruskal-Wallis test for continuous or score variables and the chi-square test for categorical variables. Effect sizes were reported for all comparisons: epsilon-squared for Kruskal-Wallis tests and Cramér V for chi-square tests. For variables with a significant Kruskal-Wallis test, pairwise post hoc Dunn tests were performed with Holm correction for multiple comparisons.

The general attitude toward AI score (0‐10) was analyzed as the main quantitative outcome. Six additional AI attitude items were analyzed separately as secondary outcomes. To limit multiplicity, P values from the 6 omnibus Kruskal-Wallis tests were adjusted using the Holm method. Post hoc Dunn-Holm comparisons were only interpreted for outcomes with a significant omnibus result after correction.

As an exploratory multivariable analysis, a linear regression model with heteroskedasticity-consistent robust SEs was fitted using the general attitude toward AI score as the dependent variable. Covariates included the final AI definition category, age, academic year, class ranking, and career aspirations (hospital-based clinical work, research, and private practice, each coded as yes or no). Results are reported as β coefficients with 95% CIs. All tests were 2-sided, and P values of less than .05 were considered statistically significant.

The qualitative component focused on free-text responses collected through the open-ended items of the survey. For the question assessing students’ definitions of AI, only the incorrect responses were analyzed, as the intention was to explore students’ misconceptions and the underlying representations associated with them. Additional free-text responses regarding feelings toward AI and expectations for AI training were also included in the qualitative dataset.

All qualitative data were compiled in Microsoft Excel and analyzed using a thematic analysis approach. Coding was performed iteratively: an initial set of codes was developed deductively based on concepts previously described in the literature (eg, nuanced optimism and replacement concerns), and additional categories were informed inductively from the data. Codes were progressively merged into broader categories through constant comparison, resulting in the identification of overarching themes. To enhance the reliability of the analysis, a second reviewer independently coded 30% of the material, and discrepancies were resolved through discussion.

Given the large and fully anonymous dataset, data saturation was not assessed, and validation by participants could not be performed. Representative quotations were selected to illustrate each theme. The study adhered to the COREQ (Consolidated Criteria for Reporting Qualitative Research) guidelines (Checklist 1).

All data were compiled in Excel version 2019 (Microsoft), and descriptive and qualitative analyses were performed using Microsoft Excel version 2019. Statistical tests were conducted using R software (version 4.3.1; R Foundation for Statistical Computing).

All authors confirm that all methods were applied in accordance with current guidelines and regulations. This study did not require specific authorization under current French regulations, as it is research evaluating the value of teaching through a student health survey. This research does not fall under clinical research and, in France, does not require ethics committee approval (French Public Health Code: articles L1110-1 to L6441-1) [35].

To ensure full compliance with ethical standards, we implemented rigorous safeguards throughout the study. No sensitive personal data, such as health information, ethnic origin, or any identifiable details, were collected at any stage. Participant anonymity was strictly preserved: Responses were untraceable, with no collection of IP addresses or other identifying metadata, and all data were analyzed in aggregated form to prevent individual recognition. Informed consent was secured through a comprehensive information letter provided to all participants prior to the survey, clearly outlining the study’s purpose, the voluntary nature of participation, and the intended use of data, with completion of the survey serving as implied consent. A request for consent was clearly written. Finally, data storage and security adhered to the General Data Protection Regulation (GDPR) requirements, with all responses encrypted, accessible only to the research team, and permanently deleted following analysis in 3 years. The email addresses were deleted immediately after the study.

This study aimed to explore medical students’ representations and attitudes toward AI in France in 2025. Our results indicate that students’ attitudes toward AI depend essentially on their knowledge of the definition of AI, as students with more accurate definitions tend to report more positive attitudes. There is also great variability in students’ attitudes toward AI, with marked differences not only between participants but also within individual responses. These differences highlight the complexity of student perceptions of AI. The themes of “nuanced optimism” and “replacement” align with observations from other studies [18,23,24,26]. New themes also emerged, such as fear of skill loss, awareness of the high ecological cost of AI use, and loss of credibility with patients. Concerns about skill loss have also been reported among educators, particularly regarding the risk of cognitive offloading and erosion of critical thinking skills, suggesting that this apprehension is shared across different stakeholders in medical education [10,11]. Ecological concerns, meanwhile, align with broader societal discussions on the energy consumption and environmental impact of large-scale AI systems, transcending the specific context of medical education [36].

Additionally, the results reveal how AI is used for learning and searching for information on the internet. However, the need for an ethical framework and responsible AI use remains a major concern for students, even though the World Health Organization (WHO) recommendations have been in place since 2021 [2].

Our study reveals 2 notable developments compared to previous work. First, only a small percentage of students reported not knowing how to define AI, compared to half of the participants in the studies by Teng et al [21]. In contrast, in our study, half of the students gave incorrect answers, suggesting that students now develop representations of AI, even if they are often erroneous, rather than being completely unaware. Their concerns about ethical considerations, AI management, and its impact on their careers remain similar to those observed by Teng et al [21]. As noted by Amiri et al [19], students’ prior knowledge strongly influences their perceptions. However, neither the level of education nor the academic year showed a significant effect [19]. These results are likely related to self-assessment of their ranking within their class. Finally, the theme of “replacement” persists [19,21], as does ambivalence between perceived AI utility and its risks. Finally, unlike in some studies [18,23,26], no formal training request emerged, probably due to students’ self-training or the still-limited clinical use of AI.

We also highlighted a form of critical consideration among students toward AI. Indeed, students expressed ethical concerns and reservations about using AI in medical contexts [20]. For example, they emphasized the need for a precise framework and established ethical rules, but this critical attitude seems less present when it comes to their own learning. Students show some mistrust toward AI as a clinical tool, recognizing its limitations and the need for human supervision. However, this critical consideration is less evident in their use of AI for learning, where they seem more inclined to accept the provided information without always questioning it. This dichotomy suggests that students perceive greater risks related to AI in medical practice than in their own learning process. It would, therefore, be crucial to encourage a more critical approach to AI use in learning, emphasizing the importance of verifying and validating the obtained information to develop a more thoughtful and responsible use of these technological tools [37].

AI use by medical students is still relatively unexplored compared with studies in other educational fields [30,32,38]. Our study revealed that medical students use AI in various ways to meet their learning needs, primarily for writing assistance, information searches, or personalized learning. It seems that this use in their learning could lead to excessive and uncritical use of information. Moreover, it is crucial to explore the impact on skill acquisition and dependence on AI in their professional environments. However, this adoption of AI in their learning process could potentially result in excessive and uncritical use of the provided information, risking the creation of dependence on these tools. On the other hand, this could also lead to the loss of essential skills, particularly regarding clinical judgment and decision-making autonomy, which are crucial in medical practice. Therefore, it is essential to explore these potential impacts on both the acquisition of critical medical skills and students’ future professional practice.

Our study had certain limitations. First, the response rate relative to the number of enrolled students was less than 10%, which limits the generalizability of our findings. The use of voluntary participation may have introduced self-selection bias, as students with stronger preexisting interest in AI, whether positive or negative, may have been more likely to respond. While participation exceeded 10% at some universities, the low overall rate suggests that our results may disproportionately reflect the attitudes of a more engaged subgroup rather than the broader student population. Second, we chose the Teng et al [21] survey because it was designed 4 years ago, allowing for comparison. Its mixed methodology offered the advantage of exploring both the underlying reasons for students’ attitudes and their modes of AI use [20]. However, regarding the evaluation of the definition, it is important to note that it is not possible to assess students’ knowledge based solely on their ability to define AI. Other surveys that evaluate AI knowledge more globally could have been used, but their length might have discouraged participation and thus reduced the response rate. A key limitation of our study is the absence of demographic data, including gender or sexual orientation. We deliberately chose not to collect this information. This decision was motivated by two ethical concerns: (1) the need to avoid reductive or exclusionary categorizations of gender, which is a spectrum of identities extending beyond binary frameworks; and (2) the risk of compromising participant anonymity, as detailed demographic data could have enabled indirect identification in smaller faculty samples. Finally, although several associations reached statistical significance due to the large sample size, the magnitudes are uniformly small, and practical implications should be interpreted cautiously.

Our study describes the current perceptions, attitudes, and uses of AI among French medical students in 2025. Students report generally positive attitudes toward AI despite limited conceptual knowledge of the definition of AI, alongside concerns related to ethics, ecological impact, and potential skill loss. Many already use AI tools for learning in a largely self-directed manner. Further longitudinal research is needed to better understand the implications of these practices for medical education and professional development. These findings provide a descriptive baseline for future evaluations of AI-related training and its impact on learning and professional development.