Artificial Intelligence in Medical Education: A Cross-Sectional Study of Usage Patterns and Perceptions Among Medical Students at a Brazilian Public University

Abstract

Purpose

To describe self-reported artificial intelligence (AI) use and perceptions of benefits, risks, and curricular integration among medical students with previous AI contact at a Brazilian public university.

Methods

This descriptive cross-sectional survey included a convenience sample of 189 students aged 18 years or older in Years 1–4 of medical school who reported previous contact with AI tools applied to medical education. Data were collected from March to July 2026 using a researcher-developed electronic questionnaire without formal validation and summarized descriptively using available responses.

Results

Daily or weekly academic AI use was reported by 76.2% (144/189). Exposure to AI as a learning resource in a course was reported by 38.3% (72/188). Perceived improvement in differential diagnosis was reported by 57.8% (108/187), whereas 60.3% (114/189) perceived a risk of harmful dependence affecting human clinical judgment. Optional and controlled curricular integration was preferred by 72.0% (136/189).

Conclusion

Among respondents with previous AI contact, frequent use coexisted with perceived benefits and concerns about dependence. These findings can support local discussion and further educational evaluation but do not establish objective learning effects or determine appropriate curricular approaches.

Introduction

The integration of artificial intelligence (AI) into medical education has expanded rapidly in recent years, particularly following the widespread availability of generative AI and large language models.^1,2 Studies indicate that medical students use these technologies to access educational resources, summarize content, clarify questions, simulate clinical situations, and explore clinical problems.^1,2 AI may therefore complement conventional teaching strategies and provide new ways for students to engage with medical knowledge.

These potential benefits coexist with concerns about how AI is used in undergraduate medical education. Although generative tools can expand access to learning resources, excessive or dependent use may reduce cognitive engagement and undermine the development of critical thinking.² Generative AI may support cognitive development when incorporated into educational strategies that promote reflection and critical appraisal; however, it may threaten learner autonomy when it replaces reasoning processes that students should perform independently.¹ Furthermore, AI-generated content may contain plausible but inaccurate, biased, or otherwise misleading information, reinforcing the need for critical appraisal, information verification, and appropriate educational guidance.^1,2

These issues are particularly relevant because medical education requires the progressive development of clinical reasoning, decision-making autonomy, and humanistic competencies. Although AI may support information organization, clinical simulation, and the exploration of diagnostic hypotheses, its educational use should preserve students’ active engagement in learning. Medical students have also expressed concerns about potential effects on the humanistic dimensions of medicine, the physician–patient relationship, and trust in clinical interactions.³ The integration of AI into medical education should therefore consider its intended purpose, conditions of use, and the competencies that educational activities are designed to develop.

AI adoption also has curricular implications. Although educational initiatives involving AI are emerging, consolidated curricular frameworks and pedagogical approaches for integrating these technologies throughout medical education remain limited.⁴ The literature recommends that medical curricula address ethical considerations such as patient autonomy, bias, and transparency.⁵ Medical students have also expressed interest in structured training on the ethical implications and clinical applications of AI.³

Brazilian studies have already documented favorable attitudes towards AI alongside concerns about over-reliance and the potential loss of medical skills.⁶ Kanashiro et al also described educational use, perceived learning benefits, and concerns about the reliability of generated information among medical students at another Brazilian institution.⁷ These findings provide a basis for comparison within Brazil and indicate that the contribution of further studies should be defined by the educational setting, population, and dimensions examined.

Accordingly, this study aimed to describe reported AI use, course-related experiences, perceived benefits and risks, and preferences for curricular integration among students with previous AI contact at a public medical school in the Brazilian Legal Amazon. Its contribution is a context-specific descriptive account of these dimensions within one educational setting. The study does not test relationships between these dimensions, estimate AI use across all students in the program, or establish the effectiveness of particular curricular approaches.

Materials and Methods

Study Design, Setting, and Period

This observational, descriptive cross-sectional study used a quantitative approach and included medical students from the Palmas Campus of the Federal University of Tocantins (UFT), Brazil, in the Brazilian Legal Amazon. Data were collected between March and July 2026. This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement for cross-sectional studies.⁸ The completed checklist is provided as the STROBE checklist (Supplementary Material). The study considered experiences involving AI tools in courses, academic activities, laboratory practices, and other educational experiences related to the undergraduate medical program. This scope did not require institutionally provided tools or use on university premises. However, the screening question did not distinguish personal use from institutionally introduced AI, and no minimum frequency or level of exposure was specified.

The medical program comprises twelve semesters over six years. Semesters 1–2 correspond to Year 1, semesters 3–4 to Year 2, semesters 5–6 to Year 3, and semesters 7–8 to Year 4. The present study covered these first four years and did not include the clerkship semesters 9–12. The published curriculum includes basic sciences in the early semesters, community health experiences before clerkship, and progressively more clinically oriented coursework in later semesters.⁹ Academic semester indicates curricular position rather than an individual measure of clinical competence.

During the study period, AI was not formally integrated into the UFT medical curriculum. Faculty members were encouraged to use AI tools to improve teaching, while students used these tools according to their own needs and knowledge. This contextual description reflects the investigators’ account of institutional practice rather than a systematic curricular audit.

Study Population and Sample

The source population comprised 336 students enrolled in semesters 1–8 (Years 1–4) of the medical program during the study period. All students in these semesters were invited to participate. Non-probability convenience sampling was used. No formal sample-size calculation was performed. The analytical sample comprised respondents who met the eligibility criteria, reported previous contact with AI, continued beyond the screening question, and provided age information confirming eligibility. The semester-based source population was not an independently verified count of age-eligible students.

A total of 224 responses were initially recorded. The screening question was, “Have you had previous contact with artificial intelligence tools applied to medical education?” An affirmative self-report determined access to the subsequent questions on use and perceptions; it did not establish formal curricular training, competence in AI, or a minimum frequency or level of exposure. The programmed branching was used because the subsequent questions focused primarily on participants’ reported experiences with AI, including frequency of use, tools used, confidence when relying on AI, perceived effects of use, and possible dependence. Twenty-eight respondents selected the negative response and exited according to the programmed survey flow. These were planned terminations before the main questionnaire sections rather than exclusions resulting from inconsistent or invalid records.

Although Supplementary Material 1 lists age, sex, and semester before screening, these data were not retained for the 28 respondents who reported no previous AI contact. Therefore, baseline comparisons with this group were unavailable. Examples of tools appeared in subsequent questions and should not be interpreted as a standardized definition presented before the screening item.

Among the 196 respondents who continued, seven records were removed during data preparation: five because age information was missing and eligibility could not be confirmed, and two because the reported age was below 18 years. The final analytical sample comprised 189 students, corresponding to 56.3% (189/336) of the invited semester-based population.

Eligibility Criteria

Students were eligible if they were aged 18 years or older, were regularly enrolled in semesters 1–8 (Years 1–4) of the institution’s medical program, and voluntarily agreed to participate by providing electronic informed consent.

Records indicating age below 18 years were removed because they did not meet the stated age criterion. Records without age information were removed because age eligibility could not be confirmed. Missing responses to other questionnaire items did not result in complete exclusion of a participant; the corresponding variables were analyzed using the available responses.

Data Collection Instrument

Data were collected using an anonymous, self-administered electronic questionnaire hosted on Google Forms. Item development drew on relevant literature, including the study by Jackson et al,³ and on questions arising from the student researcher’s own experience and interests as a medical student. No independent expert review or formal validation was conducted.

Before the main data collection, an informal pretest involved members of the research team, including a researcher affiliated with another medical school, and two student teaching assistants from the first and fourth semesters, respectively. The pretest assessed question comprehension and the functioning of the electronic form. No questions were changed, and the two teaching assistants’ pretest responses were not included in the 189 analyzed records. This limited exercise did not establish content or psychometric validity or equivalent comprehension across training stages.

The instrument included closed-ended single-response questions, multiple-response questions, a numerical rating item from 0 to 10, and one open-ended question. In addition to participant characteristics and screening, its descriptive domains comprised patterns of use (items 6–7), clinical and relational perceptions (items 8–12 and 19), curricular preferences (item 14), course and assessment experiences (items 17–18 and 20), perceived dependence and future use (items 21–22), and a single positive-influence rating (item 23). These heterogeneous items were not combined into an overall scale. The domains are not validated measures of AI literacy, dependence, or clinical competence.

The open-ended question asked participants how AI should be used in medical education. Although this question was included in the administered questionnaire, its responses were not analyzed in the present study, which was limited to a descriptive quantitative analysis. The complete questionnaire is provided in Supplementary Material 1.

Items addressing clinical reasoning, diagnostic hypothesis formulation, differential diagnosis, autonomy, confidence, clinical skills, and relational aspects were interpreted as self-reported perceptions. They did not objectively assess learning, diagnostic ability, or clinical competence. Their interpretation may vary with training stage, particularly when early-year students answer questions about clinical decisions. The wording and response options of the administered questionnaire are retained in Supplementary Material 1.

Data Collection Procedures

Recruitment notices were delivered in classrooms, and reminders containing a QR code linked to the questionnaire were displayed in teaching spaces. The electronic form was configured to permit only one response per participant to reduce the possibility of duplicate submissions.

Before accessing the questionnaire, participants received information about the study objectives, voluntary participation, confidentiality, and potential risks and benefits. Access to the questions was conditional on electronic agreement with the informed consent form. No direct personal identifiers were requested, and Google Forms was configured not to collect email addresses.

After data collection was completed, responses were exported to an electronic database and assessed for eligibility, completeness, and consistency.

Potential Sources of Bias

Recruitment notices and QR-code reminders were used to facilitate participation among students in the eligible academic semesters. Restricting the electronic form to one response per participant reduced the possibility of duplicate entries, while the anonymous, self-administered format removed direct interviewer administration. Nevertheless, voluntary participation may have introduced selection bias. The self-reported nature of the responses may also have been affected by recall, interpretation, and social desirability biases. In addition, the absence of formal questionnaire validation may have introduced measurement error and limited the comparability of the findings with those of other studies.

Study Variables

Sociodemographic and academic variables included age, sex, academic semester, academic performance coefficient category, and participation in teaching assistantships, scientific initiation activities, or student academic societies. Age was recorded in years and standardized to numerical values. Sex and participation in complementary academic activities were nominal categorical variables; semester and substantive academic performance categories were ordinal. “Prefer not to answer” was retained as a separate response option for academic performance. Semesters 1–2 corresponded to Year 1, 3–4 to Year 2, 5–6 to Year 3, and 7–8 to Year 4.

Previous contact with AI was used as the screening variable. Among participants who answered affirmatively, categorical variables included frequency of AI use, types of tools used, exposure to AI in academic courses, and use of AI to address simulated or real clinical cases. The items concerning tools used and skills potentially compromised allowed multiple responses.

Additional categorical variables assessed perceptions of possible impairment in diagnostic hypothesis formulation, development of traditional clinical reasoning, confidence when relying on AI, relationships with patients or interactions during clinical simulations, technological dependence, perceived improvement in differential diagnosis, comfort with future AI use, and preferred approaches to incorporating AI into the medical curriculum.

The perceived positive influence of AI on medical education was assessed using a single numerical item ranging from 0 to 10, with higher values indicating a more positive perceived influence. An arbitrary threshold of ≥7 was used solely to summarize the proportion of respondents selecting ratings toward the upper end of the scale. This threshold was not validated as a classification of benefit or competence, and scores below 7 were not interpreted as necessarily indicating neutral or negative perceptions.

Data Management and Statistical Analysis

Responses were assessed for eligibility, completeness, consistency, and standardization. Age entries containing additional text were converted to numerical format while preserving the reported value. All other variables were retained according to the response options provided in the questionnaire.

Only descriptive statistical analyses were conducted. Categorical variables were summarized using absolute and relative frequencies. For multiple-response questions, each response option was counted separately, using the number of participants with an available response to that item as the denominator; consequently, percentages could total more than 100%.

The age distribution was assessed by visual inspection of a histogram. Because the distribution was right-skewed, age was summarized using the median and interquartile range. The positive-influence rating was summarized using the mean and standard deviation, median and interquartile range, minimum and maximum, and the number and proportion of ratings of 7 or higher.

Missing data were not imputed. Percentages and descriptive measures for each variable were calculated using the number of available responses, and the corresponding denominator was reported whenever it differed from the total analytical sample. The open-ended responses were not subjected to qualitative or quantitative analysis in the present report.

Because convenience sampling was used, no sampling weights were applied. No inferential, subgroup, or sensitivity analyses were performed. Database organization and descriptive statistical analyses were conducted using Microsoft Excel for Microsoft 365 (Microsoft Corporation, Redmond, WA, USA).

During manuscript revision, an additional computational verification of the tabulated frequencies and descriptive summaries was performed using Python (version 3.12.14) and NumPy (version 2.3.5). Quartiles were calculated using linear interpolation, and the sample standard deviation was calculated using an n−1 denominator. This procedure was used only to verify the descriptive results and did not constitute an additional statistical analysis.

Results

Participant Flow

Of the 336 students invited, 224 submitted responses, corresponding to an initial participation proportion of 66.7% (224/336). Among these respondents, 87.5% (196/224) reported previous AI contact and continued to the subsequent questionnaire sections, whereas 12.5% (28/224) reported no previous contact and exited according to the survey flow.

Among the 196 respondents who continued, five records lacked age information and two indicated an age below 18 years. The final analytical sample comprised 189 students, equivalent to 56.3% (189/336) of the invited semester-based population. Initial participation and analytical inclusion are distinct proportions. An eligibility-adjusted response rate cannot be established from these counts because age eligibility was not determined for all invited students.

All participants in the analytical sample had answered affirmatively to the screening question. Missing data for the closed-ended items included in the analysis ranged from 0% to 1.6%. No missing data were imputed, and each variable was analyzed using the available responses.

Sociodemographic and Academic Characteristics

Participants had a median age of 22 years (interquartile range, 21–24 years). Female students accounted for 57.1% (108/189). All included semesters were represented, with the largest proportion enrolled in the third semester (20.7%; 39/188). Most respondents reported an academic performance coefficient above 8.0 (86.7%; 163/188) and participation in teaching assistantships, scientific initiation activities, or student academic societies (86.8%; 164/189). Table 1 presents these characteristics and the semester-to-year mapping.

Table 1.

Sociodemographic and Academic Characteristics of Participants (n=189)

Characteristic	n (%) or Summary Measure
Sex
Female	108 (57.1)
Male	81 (42.9)
Academic semester (valid n=188)
1st Semester (Year 1)	13 (6.9)
2nd Semester (Year 1)	24 (12.8)
3rd Semester (Year 2)	39 (20.7)
4th Semester (Year 2)	17 (9.0)
5th Semester (Year 3)	31 (16.5)
6th Semester (Year 3)	27 (14.4)
7th Semester (Year 4)	24 (12.8)
8th Semester (Year 4)	13 (6.9)
Academic performance coefficient (valid n=188)
Below 6	0 (0.0)
6–7	1 (0.5)
7–8	7 (3.7)
>8	163 (86.7)
Prefer not to answer	17 (9.0)
Participation in teaching assistantships, undergraduate research, or student academic societies
Yes	164 (86.8)
No	25 (13.2)
Age, years
Median (Q1-Q3)	22 (21–24)

Notes: Data are presented as n (%) unless otherwise indicated. Percentages were calculated using the available responses for each variable; valid denominators are shown when different from n=189. Missing data were not imputed. Semesters 1–2 correspond to Year 1, 3–4 to Year 2, 5–6 to Year 3, and 7–8 to Year 4. Academic performance coefficient categories are presented as offered in the questionnaire. “Prefer not to answer” was retained as a separate response category. Percentages may not sum to 100% because of rounding.

Abbreviations: Q1, first quartile; Q3, third quartile.

Curricular Experience and Patterns of AI Use

Daily academic AI use was reported by 52.4% (99/189) and weekly use by 23.8% (45/189), giving 76.2% (144/189) reporting daily or weekly use. Exposure to AI as a learning resource in a course was reported by 38.3% (72/188). This item described a course-related experience rather than verified formal integration of AI into the curriculum.

All respondents to the tools question reported using language models such as ChatGPT, Gemini, or Claude (100%; 188/188). Clinical simulators, adaptive learning platforms, NotebookLM, and diagnostic support software were reported less frequently. The reported categories reflect questionnaire responses and are not mutually exclusive technological classes.

Use of AI to address simulated or real clinical cases was reported by 69.4% (129/186), and participation in academic assessments directly supported by AI by 22.0% (41/186). Course-related experiences and patterns of use are summarized in Table 2.

Table 2.

Self-Reported Course-Related Experiences and Patterns of Artificial Intelligence Use (n=189)

Characteristic	n (%)
AI as a learning resource in a course (valid n=188)
Yes	72 (38.3)
No	100 (53.2)
Unsure	16 (8.5)
Frequency of academic AI use
Daily	99 (52.4)
Weekly	45 (23.8)
Occasionally	41 (21.7)
Rarely	3 (1.6)
Never	1 (0.5)
Use of AI in simulated or real clinical cases (valid n=186)
Yes	129 (69.4)
No	57 (30.6)
Participation in academic assessments directly supported by AI (valid n=186)
Yes	41 (22.0)
No	145 (78.0)
AI tools used; multiple responses allowed (valid n=188)
Language models	188 (100.0)
Clinical simulators	23 (12.2)
Adaptive learning platforms	9 (4.8)
NotebookLM	4 (2.1)
Diagnostic support software	1 (0.5)

Notes: Data are presented as n (%). Percentages were calculated using the available responses for each item; valid denominators are shown when different from n=189. Missing data were not imputed. The tools item allowed multiple responses, so percentages may sum to more than 100%. Tool categories reflect the reported responses and may overlap technologically. All four respondents reporting NotebookLM also selected language models. Course-related experiences do not establish formal curricular integration.

Abbreviation: AI, artificial intelligence.

Perceived Benefits for Medical Education

Perceived improvement in differential diagnosis was reported by 57.8% (108/187): slight improvement by 35.3% (66/187) and substantial improvement by 22.5% (42/187). Respondents were unsure in 31.0% of cases (58/187), while 10.2% (19/187) reported no change and 1.1% (2/187) reported worsening.

The positive-influence rating had a mean of 7.0 (SD, 2.3), median of 7 (interquartile range, 6–8), and range of 0–10. Scores of 7 or higher were reported by 74.1% (140/189) (Table 3).

Table 3.

Self-Reported Perceptions of the Benefits, Risks, and Curricular Integration of Artificial Intelligence (n=189)

Characteristic	n (%) or Summary Measure
Potential impairment of diagnostic hypothesis formulation
Strongly agree	4 (2.1)
Partially agree	48 (25.4)
Neutral	38 (20.1)
Partially disagree	55 (29.1)
Strongly disagree	44 (23.3)
Potential impairment of traditional clinical reasoning
Yes	56 (29.6)
No	112 (59.3)
Unsure	21 (11.1)
Confidence when relying on AI (valid n=188)
More confident	35 (18.6)
Neither more nor less confident	64 (34.0)
Less confident	89 (47.3)
Perceived effect on relationships with patients or interactions during clinical simulations (valid n=187)
Positive	73 (39.0)
Neutral	91 (48.7)
Negative	23 (12.3)
Potential risk of harmful dependence on AI affecting human clinical judgment
Yes	114 (60.3)
No	36 (19.0)
Unsure	39 (20.6)
Perceived improvement in differential diagnosis (valid n=187)
Substantial improvement	42 (22.5)
Slight improvement	66 (35.3)
No change	19 (10.2)
Worsened	2 (1.1)
Unsure	58 (31.0)
Comfort with AI-assisted clinical decision-making during clinical clerkship or residency training
Comfortable	42 (22.2)
Not comfortable	37 (19.6)
Depends on the situation	110 (58.2)
Preferred approach to incorporating AI into the medical curriculum
Regular and mandatory	15 (7.9)
Optional and controlled	136 (72.0)
Limited to specific tools	35 (18.5)
Should not be incorporated	3 (1.6)
Perceived positive influence of AI on medical education
Mean (SD)	7.0 (2.3)
Median (Q1–Q3)	7 (6–8)
Minimum–maximum	0–10
Score ≥7	140 (74.1)

Notes: Data are presented as n (%) unless otherwise indicated. Percentages were calculated using the available responses for each item; valid denominators are shown when different from n=189. Missing data were not imputed. The positive-influence item ranged from 0 to 10, with higher values indicating a more positive perceived influence. The arbitrary threshold of ≥7 was used solely as a descriptive summary and was not validated as a classification of benefit or competence. Scores below 7 do not necessarily indicate neutral or negative perceptions. Percentages may not sum to 100% because of rounding.

Abbreviations: AI, artificial intelligence; Q1, first quartile; Q3, third quartile; SD, standard deviation.

Perceptions of Clinical Reasoning, Autonomy, and Relational Aspects

Regarding potential impairment of diagnostic hypothesis formulation, 52.4% (99/189) partially or strongly disagreed, 27.5% (52/189) partially or strongly agreed, and 20.1% (38/189) were neutral. For perceived impairment of traditional clinical reasoning, 59.3% (112/189) answered no, 29.6% (56/189) answered yes, and 11.1% (21/189) were unsure.

When relying on AI for clinical decisions, 47.3% (89/188) reported feeling less confident, 34.0% (64/188) were indifferent, and 18.6% (35/188) felt more confident.

A risk of harmful dependence on AI affecting human clinical judgment was perceived by 60.3% (114/189). In the multiple-response item, autonomous decision-making was selected by 78.8% (149/189), the ability to improvise in real clinical cases by 76.7% (145/189), clinical reasoning by 51.3% (97/189), and empathy and communication by 24.9% (47/189). The “None of the above” option was selected by 3.2% (6/189) (Figure 1).

Figure 1.

Skills perceived by medical students as potentially compromised by excessive AI use (n=189). Labels show the number selecting each option divided by the number of respondents, followed by the percentage. Participants could select more than one response; therefore, percentages may sum to more than 100%. All six respondents selecting “none of the above” selected that option alone.

Abbreviation: AI, artificial intelligence.

For relationships with patients or interactions during clinical simulations, 48.7% (91/187) perceived a neutral effect, 39.0% (73/187) a positive effect, and 12.3% (23/187) a negative effect. Table 3 presents these perceptions and their valid denominators.

Curricular Integration and Future Clinical Use

Optional and controlled curricular integration was preferred by 72.0% (136/189), integration limited to specific tools by 18.5% (35/189), regular and mandatory integration by 7.9% (15/189), and no integration by 1.6% (3/189).

Comfort with future AI-assisted clinical decisions during clerkship or residency depended on the situation for 58.2% (110/189); 22.2% (42/189) would feel comfortable and 19.6% (37/189) would not (Table 3).

Discussion

Among respondents with previous AI contact, frequent academic use coexisted with perceived benefits and concerns about dependence, autonomy, and clinical competencies. This combination suggests that familiarity with AI does not necessarily translate into unreserved acceptance of its use. Rather, students appeared to recognize both its practical value and potential limitations. Because the analytical sample included only respondents with previous AI contact, these findings should be interpreted as reflecting the perceptions and experiences of students already exposed to AI rather than those of the entire medical student population.

The predominance of language models is consistent with their accessibility and broad range of educational applications described in the literature.^1,2 Their versatility may also help explain their prominence, as the same tool can be used for content explanation, summarization, question generation, writing support, and assistance with clinical reasoning exercises. In Germany, Fußhöller et al reported ChatGPT use for learning alongside concerns about misinformation and limited confidence in AI-related skills.¹⁰ Direct numerical comparisons nevertheless require caution because their measure of having used ChatGPT for medical education differs from daily or weekly AI use among respondents selected for previous AI contact in the present study. Differences in survey period, training stage, access to AI tools, and question wording may also contribute to variation across studies.

Reported course-related AI experiences should also be distinguished from formal curricular integration. According to the investigators’ account, the UFT medical curriculum did not formally incorporate AI during the study period, although faculty members were encouraged to use AI tools to improve teaching. Thus, students may encounter AI within individual courses or assessments even in the absence of an institutionally defined AI curriculum. This distinction is educationally relevant because informal exposure does not necessarily provide structured training in verification, critical appraisal, appropriate delegation of tasks, or limitations of AI-generated content. The questionnaire did not independently audit curricular provision or determine whether each reported use episode was self-directed, teacher-guided, or institutionally organized, and therefore cannot establish the extent or quality of institutional integration.

Perceived benefits are consistent with favorable evaluations reported in Brazilian and international surveys, including those by Kanashiro et al and Sami et al.^7,11 Such perceptions may reflect the immediate usefulness of AI for obtaining explanations, organizing information, generating study materials, and supporting academic tasks. However, perceived improvement in differential diagnosis should not be interpreted as evidence of improved diagnostic accuracy. Laupichler et al assessed AI literacy using pre-validated self-report instruments, whereas the present questionnaire did not assess that construct equivalently.¹² Similarly, Kubrusly et al examined patterns of AI engagement in relation to self-perceived competence in a Brazilian sample.¹³ Together, these distinctions reinforce that adoption, perceived benefit, confidence, and competence represent different constructs and should not be treated as interchangeable outcomes.

Concern about dependence is also consistent with findings from other settings. Giavina-Bianchi et al described favorable attitudes toward AI alongside concerns about over-reliance and possible loss of medical skills in a Brazilian medical school.⁶ Zhang et al reported widespread AI use together with limitations in students’ self-reported critical evaluation of AI-generated information across Chinese medical schools.¹⁴ The coexistence of frequent use and concern about dependence may therefore reflect a broader tension in medical education: students may value the efficiency and accessibility of AI while simultaneously recognizing the risks of delegating reasoning, verification, or judgment to automated systems. In the present study, however, this interpretation remains hypothetical because no cross-tabulation was performed to determine whether the same individuals who used AI frequently were also those expressing greater concern about dependence.

The wording of the questions also requires attention when interpreting this apparent tension. A question concerning possible impairment from AI use in general is conceptually different from one addressing potential harm from excessive use. Students may therefore endorse both frequent use and concern about overuse without holding inconsistent views. Rather than suggesting rejection of AI, these responses may indicate an emerging awareness that usefulness depends on the conditions under which AI is used. From an educational perspective, this distinction is important because it shifts the discussion from whether students should use AI to how they should evaluate, verify, and appropriately delimit AI-assisted work.

Interpretation of relational and clinical items requires particular caution. The question concerning relationships combined actual patient encounters with clinical simulations, which involve different levels of responsibility and interpersonal complexity. Students in earlier years may have responded based on impressions or expectations rather than substantial clinical experience, and several items did not include a not-applicable option. Although concerns about the effects of AI on humanistic dimensions of medicine have been described elsewhere, the present findings cannot establish changes in empathy, communication, or real patient care.³ Nevertheless, they suggest that future studies should distinguish simulated from real clinical contexts and account explicitly for stage of training when examining students’ perceptions of AI-related effects on clinical interaction.

The preference for optional and controlled integration should likewise be interpreted as an expression of students’ views rather than evidence that this curricular model is superior. Montrezor and Taxini described a structured first-year physiology activity in which students critically appraised AI-generated content.¹⁵ This type of guided educational experience differs from the heterogeneous and partly informal experiences captured in the present study. Even so, it illustrates how AI exposure can be embedded within activities that emphasize appraisal rather than passive acceptance of generated output. Framework and ethics literature has similarly highlighted curricular domains such as critical appraisal, bias, transparency, professional responsibility, and human oversight.^4,5 Taken together with the present findings, these considerations suggest that the educational priority may not simply be increasing exposure to AI, but helping students develop explicit criteria for when, how, and to what extent AI should inform academic and clinical reasoning.

Several limitations constrain interpretation. The single-center convenience sample and voluntary participation limit generalizability and may have favored students with greater interest in AI, although the direction and magnitude of this potential selection bias cannot be determined. The main analyses were restricted to respondents who reported previous AI contact and therefore do not describe all students enrolled in the program. No use or perception outcomes were obtained for respondents who exited at screening, and their age, sex, and semester data were not retained; consequently, they could not be characterized or compared with the analytical sample.

The questionnaire was not formally validated, and the informal pretest did not establish equivalent comprehension across training stages. The screening wording may have been interpreted differently among respondents, and some clinical items may have been particularly difficult for students in earlier years. In addition, some questions combined real and simulated experiences or lacked a not-applicable response option. These features may have introduced measurement error that cannot be quantified from the available data. Responses reflect perceptions and may also be influenced by recall and social desirability. Available-case analyses resulted in different item denominators, and the cross-sectional design, absence of objective outcomes, and lack of subgroup or association analyses preclude conclusions about changes in learning, actual dependence, or clinical competence.

Future studies should include students with and without previous AI contact and use questions tailored to different stages of training. Cognitive interviews, independent expert assessment, and pilot testing should precede wider use of a revised instrument. Multicenter sampling would broaden the range of educational settings represented, while longitudinal or interventional designs incorporating objective educational and clinical outcomes would be needed to evaluate whether different patterns of AI use affect learning, reasoning, or competence over time. Within its descriptive scope, the present study identifies a relevant educational tension: students may value AI as an academic resource while also recognizing the need to preserve independent reasoning and professional judgment. This tension provides a useful basis for local curricular discussion and for subsequent studies evaluating how AI can be incorporated into medical education with appropriate critical oversight.

Conclusion

Among respondents with previous AI contact at this Brazilian public medical school, academic AI use was frequent and coexisted with perceived benefits, concerns about dependence, and a preference for optional and controlled curricular integration. These descriptive findings can inform local educational discussion and help define priorities for further evaluation of AI use in medical training. Although they do not establish objective effects on learning or clinical performance or determine specific curricular content, teaching methods, or requirements, they highlight the need to examine how AI can be incorporated while preserving critical appraisal, independent reasoning, and appropriate professional judgment.

Funding Statement

This study received support from the Institutional Scientific Initiation Scholarship Program of the Federal University of Tocantins (PIBIC/UFT), through a scientific initiation scholarship awarded to AVFL under Call No. 129/2025. The funding body had no role in the study design, data collection, data analysis, interpretation of the findings, preparation of the manuscript, or decision to submit the article for publication.

Declaration of Generative AI and AI-Assisted Technologies

The authors used ChatGPT (OpenAI; model: GPT-5.6) and Codex (OpenAI; model: gpt-6-astra) for language editing, graphical support, consistency checks, assistance with literature searching, and preparation of manuscript revisions and responses to reviewers. These tools supported the organization and wording of the material and the preparation of computational procedures used to check descriptive results. The authors reviewed and edited the assisted content and independently verified the accuracy of the manuscript and its references. The authors confirm the originality and accuracy of the final manuscript and take full responsibility for the integrity of its entire content. They reviewed the terms of use applicable to the tools and confirm that their use of the content is suitable for publication. These tools were not used to generate or alter primary study data.

Abbreviations

AI, artificial intelligence; IQR, interquartile range; SD, standard deviation; STROBE, Strengthening the Reporting of Observational Studies in Epidemiology; UFT, Federal University of Tocantins.

Data Sharing Statement

The individual-level data are not publicly available because of the confidentiality commitments established in the study’s ethical protocol. Aggregated data supporting the findings are presented in the article and its tables and figures.

Ethics Approval and Informed Consent

The study was approved by the Research Ethics Committee of the Federal University of Tocantins (Approval No. 7934172; CAAE No. 88243325.00000.5519). All participants voluntarily provided electronic informed consent before completing the questionnaire. The study was conducted in accordance with the Declaration of Helsinki and Brazilian National Health Council Resolution No. 466/2012. Participant confidentiality and the rights to decline participation or withdraw without penalty were ensured.

Consent for Publication

Not applicable. The manuscript contains no identifiable personal information, images, or recordings.

Author Contributions

AVFL and SAF conceived the study, collected and organized the data, reviewed the literature, and drafted the manuscript. ACGR supervised the study and contributed to the methodology, interpretation of the findings, and critical revision of the manuscript. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.