Diagnostic reasoning with and without AI: automation bias in pre-clerkship medical students

Abstract

Background

As large language models become increasingly integrated into clinical workflows, medical students need structured opportunities to learn how to engage critically with artificial intelligence (AI) during diagnostic reasoning. Empirical evaluation of automation bias and other risks of AI use in pre-clerkship training remains limited.

Methods

We piloted a two-component exercise for second-year pre-clerkship medical students: an introductory lecture on AI capabilities and limitations followed by a custom-built web application integrating an AI chatbot into a diagnostic reasoning case in which students ranked their differential diagnoses before and after AI access and rated the perceived influence of AI on their reasoning. Diagnostic accuracy was scored against predefined criteria. Descriptive and inferential statistics were calculated.

Results

In a sample of 185 students, AI use was associated with increased diagnostic accuracy (Wilcoxon signed-rank Z = -4.21, P < .001), with the greatest benefit concentrated among students with lower performance before AI use. Both students whose accuracy improved (n = 65) and those whose accuracy worsened (n = 26) after AI use rated AI as more influential than students whose accuracy did not change (Z = -2.74, P = .006 and Z = -3.23, P = .001, respectively), suggesting that perceived influence was related to whether AI changed students’ rankings, regardless of whether the change was beneficial or detrimental.

Conclusions

Students benefited most from AI when their baseline diagnostic accuracy was weakest but did not appear to reliably distinguish helpful from harmful AI influence, consistent with automation bias. Foundational coursework on AI and its limitations, paired with AI-integrated case practice and faculty-led debriefing, offers one training approach to address this challenge.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12909-026-09979-x.

Background

Clinical reasoning is a complex cognitive process through which clinicians gather and interpret clinical information, generate and refine diagnostic hypotheses, construct problem representations, and make decisions regarding diagnosis and management [1]. It is a core competency in medical education and is increasingly influenced by artificial intelligence (AI) [2]. As large language models (LLMs) become further integrated into clinical workflows [3], medical trainees who over-rely on AI to perform clinical reasoning face significant risks, including deskilling, never-skilling, and mis-skilling [4].

Automation bias, which refers to the tendency to over-rely on the outputs of automated systems, even when those outputs may be incorrect, has been documented in healthcare through studies of clinical decision support systems that predated the era of generative AI, demonstrating risk of introducing new errors that may go unrecognized despite overall automation-related performance gains [5]. More recent empirical studies have extended these findings to AI-based decision support tools. For example, in a recent prospective study, diagnostic correctness of mammogram interpretations by radiologists, particularly those with limited experience, was found to be significantly impacted by incorrect suggestions of an AI system [6]. In another study on diagnosing a complication of wound healing with AI assistance, less experienced healthcare professionals showed greater susceptibility to incorrect suggestions from AI [7].

Building on this pattern, concerns have been raised that early-career medical trainees may be especially vulnerable to automation bias as they develop their diagnostic reasoning skills [8]. However, empirical studies specifically evaluating automation bias in medical students using LLMs during the formation of their diagnostic reasoning skills remain limited [8]. To address this gap, we conducted a pilot AI integration exercise with medical students enrolled in a pre-clerkship clinical reasoning course to examine whether AI access was associated with changes in diagnostic accuracy and whether students could recognize when AI helped or hindered their performance.

Methods

We employed a single-group pre-post design to evaluate whether AI access was associated with changes in performance on a diagnostic reasoning exercise and students’ perceived influence of AI on their diagnostic accuracy. The pilot exercise was conducted within the Introduction to Clinical Reasoning course at Tufts University School of Medicine as part of a broader institutional AI integration initiative [9].

Participants

All second-year pre-clerkship medical students enrolled in the Introduction to Clinical Reasoning course were eligible to participate (N = 202). The study was determined exempt by the Tufts University Health Sciences Institutional Review Board (ID MODCR-06-11856).

The educational intervention

The exercise comprised two components: an introductory lecture and an AI-integrated clinical reasoning case.

First, students attended a lecture, “Artificial Intelligence in Medicine: A Pre-Clerkship Primer,” delivered to all participants in August 2025. The lecture introduced foundational AI concepts and terminology, underlying technology including neural networks and the transformer architecture, current and emerging clinical applications of LLMs, and key limitations of these technologies, including hallucinations, bias, patient privacy considerations, and a lack of genuine clinical reasoning. The lecture also addressed the use of AI in medical education and explicitly introduced the risk of automation bias.

Second, we developed a custom-built web application (Figure S1) that integrated an AI chatbot into a three-part acute kidney injury (AKI) case requiring iterative Bayesian reasoning. The case was designed and authored by one of the study team members (RH), a board-certified attending nephrologist, and peer-reviewed by two other study team members (SE and RT), co-directors of the course and internal medicine-trained physicians (full case text available in the supplementary materials). As new clinical data emerged at each stage, the relative likelihood of competing diagnoses shifted, requiring students to re-rank their differential, with acute interstitial nephritis emerging as the most likely diagnosis by the final part. Within each part of the case, students first ranked their differential diagnoses without AI access (pre-AI phase); the integrated AI chatbot then became available for students to refine their reasoning and adjust their rankings on a voluntary basis (post-AI phase). Students could submit up to 10 prompts across the whole case. Students completed the case independently online and then attended an in-class session in November 2025 to review their AI interactions with classmates and faculty.

AI chatbot configuration

The chatbot was powered by GPT-4o (model version gpt-4o-2024-08-06, accessed via the OpenAI application programming interface), which was the most capable model available at implementation in August 2025; default model parameters were used.

The chatbot was configured through a prompt template, instructing the model to act as an expert assisting a medical student, limiting responses to three sentences to encourage focused, interpretable outputs, and supplying prior dialogue from the same case part as context to enable continuity across multiple prompts.

The full prompt template was as follows, with dynamic content shown in brackets: “Your role is ‘assistant’, an expert helping a medical student (‘user’). Here is the question from the user: [user question]. Limit response to 3 sentences max. For reference, here is an array of questions and responses from your prior dialogue: [prior dialogue]. If there are no questions or responses, this is the first question of your dialogue.”

The chatbot did not have automatic access to the clinical case data; students were required to select and include relevant clinical information in their prompts, mirroring the real-world challenge of crafting effective but appropriately de-identified queries.

Throughout this manuscript, ‘AI chatbot’ refers specifically to this GPT-4o-powered interface; ‘AI’ is used as shorthand for the chatbot when describing the exercise and more broadly when referring to the field.

Outcome measures

The primary outcome was diagnostic accuracy. Students ranked differential diagnoses in order of likelihood, with rank 1 indicating the most likely diagnosis, from a pre-determined list of five diagnoses; students had the opportunity to update these rankings pre- and post-AI at each part of the case (Figure S1). For each part of the case, these rankings were scored against predefined criteria (Fig. 1). Each criterion checked whether a diagnosis was ranked at a particular position, within a particular range, or as more likely than another specific diagnosis. All diagnostic accuracy criteria were weighted the same; each was scored 0 or 1 and summed into a composite score for each part, and per-part scores were summed into an overall composite score across the full case. Scoring was based solely on students’ diagnostic rankings; optional written comments provided by students were not included in these diagnostic accuracy outcome measures.

Fig. 1.

Diagnostic accuracy scoring criteria for each part of the case. Differential diagnoses were ranked by likelihood (rank 1 = most likely)

The secondary outcome was students’ perceived influence of AI on their reasoning. Immediately after the case, students responded to the rating question “How much did the AI chat influence your differential diagnosis rankings?” on a 5-point scale (1 = least, 5 = most).

Statistical analysis

Wilcoxon signed-rank tests were used to examine pre- to post-AI changes in composite score, with effect sizes reported as r. McNemar tests were used to examine changes in individual diagnostic accuracy criteria from pre- to post-AI, with effect sizes reported as Cohen’s g.

To assess whether score changes from pre- to post-AI were related to baseline performance, and specifically whether gains were concentrated among certain learners, students were grouped by their pre-AI composite score within each part of the case. Because the maximum composite score varied across case parts (0–2 for Part 1, 0–3 for Part 2, and 0–4 for Part 3), the number of baseline groups also varied (3, 4, and 5 groups, respectively). Score changes were calculated as the post-AI score minus the pre-AI score, with positive values indicating improvement, negative values indicating worsening, and zero indicating no change. Kruskal-Wallis tests were used to compare change scores across baseline score groups. Spearman rank correlation (ρ) assessed the relationship between overall pre-AI composite score and the score change following AI use.

To examine whether students’ perceived AI influence at the end of the exercise was associated with pre- to post-AI changes in their diagnostic accuracy, students were classified into three groups based on the direction of change in their overall composite score (summed post-AI score across Parts 1–3 minus summed pre-AI score across Parts 1–3): improved, worsened, or no change. Differences across these groups were tested with the Kruskal-Wallis test, with Mann-Whitney U tests for pairwise comparisons. Effect sizes (r) were calculated for Mann-Whitney U tests.

Statistical significance was set at P < .05. All analyses were performed in IBM SPSS (Version 31.0).

Post-hoc review of student prompts and AI responses

Following the primary analyses, the case author (RH) conducted a post-hoc qualitative review of student prompts and AI responses. Two subsets of prompts and AI responses were reviewed to explore preliminary patterns: prompts and corresponding AI responses for students whose ranking of acute interstitial nephritis moved from incorrect to correct following AI use in Part 3, as well as prompts and corresponding AI responses for students whose ranking of glomerulonephritis moved from correct to incorrect following AI use in any case part. This review was exploratory and was not part of the pre-specified analysis plan.

Results

Participants and engagement with AI

Of 202 students who participated in the exercise, 185 who had completed the case were included in the analysis. Students who opted out of sharing their data for study purposes (n = 6), who did not submit any case data (n = 4), or who submitted partial data but did not complete the case (n = 7) were excluded from the data analysis.

Students’ engagement with the AI chatbot and revision of differential rankings were high overall but declined across the three case parts (Figure S2). The proportion of students submitting at least one prompt fell from 92% (170 students) in Part 1 to 76% (140) in Part 2 and 69% (127) in Part 3, and the proportion who revised their differential rankings after AI access declined similarly, from 72% (133 students) to 50% (92) to 34% (63) in Parts 1–3, respectively.

Students submitted 884 prompts in total (515 in Part 1, 192 in Part 2, and 177 in Part 3), for a mean of 4.77 prompts per student by the end of the case with a standard deviation of 2.52. No student reached the 10-prompt limit; the maximum submitted by any single student was nine.

Diagnostic accuracy improved after AI use, but the extent depended on baseline performance

As shown in Table 1, across the full case, 65 students (35%) improved their overall composite diagnostic score following AI use, whereas 26 students (14%) declined and 94 (51%) showed no change (Wilcoxon signed-rank Z = -4.21, P < .001). Composite scores improved significantly within each case part as well (all P < .05).

Table 1.

Composite diagnostic reasoning scores pre- vs. post-AI by case part (N = 185)

Case Part	Composite Score Change
	Improved, No. (%)	Worsened, No. (%)	No Change, No. (%)	Wilcoxon Z	P value	Effect Size (r)b
Part 1 (score: 0–2)	37 (20)	19 (10)	129 (70)	-2.56	.010a	0.19
Part 2 (score: 0–3)	35 (19)	13 (7)	137 (74)	-2.92	.004a	0.21
Part 3 (score: 0–4)	31 (17)	16 (9)	138 (74)	-1.98	.048a	0.15
Overall (score: 0–9)	65 (35)	26 (14)	94 (51)	-4.21	< .001a	0.31

^a Statistically significant (P < .05; Wilcoxon signed-rank test)

^b Effect size (r): small = 0.10, medium = 0.30, large = 0.50

At the level of individual diagnostic accuracy criteria, three of the nine criteria assessed improved significantly after AI use. In Part 1, the proportion of students correctly ranking pre-renal azotemia as more likely than acute tubular necrosis increased from 62% to 71% (P = .002). In Part 2, the proportion correctly ranking acute tubular necrosis as more likely than pre-renal azotemia increased from 45% to 55% (P = .001). In Part 3, the proportion correctly ranking acute interstitial nephritis, the correct final diagnosis, as most likely increased from 65% to 81% (P < .001), with 32 students (17%) moving to this diagnosis and 3 (2%) moving away from it. The remaining six criteria did not change significantly (Table 2).

Table 2.

Diagnostic criteria performance pre- vs. post-AI by case part (N = 185)

Diagnostic Criterion	Criterion Met Pre-AI, No. (%)	Criterion Met Post-AI, No. (%)	P value	Effect size (Cohen’s g)b
Part 1
Glomerulonephritis ranked 4th or 5th	140 (76)	144 (78)	0.597	0.06
Pre-renal azotemia ranked above acute tubular necrosis (ATN)	114 (62)	132 (71)	.002a	0.28
Part 2
Glomerulonephritis ranked 4th or 5th	135 (73)	132 (71)	0.581	0.12
Post-renal AKI ranked 4th or 5th	92 (50)	100 (54)	0.185	0.14
ATN ranked above pre-renal azotemia	83 (45)	102 (55)	.001a	0.29
Part 3
Glomerulonephritis ranked 4th or 5th	130 (70)	123 (67)	0.092	0.27
Post-renal AKI ranked 4th or 5th	117 (63)	120 (65)	0.549	0.14
Acute interstitial nephritis ranked 1st	121 (65)	150 (81)	< .001a	0.41
ATN ranked 1st or 2nd	92 (50)	84 (45)	0.077	0.25

^a Statistically significant (P < .05; McNemar test)

^b Effect size (Cohen’s g): 0.05 = small, 0.15 = medium, 0.25 = large

Score changes differed across baseline performance groups within each case part (Kruskal-Wallis H = 42.74, 20.07, and 36.53 for Parts 1–3, respectively; all P < .001). Full pre-AI to post-AI score transition matrices for each case part are provided in Supplementary Table S1. Baseline composite score was negatively correlated with post-AI score change (Spearman ρ = -0.306, P < .001), indicating that gains were concentrated among students with lower initial diagnostic accuracy. Students who entered the case with higher baseline accuracy generally maintained their performance.

Students did not reliably distinguish beneficial from detrimental AI influence

We compared perceived AI influence on diagnostic reasoning across the three performance groups: students whose overall composite score improved (n = 65), worsened (n = 26), or did not change (n = 94) following AI use.

As shown in Table 3, perceived influence differed significantly across performance groups (Kruskal-Wallis H = 13.61, P < .001). Pairwise comparisons using Mann-Whitney U tests showed that both students whose diagnostic accuracy improved after AI use and those whose accuracy worsened rated AI as significantly more influential than students whose diagnostic rankings did not change (Z = -2.74, P = .006 and Z = -3.23, P = .001, respectively). However, there was no significant difference in perceived AI influence between improved and worsened groups (Z = -1.02, P = .308).

Table 3.

Perceived AI influence on diagnostic reasoning by performance outcome (N = 185)

Performance Outcome	No.	Mean Rank	Pairwise Comparison	Z	P value	Effect Size (r)b
Improved diagnosis	65	102.95	vs. No Change	-2.74	.006a	0.22
No change in diagnosis	94	79.80	vs. Worsened	-3.23	.001a	0.30
Worsened diagnosis	26	115.87	vs. Improved	-1.02	0.308	0.11

^a Statistically significant (P < .05; Mann-Whitney U test)

^b Effect size (r): small = 0.10, medium = 0.30, large = 0.50

Post-hoc review of student prompts and AI responses

The largest statistically significant criterion-level improvement after AI use was the correct ranking of acute interstitial nephritis, the most likely final diagnosis, which increased from 65% to 81% in Part 3 (P < .001, Cohen’s g = 0.41; Table 2). Of the 31 students whose composite score improved in Part 3 (Table 1), 27 (87%) did so specifically by correctly ranking acute interstitial nephritis as the most likely diagnosis in the post-AI phase after initially ranking it lower pre-AI. We conducted a post-hoc review of all AI prompts and responses from Part 3 for these 27 students.

In all 27 instances, the prompt and AI response identified the connection between acute interstitial nephritis and a white blood cell cast, a key clinical finding that strongly suggests acute interstitial nephritis over other etiologies and that is a new finding in Part 3 of the case. In 26 of these 27 instances, the student’s prompt specifically included “white blood cell cast,” and the AI response specifically named “acute interstitial nephritis,” as illustrated in the following representative example:

Student prompt

“Now the urine sediment shows 30 white blood cells /hpf, and a white blood cell cast is identified. The urine culture returns negative. With this new information, what is the likely diagnosis?”

AI response

“The presence of 30 white blood cells per high power field and a white blood cell cast in the urine sediment suggests acute interstitial nephritis (AIN), especially considering the negative urine culture. The recent history of antibiotic use, specifically cephalexin, aligns with this diagnosis, as medications are common triggers for AIN.”

In the remaining instance, the same association between white blood cell casts and acute interstitial nephritis appeared in the AI response, but the logic was reversed: the student asked about urinalysis findings for acute interstitial nephritis in general rather than about how to interpret a white blood cell cast:

Student prompt

“what is the urine analysis in acute interstitial nephritis”.

AI response

“In acute interstitial nephritis, urinalysis often shows white blood cell casts, sterile pyuria (presence of white blood cells in urine without bacteria), and possibly eosinophiluria (eosinophils in the urine). There may also be mild proteinuria and hematuria. These findings indicate inflammation in the renal interstitium, commonly due to drug-induced hypersensitivity reactions or infections.”

We also reviewed prompts and AI responses among students whose diagnostic accuracy worsened after AI use. Glomerulonephritis was selected as the focus of this post-hoc review because it is the only diagnosis with the same scoring criterion consistently evaluated in all three case parts (Fig. 1). Throughout the case, the absence of hematuria on the initial urinalysis was the key clinical finding that made glomerulonephritis unlikely, as hematuria is a highly sensitive finding for glomerulonephritis such that its absence argues strongly against the diagnosis. The absence of hematuria thus makes the ranking of glomerulonephritis a useful lens for examining how students handled pertinent negatives when prompting the AI chatbot.

Across the case, 44 unique students had at least one part where their pre- to post-AI composite score worsened (48 total instances of post-AI worsening in diagnostic accuracy across all parts, but 4 students worsened in more than one part). Of these 44, 24 (55%) had a pre-AI ranking that correctly placed glomerulonephritis at 4th or 5th but then shifted to an incorrect post-AI ranking (14 in Part 1, 4 in Part 2, and 6 in Part 3). In 22 of these 24 instances, the AI response specifically mentioned glomerulonephritis as a possible diagnosis, but in 20 of these same 24 instances, the student’s prompt did not specifically mention the absence of hematuria, the key pertinent negative for excluding glomerulonephritis in the case. In 15 of these 20 prompts omitting the absence of hematuria, the student specifically mentioned other urine findings, such as proteinuria. A representative example is shown below from a student who inquired about proteinuria but did not mention the absence of hematuria:

Student prompt

“which causes of acute kidney injury have elevated protein in urine”.

AI response

“Causes of acute kidney injury (AKI) that can result in elevated protein in the urine include glomerulonephritis, acute tubular necrosis with secondary tubular proteinuria, and conditions leading to nephrotic syndrome like minimal change disease. These conditions involve damage to the glomeruli or tubules, leading to protein leakage into the urine. Proper diagnostic evaluation, including urine analysis and kidney function tests, is necessary to determine the underlying cause of AKI.”

Among these 20 instances where the AI response mentioned glomerulonephritis but the student’s prompt omitted the absence of hematuria, only 4 AI responses noted that hematuria would typically be expected in glomerulonephritis; the remaining 16 did not raise hematuria at all.

Discussion

In this pilot AI integration exercise, pre-clerkship medical students completed a three-part iterative diagnostic reasoning case, ranking differential diagnoses before and after optional access to an AI chatbot in each part. Students’ engagement with AI was high in this exercise, and our analysis revealed two main findings. First, AI access was associated with an improvement in students’ diagnostic accuracy, with the largest improvement centered on the correct final diagnosis, acute interstitial nephritis, and with gains concentrated among students who entered the case with lower baseline accuracy. Second, students did not appear to distinguish beneficial from detrimental AI influence. Those whose accuracy improved after AI use and those whose accuracy worsened both rated AI as more influential than students whose rankings did not change. Together, these findings suggest that AI can support early learners in a diagnostic reasoning exercise while also raising concern for automation bias.

The finding that gains after AI use were concentrated among students with lower initial diagnostic accuracy, while those with higher baseline accuracy largely maintained their performance, is consistent with a ceiling effect. One possible interpretation is that AI helped lower-baseline students reweight the relative likelihood of competing diagnoses while higher-baseline students had already made such a judgment on their own. This pattern echoes prior empirical work in which less-experienced medical professionals were most helped by correct AI suggestions as well as most vulnerable to incorrect ones [6, 7]. Our findings extend that observation to pre-clerkship medical students using a general-purpose LLM during a diagnostic reasoning exercise.

Our post-hoc review of student prompts and AI responses explored how the content of students’ prompts may have shaped whether AI use improved or worsened diagnostic accuracy. Among students whose ranking of acute interstitial nephritis moved from incorrect to correct after AI use in Part 3, prompts and AI responses consistently engaged the key pertinent positive finding, the white blood cell cast, suggesting that AI was able to effectively support diagnostic accuracy when the prompt included the appropriate clinical information. Conversely, among students whose accuracy worsened from pre- to post-AI, the predominant pattern was an error of omission: over half (24 of 44, 55%) of students whose composite score worsened after AI use did so specifically by shifting glomerulonephritis from a correct pre-AI ranking to an incorrect post-AI ranking, and in the large majority of these instances (20 of 24, 83%), the student’s prompt did not include the absence of hematuria, the key pertinent negative for excluding glomerulonephritis, as hematuria is expected in nearly all cases of glomerulonephritis such that its absence argues strongly against the diagnosis. In most of these instances (15 of 20, 75%), the student’s prompt did include other urine findings, such as proteinuria, which can also be present in glomerulonephritis; however, the AI chatbot responded to these findings without the context of the absent hematuria, which still would have made glomerulonephritis unlikely even in the presence of proteinuria. In only a minority of these instances (4 of 20, 20%) did the AI response note that hematuria would typically be expected in glomerulonephritis; in the remaining instances, the AI response did not mention hematuria at all. These observations highlight two distinct challenges. First, students may fail to include the specific clinical findings needed to differentiate between diagnoses, particularly key pertinent negatives, in their prompts. Second, the LLM used for the AI chatbot in this exercise (GPT-4o) did not consistently identify what was missing from the student’s prompt, potentially reinforcing rather than correcting incomplete prompts. Whether this behavior is model-specific or a broader limitation of LLMs at the time of implementation remains an open question. For educators, the first observation raises the importance of future work on prompt construction as a component of AI-integrated exercises like this one; the second may evolve as LLM capabilities advance.

Our second main finding concerned students’ perception of AI’s influence. Students’ perceptions appeared to reflect whether AI use was tied to a change in diagnostic rankings rather than whether that change was beneficial or detrimental. This dissociation between recognizing AI’s influence and judging its accuracy is a pattern associated with automation bias [5], and our findings provide preliminary educational evidence compatible with this concern. Notably, the students had been introduced to automation bias as a concept in the lecture preceding the case. Conceptual awareness alone may not be sufficient for students to detect automation bias in their own reasoning. A recent randomized controlled trial in fourth-year medical students found that a safety disclaimer about ChatGPT’s potential errors did not alter students’ use of AI advice [10], suggesting that a warning alone may not be enough to help students judge AI’s accuracy in clinical training.

Empirical evaluation of AI-induced reasoning shifts in medical trainees has been limited [8, 11]. In a recent study of 40 first-year medical students using ChatGPT in team-based clinical reasoning, students aligned their answers with AI suggestions whether or not the suggestions were correct, and subsequent peer discussion did not correct the AI-induced errors [12]. Our study adds to this limited evidence base with a larger cohort of second-year pre-clerkship students engaged in an iterative diagnostic reasoning case across multiple stages. To our knowledge, few studies have examined whether medical students can distinguish beneficial from detrimental AI influence on their diagnostic accuracy by pairing objective performance data with self-rated AI influence in the same exercise.

Although our findings arise from a single educational exercise focused on diagnostic accuracy and should not be generalized directly to clinical environments, the broader concern that uncritical AI use during training could affect downstream clinical reasoning and patient safety has been raised independently [4], and our findings are consistent with that concern. For medical students at risk of automation bias early in training, foundational coursework on AI and its limitations should be paired with applied exercises and structured debriefing with expert faculty to help students recognize when AI helps and when it hinders their performance. This curricular approach is feasible across institutions and clinical domains.

Our study has several limitations. Our primary outcome of diagnostic accuracy reflects only one component of the complex process of clinical reasoning, limiting the generalizability of our findings to clinical reasoning more broadly. The exercise was conducted within a single pre-clerkship clinical reasoning course at a single institution, used a single clinical case focused on acute kidney injury, and relied on a single LLM (GPT-4o), the most capable model available at implementation. The case was unique to this exercise, and its scoring criteria have not been standardized or validated outside of this study; additionally, all diagnostic accuracy criteria were weighted equally, which may not fully capture the relative clinical importance of individual diagnostic distinctions. Perceived AI influence was assessed with a single rating item administered once at the end of the case rather than after each part. Our pre-post design lacked a control group without AI access, limiting causal inference about the source of observed changes in diagnostic accuracy. Other factors, such as repeated engagement with the case material or learning during the exercise, may also have contributed to the observed changes, and the design does not allow these to be distinguished from the effect of AI use. AI use was voluntary within each case part, and students who chose to engage with the AI chatbot may have differed systematically from those who did not. Student engagement with the AI chatbot declined across case parts, potentially limiting comparisons in later parts of the case. The intervention was implemented over a relatively long period of time, with up to 12 weeks between the foundational lecture and the AI-integrated case, and the relative contribution of each component (lecture, applied exercise, and faculty-led debriefing) was not directly evaluated. We did not apply a multiple-comparisons correction across the pairwise tests. Finally, the prompt review was manually conducted as a post-hoc exploratory analysis by a single member of the study team and was not part of the pre-specified analysis plan; its findings should be interpreted as hypothesis-generating rather than confirmatory.

Our findings raise several directions for future research. First, similar studies across additional LLMs, clinical domains, and learner stages will be important for generalizability. Second, future work should directly evaluate the effect of faculty-led debriefing through randomized comparison of curricular approaches with and without structured debriefing. Finally, a more comprehensive analysis of students’ AI interactions, including prompts submitted and AI responses across all parts of the case, may help clarify the mechanisms by which AI improved or worsened diagnostic accuracy and inform more targeted instruction in prompt literacy.

Conclusions

As LLMs become increasingly integrated into clinical workflows and educational environments, educators must adapt medical school curricula for this new era. The challenge is not whether to incorporate AI into medical education but how to do so in a way that builds learners’ understanding of the technology and its limitations while strengthening, rather than bypassing, the development of independent critical thinking in medicine [4]. Our findings suggest that AI can support pre-clerkship learners in a diagnostic reasoning exercise while also raising concern for automation bias, underscoring the need for educational approaches that develop both AI literacy and the critical judgment to evaluate AI outputs. Foundational coursework on AI and its limitations, paired with applied exercises and structured debriefing with expert faculty, offers one such curricular approach.

Abbreviations

AI

Artificial intelligence

AKI

Acute kidney injury

ATN

Acute tubular necrosis

LLM

Large language model

Authors’ contributions

RH, SE, and RT conceived and designed the study. RH delivered the foundational lecture, authored the acute kidney injury case, and developed the custom-built web application. SE and RT contributed to the case design and the final web application. MB, SE, and RT administered the pilot exercise and oversaw data collection. SR performed the statistical analysis. RH, MB, and SR interpreted the data. RH wrote the first draft of the manuscript. MB and SR contributed substantially to the writing. All authors contributed to manuscript revision and read and approved the final manuscript.

Funding

Not applicable.

Data availability

Data supporting the conclusions of this article are included within the article and supplementary materials. The dataset is available from the corresponding author, RH, upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was determined exempt by the Tufts University Health Sciences Institutional Review Board under protocol ID MODCR-06-11856. Research participation was voluntary, and all participants included in the analysis provided informed consent in accordance with institutional policy. All procedures performed in studies involving human participants were in accordance with the ethical standards of the Tufts University Health Sciences Institutional Review Board and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.