This study was conducted at Singapore General Hospital under the SingHealth Anesthesiology Residency Program. It was reviewed by the Singapore General Hospital Research Office and was granted exemption from formal institutional review board review, as it involved routine educational activities and the use of anonymized resident data. Participation was voluntary, and no identifiable personal data were collected. Completion of the TAM survey was considered implied consent. All data were deidentified prior to analysis, and no compensation was provided to the participants.

Participants were recruited in a systematic cohort-based approach within the SingHealth Anesthesiology Residency Program at Singapore General Hospital. All anesthesiology residents actively preparing for an upcoming Master of Medicine (MMed) anesthesiology examination were scheduled to complete a routine, in-person mock short answer question (SAQ) assessment and were invited to participate in the study. Eligibility required active enrollment in the residency program with at least 32 months of anesthesia-related training experience, consistent with the criteria for sitting in the MMed examination. No additional exclusion criteria were applied, as the cohort was already homogeneous in the training level and examination eligibility. At the start of the session, a member of the program faculty provided a standardized study briefing, explaining its objectives, procedures, and voluntary nature of participation. Residents were informed that participation would not influence their training evaluation or examination outcomes. Implied consent was obtained following this briefing. A total of 21 residents, representing the full cohort of candidates preparing for that year’s MMed examination, consented and participated in the study. The cohort comprised residents in years 2 to 4 of training, with a gender distribution of 14 females and 7 males.

PASS was developed using OpenAI’s GPT-4 model and customized via the ChatGPT Custom GPT interface to provide grading and detailed feedback for SAQ-style responses. Two study investigators (LYH and KY) conducted 3 iterative rounds of pilot testing by using simulated essay responses modeled after previous MMed anesthesiology examination questions. Each round involved a systematic review of PASS outputs by both investigators and 2 faculty anesthesiologists to evaluate the clinical accuracy, feedback clarity, and grading consistency. Refinements across iterations focused on clarifying role-based prompts, improving alignment with examiner expectations, and reducing hallucinated or ambiguous statements in the generated feedback. After each testing round, the prompt was revised to enhance the specificity of the instructions and the standardization of the grading language. The final prompt version, presented in Table S1 in Multimedia Appendix 1, reflected cumulative adjustments derived from these pilot evaluations and represented the configuration used in the study.

The reference materials in PASS included the 2024 MMed anesthesiology syllabus (PDF, 32 pages), examiner feedback reports (n=5, 1-2 pages each), past SAQ papers (2018-2024), and standard textbooks such as Miller’s Anesthesia (9th edition) and Yao & Artusio’s Anesthesiology: Problem-Oriented Patient Management (9th edition).

Prompt engineering followed the principles outlined by Meskó [15], focusing on specificity and contextualization. Role-playing prompts were used to instruct the LLM to adopt the perspective of an anesthesiologist examiner, ensuring clinically accurate and assessment-calibrated evaluations.

Participants completed a closed-book mock examination of 4 SAQs. These were adapted from prior internal mock examinations aligned with recent MMed anesthesiology examination themes and reviewed by another anesthesiology faculty to ensure content validity. Examiner 1 developed a standardized mark scheme and graded all 84 scripts as per standard assessment practices. Each SAQ constituted 1 script, with 21 participants producing 84 total responses.

Additionally, 2 independent human examiners (examiners 2 and 3) also graded all the scripts based on the standardized mark scheme. Examiners 2 and 3 independently evaluated the PASS grading iterations alongside examiner 1. Two study investigators generated 3 separate PASS grading iterations per script in parallel chat windows. PASS was not provided with the official grading rubric and relied solely on its internal knowledge base. SAQ responses were graded on a 0-8 scale and stratified into poor (0-4.5), average (5-5.5), and good (6-8) categories. The stratified gradings were then analyzed for grading IRR, with examiner 1’s scores serving as the reference standard when required.

Participants subsequently evaluated both PASS and standard GPT-4.0 feedback by using a customized TAM survey [16]. TAM is grounded in the theory of reasoned action, which posits that an individual’s behavioral intention to use a technology is shaped by their attitudes toward the technology and by perceived usefulness. This framework has been widely applied to understand user acceptance of new technologies, including those used in medical education [14,17,18]. Four dimensions were rated on a 5-point Likert scale (1=least agreement, 5=strongest agreement): (1) usefulness for examination preparation, (2) efficiency in study preparation, (3) ease of use, and (4) likelihood of future use. Participation in the survey was voluntary and anonymous. Incomplete surveys were excluded from the analysis.

PASS responses and feedback to participant essays were reviewed independently by study investigators and participants to detect hallucinations and deviations. Hallucinations were defined as “any incorrect information or statements that may cause moderate to major patient harm,” while deviations were defined as “variations from usual clinical practice unlikely to result in patient harm.” The frequency of hallucinations and deviations detection was compared between groups.

All statistical evaluations were performed in the Excel and Python 3.8 environment. The TAM scores for PASS versus GPT-4 were compared using paired 2-sided t tests. Fleiss κ was used to calculate IRR within PASS and human examiner groups, and agreement between individual examiners was evaluated using Cohen κ analysis. Chi-square tests were used to compare hallucination and deviation detection between examiners and candidates. As this was a proof-of-concept feasibility study, no a priori power calculation was performed.

This paper adheres to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guideline [19], and a completed STROBE checklist has been included in Multimedia Appendix 2.

Of the 21 participants, 17 (81%) completed the TAM survey, giving a total of 136 survey responses. Four participants did not complete the TAM survey, and no data imputation was performed. User perceptions of PASS were positive. Mean ratings for PASS were as follows: usefulness (4.25, SD 0.50), efficiency (4.12, SD 0.61), ease of use (4.56, SD 0.63), and likelihood of future use (4.13, SD 0.75). Paired 2-sided t tests comparing PASS to standard GPT-4 revealed statistically significant differences in 3 of the 4 domains. PASS was rated significantly higher than GPT-4 in usefulness (mean difference=0.81, 95% CI 0.66-0.96; P<.001), efficiency (mean difference=0.71, 95% CI 0.53-0.88; P<.001), and likelihood of future use (mean difference=0.54, 95% CI 0.39-0.70; P<.001). There was no significant difference between PASS and GPT-4 in ease of use (mean difference=0.06, 95% CI –0.07 to 0.18; P=.35). A summary of the TAM results can be found in Table 1 and Figure 1.

A total of 504 scripts were marked by 6 examiners (3 humans, 3 PASS). PASS examiners demonstrated moderate internal IRR (κ=0.522), whereas human examiners exhibited fair internal IRR (κ=0.275), based on the benchmarks of Landis and Koch [20]. Individually, the agreement of PASS and human examiners with examiner 1 were as follows: PASS 1 (κ=0.237), PASS 2 (κ=0.193), PASS 3 (κ=0.140), examiner 2 (κ=0.149), and examiner 3 (κ=0.470). Among these, examiner 3 demonstrated the highest agreement with examiner 1, while PASS 3 exhibited the lowest. The complete interexaminer agreement matrix can be found in Figure 2. When analyzed as a group (Table 2), PASS examiners approached a moderate level of agreement with the combined group of human examiners (examiners 1, 2, and 3) (κ=0.357, 95% CI 0.156-0.558). Pairwise agreement between PASS and individual human examiners ranged from κ=0.159 (PASS vs examiner 1, slight) to κ=0.429 (PASS vs examiner 3, moderate).

A total of 316 PASS-generated responses were evaluated for hallucinations and deviations. Examiners conducted 78.8% (249/316) of the evaluations, while candidates accounted for 21.2% (67/316). Across all responses, 67 hallucinations and 189 deviations were labeled. Each hallucination or deviation was labeled by independent reviewers. No statistically significant difference was observed in the hallucination detection rates between candidates and examiners (P=.21). Candidates identified 10 hallucinations across 67 responses (15%), while examiners detected 57 hallucinations across 249 responses (22.9%). In contrast, examiners identified significantly more deviations than candidates. Candidates detected 21 deviations (31.3%), whereas examiners identified 168 deviations (67.5%)—a difference that was highly significant (P<.001) (Table 3).

This proof-of-concept study demonstrates strong end user acceptance of a specialized LLM, supporting the potential role of tailored AI tools in high-stakes medical examination preparation. Younger learners increasingly prefer technology-driven educational tools that provide prompt feedback, personalized learning experiences, and interactive engagement [21,22]. Specialized LLM platforms such as our case example PASS offers an innovative approach to meet these expectations.

Our findings are consistent with prior research showing that LLM-based educational tools are generally well-received across educational contexts [14,23,24]. However, user acceptance specifically for high-stakes examination preparation, where performance directly impacts career progression, remains underexplored [25]. Given that user acceptance is a key factor in the successful adoption of new technologies [26], our results add evidence that specialized LLMs can be successfully incorporated as study aids for postgraduate medical examinations.

Notably, PASS, a specialized version of GPT-4 customized with domain-specific content, received significantly higher ratings in usefulness, efficiency, and likelihood of future use compared to standard GPT-4. This supports the argument that domain-specialized enhancements improve both learner experience and perceived educational value [27].

In our study, PASS demonstrated a higher internal grading IRR (κ=0.522) compared to human examiners (κ=0.275), suggesting that LLM-based grading may provide more standardized evaluations. When analyzed as a group, PASS examiners showed statistically significant agreement with human examiners (κ=0.357), indicating that PASS grading trends align with human assessment.

Timely and accurate grading feedback is crucial for effective examination preparation, as it helps candidates refine their responses and align their answering techniques with examiner expectations [28]. However, faculty availability is often limited, restricting or delaying access to grading opportunities. Additionally, human grading is inherently variable, influenced by factors such as individual biases and grading fatigue [29]. Our results suggest that specialized LLMs could offer scalable, consistent supplemental grading and feedback in postgraduate education, potentially alleviating faculty workload and enabling self-directed formative assessment.

However, PASS’s grading agreement with humans remained only moderate, indicating that its reliability may not yet be sufficient for formal grading. This finding aligns with previous research highlighting the limitations of LLMs as formal grading tools [30]. Despite these constraints, specialized LLMs remain valuable as self-assessment tools, allowing students to estimate their performance during examination preparation.

PASS was intentionally not provided with the official marking rubric to reflect the realistic conditions of learner-driven use. This design choice likely contributed to moderate agreement levels, as rubric-guided calibration could enhance consistency between AI and human grading. Future studies may investigate whether standardized grading frameworks improve IRR and alignment with human examiners.

Hallucinations in LLM generated feedback are a well-documented concern [1]. In medical education, this issue is particularly critical, as even minor misconceptions can lead to significant errors in clinical management, potentially resulting in adverse patient outcomes [31]. While previous studies have discussed potential ways to mitigate the effect of hallucinations [32], there is a notable gap in research comparing the detection rate of hallucinations between tutors and learners.

Encouragingly, our study shows that candidates were able to identify hallucinations at a rate comparable to examiners. The ability of candidates to recognize and filter out major erroneous feedback on their own suggests that specialized LLMs show potential for use under supervision in examination preparation. However, it is important to note that all candidates in our study had at least 32 months of anesthesia experience, which may have contributed to their ability to detect errors effectively. Less experienced learners may not demonstrate the same level of discernment and could require closer supervision when using AI-generated feedback.

Additionally, our results show that candidates were significantly less likely than examiners to detect deviations from standard clinical practice. Failure to recognize clinically relevant, even if minor, deviations from standard practice may inadvertently reinforce incorrect reasoning patterns. If undetected, such deviations could perpetuate unsafe habits or incomplete understanding, which, in a clinical context, might affect patient management. Incorporating structured faculty review or critical appraisal exercises into curricula could help learners identify and mitigate such errors when using AI feedback.

Our findings suggest important considerations for future use of LLMs in education. Educators must assess whether learners possess the critical skills to evaluate AI-generated content before incorporating LLM feedback into curricula. Future research should explore hallucination detection across learner levels and strategies to build critical evaluation skills alongside AI-assisted learning.

This study has several limitations. First, the small, single-center cohort of 21 participants may limit the generalizability of our findings. All participants were experienced anesthesiology residents from the same training program and were approaching their final examinations. Consequently, the results may not extend to less experienced trainees, learners from other specialties, or different educational settings. Junior trainees may have greater difficulty discerning hallucinations or subtle deviations in AI-generated feedback, underscoring the continued importance of supervised use and faculty guidance. Future multicenter studies involving a broader range of specialties, experience levels, and training environments will be essential to evaluate the scalability and educational impact of specialized LLMs. Second, no formal power calculation was conducted, given the exploratory, proof-of-concept nature of the study. In addition, intraexaminer reliability was not assessed, which may have influenced the variability in human grading. Finally, this study did not examine the direct effect of using PASS on actual examination performance or learning outcomes. Assessing these educational and behavioral impacts should be a key focus of subsequent research to determine the long-term value and safety of integrating specialized LLMs into postgraduate medical education.

Building on these preliminary findings, future research should focus on multicenter validation of specialized LLMs across different institutions, specialties, and learner experience levels to better define their generalizability. Further work is also needed to evaluate the longitudinal educational impact of LLM-assisted feedback on examination performance, critical thinking, and clinical reasoning. Incorporating standardized grading rubrics or structured prompt frameworks may improve grading reliability and enhance alignment with human examiners. In parallel, research should explore strategies to train learners in AI literacy and critical appraisal, ensuring they can appropriately interpret and verify model outputs. Ultimately, integrating specialized LLMs into postgraduate education should aim to augment, not replace, human judgment, fostering safe, reflective, and supervised use of AI in medical learning environments.

This study provides proof-of-concept evidence supporting the feasibility and user acceptance of specialized LLMs as supplemental study aids for postgraduate medical examination preparation. The specialized LLM (PASS) demonstrated strong user acceptance, outperforming standard GPT-4 in perceived usefulness, study efficiency, and likelihood of future use. PASS also showed moderate grading consistency and agreement with human examiners, indicating potential value in supporting formative, self-directed learning.

While experienced learners were able to recognize major hallucinations at rates comparable to examiners, their lower detection of subtle deviations underscores the continued necessity of human oversight when using AI-generated educational feedback. These findings highlight the potential of specialized LLMs to enhance postgraduate learning under supervision while reinforcing the importance of ongoing faculty involvement, quality control, and iterative model refinement before broader integration into training programs.