Effective clinical interviewing is essential for making correct diagnoses and building strong relationships with patients [1]. Traditionally, students learn these skills through supervised practice with real or standardized patients and feedback from faculty [1]. However, this apprenticeship-style approach is time-intensive and limits opportunities for deliberate practice [2].

The assessment component itself also consumes substantial faculty and resident physician (RP) time. In competency-based medical education (CBME), faculty complete numerous workplace-based assessment forms; one Canadian study found a mean of 3 minutes, 6 seconds per entrustable professional activity form, adding approximately 18 minutes of extra documentation time for each staff member every 4-week block [3]. Multiprogram qualitative work further confirms that the cumulative “assessment burden” is now viewed as a major threat to sustainability, prompting programs to redesign processes to reduce administrative load [4].

Recently, generative artificial intelligence (AI) using large language models (LLMs) has enabled the creation of AI-based virtual patients that both converse with learners and automatically evaluate performance [2,5]. Empirical studies have shown promising results for AI assessment in free-text clinical documentation [6], script concordance testing [7], and objective structured clinical examination (OSCE) history-taking stations [8]. Many of these systems use validated rubrics such as the Master Interview Rating Scale (MIRS) to structure feedback [9]. Nevertheless, the reliability and validity of AI-generated ratings remain understudied; therefore, establishing concordance with expert evaluations is a prerequisite for educational or licensure use.

This study compared AI-based assessment (ABA) scores of clinical interview performance using GPT-o1 Pro (OpenAI; ABA-o1) and GPT-5 Pro (OpenAI; ABA-5) with human-based assessment (HBA) scores. We hypothesized that ABA scores and HBA scores would exhibit strong concordance and that ABA scoring would serve as a substitute for HBA scoring. We also hypothesized that AI would complete evaluations more rapidly, reducing the assessment burden on clinicians. A secondary aim was to evaluate agreement across participants with differing clinical experience and evaluate whether the use of AI could lead to a measurable reduction in evaluation time, thereby contributing to overall efficiency in assessment processes.

A cross-sectional validation study was conducted. This study involved 3 medical students (MSs), 3 RPs, and 2 attending physicians (APs) who participated in standardized clinical scenarios.

A man aged 27 years presenting with progressive bilateral leg weakness, particularly proximal, was scripted based on a published case of thyrotoxic periodic paralysis. The scenario, created by a general internal medicine specialist with extensive educational experience drawing directly on prior literature, included relevant clinical history (eg, recent myalgias, tremors, diarrhea, and insomnia), red flag cues (eg, acute onset, muscle weakness, and hypokalemia), and psychosocial factors (eg, recent immigration and use of herbal supplements). The case represented a classic presentation of thyrotoxic periodic paralysis caused by hyperthyroidism. The patient was implemented as an AI-simulated character using ChatGPT’s custom generative pretrained transformers.

The participants were recruited through convenience sampling complemented by snowball sampling. MSs comprised a third-year and a fifth-year student; RPs comprised 3 postgraduate year 1 residents; and APs comprised 2 board-certified physicians in internal medicine or general internal medicine in Japan, each with ≥5 years of clinical teaching experience.

Each participant conducted a history-taking encounter by speaking with an AI patient. All conversations were recorded and transcribed verbatim. As all interviews took place within the ChatGPT-based simulated patient interface, transcripts were automatically generated from dialogue logs without manual correction.

The MIRS from the University of Tennessee was originally designed to assess 27 items. In this study, 25 of these items were evaluated based on the available conversational recordings. Each item was rated on a scale from 0 to 5 (total possible score: 0‐125) covering domains such as information gathering, organization, empathy, and patient-centered communication. The excluded items were nonverbal behavior and pace and flow of the interview, which require audiovisual input to evaluate.

The main outcome was the comparison of MIRS scores.

For ABA-o1, each transcript was submitted separately to GPT-o1 Pro with a base prompt directing it to rate the encounter using the MIRS and justify each score. This process was repeated 5 times per transcript, and item-level and total scores were averaged across runs. For ABA-5, using the same base prompt, the 7 transcripts were scored in 2 batch submissions rather than individually. Run 1 included MS 1, MS 2, RP 1, and RP 2, and run 2 included RP 3, AP 1, and AP 2. For each batch, the prompt explicitly stated that it contained 4 interview transcripts (run 1) or 3 interview transcripts (run 2). For each participant within a batch, item-level and total MIRS scores were extracted from the model’s output. The base prompt and model settings were held constant across runs, and the full prompt is provided in Multimedia Appendix 1 to support reproducibility. For both ABA-o1 and ABA-5, reproducibility was further examined by scoring each transcript 5 times using different random seeds.

For HBA, 5 blinded clinical instructors independently rated each transcript using the same MIRS rubric. All assessors were board certified in general internal medicine or general medicine in Japan, actively involved in medical education, and coauthors of this study (Y Tokushima, TS, RW, TM, and FS). Item-level and total scores were averaged across the 5 raters. To further ensure calibration beyond the preparatory webinar, raters briefly discussed scoring rationales for sample transcripts to reach consensus on the interpretation of rubric items.

The secondary outcome was the comparison of assessment time, which was assessed as follows:

For all 3 methods, mean assessment time and SD were calculated, and absolute and relative time savings of ABA vs HBA were reported.

All analyses used R (version 4.3.1; R Foundation for Statistical Computing). Descriptive statistics (mean and SD and coefficient of variation [CV]) summarized the scores. Agreement was assessed using the Pearson correlation coefficient (r) for linear associations; the Lin concordance correlation coefficient (CCC) for both correlation and bias, summarizing overall agreement in a single index; and Bland-Altman analysis for bias and limits of agreement (LoA).

Reliability metrics included the Cronbach α for internal consistency, and intraclass correlation coefficients (ICCs) were calculated to quantify (1) repeatability across the 5 independent GPT-o1 Pro and GPT-5 Pro runs (stability of scores when the same model was applied repeatedly to the same transcript) and (2) interrater reliability across the 5 physician raters (agreement among different human raters). A 2-sided α of <.05 denoted significance.

Ethics approval was obtained from the Juntendo University institutional review board (approval E24-0314-U02). All participants provided written informed consent before taking part. To protect participants’ privacy and confidentiality, all interview transcripts and performance scores were deidentified prior to evaluation and analysis by assigning study IDs and removing any potentially identifying information. Only deidentified transcripts were shared with the physician raters, and results are reported in aggregate. Study data were stored on password-protected, access-restricted institutional systems, and only the research team had access. Participants received no financial compensation for participation.

Table 1 summarizes the interview scores obtained via ABA-o1, ABA-5, and HBA. Across all 7 participants, group-level means were 53.7 (SD 6.8) for HBA, 53.2 (SD 9.2) for ABA-5, and 52.1 (SD 6.9) for ABA-01. Within-participant variability (mean CV percentage) was similar for the 2 automated methods (ABA-o1=6.6%; ABA-5=6.6%) and higher for HBA (13.9%). Individual-level differences were generally small, although notable divergences arose for RP 2 when comparing HBA vs ABA-o1 (46.8 vs 53.4; Δ=6.6) and for AP 2 when comparing ABA-5 vs HBA (67.8 vs 58.8; Δ=9.0) and ABA-5 vs ABA-o1 (67.8 vs 55.6; Δ=12.2).

Agreement and reliability were evaluated across the 3 rating methods (ABA-o1, ABA-5, and HBA). Pairwise concordance with HBA was high for both AI variants: ABA-o1 vs HBA showed a Pearson correlation coefficient (r) of 0.90 (95% CI 0.78‐0.96) and CCC of 0.88; ABA-5 vs HBA showed an r of 0.87 (95% CI 0.72‐0.94) and CCC of 0.86. Concordance between the 2 AI pipelines was the highest (ABA-o1 vs ABA-5: r=0.98, 95% CI 0.95‐0.99; CCC=0.98), indicating near interchangeability of the AI variants (Table 2). Internal consistency followed the same pattern: Cronbach α was 0.81, 0.86, and 0.80 for ABA-o1, ABA-5, and HBA, respectively.

All correlations were significant (2-sided P<.001). Bland-Altman analyses comparing each ABA with HBA showed small positive mean biases (ABA-o1 vs HBA: +0.43 [SD of differences 2.70]; ABA-5 vs HBA: +1.54[SD of differences 5.17]), with 95% LoA of −4.87 to 5.72 and −8.60 to 11.68, respectively; no proportional bias was observed in either comparison (Figures 1A and 1B).

Repeatability was assessed using the ICC. ABA-o1 showed substantial repeatability across 5 independent runs (ICC(3,1)=0.77; ICC(3,5)=0.94), and ABA-5 likewise showed substantial repeatability (ICC(3,1)=0.82; ICC(3,5)=0.96). In contrast, interrater reliability among the 5 HBA physician raters was only fair on single measures (ICC(2,1)=0.38) and improved when averaging them (ICC(2,5)=0.75). Overall, both AI-based approaches yielded more stable ratings across repeated evaluations than HBA, with ABA-5 slightly more stable than ABA-o1.

Table 3 summarizes mean interview scores and SDs by training level. Across methods, APs had the highest means (HBA: 57.6, SD 1.7; ABA-o1: 56.8, SD 1.7; ABA-5: 58.5, SD 13.2). MSs were next (HBA: 56.5, SD 12.0; ABA-o1: 55.0, SD 12.2; ABA-5: 55.3, SD 13.2), in some cases approximating AP performance. RPs had the lowest means (HBA: 49.2, SD 3.6; ABA-o1: 47.1, SD 0.5; ABA-5: 48.3, SD 3.8). Therefore, the anticipated ordinal pattern (APs>RPs>MSs) was not consistently observed as MS means exceeded RP means across all methods.

Total processing time was 5 hours, 59 minutes, 35 seconds for the physician benchmark; 1 hour, 56 minutes, 38 seconds for ABA-5; and 2 hours, 31 minutes, 5 seconds for ABA-o1. Average time per case was 3 minutes, 19.9 seconds for ABA-5 (batch-to-batch SD 1 minute, 6 seconds); 4 minutes, 19 seconds for ABA-o1 (SD 3 minutes, 9 seconds); and 10 minutes, 16.4 seconds for the physicians (SD 11 minutes, 9 seconds). Relative to the physicians, total time was reduced by 67.6% with ABA-5 and 58% with ABA-o1 (Table 4).

In this validation study comparing 3 rater groups (HBA, ABA-o1, and ABA-5), ABA-o1 and ABA-5 produced interview ratings that were statistically indistinguishable from those produced via HBA yet showed markedly superior psychometric stability relative to HBA (Cronbach α: ABA-o1=0.81, ABA-5=0.86, and HBA=0.80; ICC: ABA-o1=0.77, ABA-5=0.82, and HBA=0.38). Cronbach α values of ≥0.8 indicate good internal consistency [10], and ICC(2,1) values of ≥0.75 denote good interrater reliability [11]. Agreement metrics were likewise robust as evaluative tools: the CCC assesses both correlation and bias in a single index [12], whereas Bland-Altman analysis remains the standard for visualizing bias and LoA [13]. ABA-5 was benchmarked against HBA using the same agreement framework.

Although the observed differences in reliability were significant, they may also have practical implications in educational settings. The consistently higher internal consistency and interrater reliability suggest that ABA scoring (including ABA-o1 and ABA-5) could enhance assessment efficiency and reproducibility. Depending on the context, ABA may serve not only as a scalable adjunct but also as a viable alternative to human raters in transcript-based clinical interview evaluations, although this requires significant larger-scale validation.

These findings corroborate previous work in which LLMs matched or exceeded faculty performance when scoring free-text notes [6], designing script concordance tests [7], and evaluating OSCE encounters [8]. A recent study showed that GPT-4o can produce inpatient documentation of comparable quality to that produced by resident physicians while reducing charting time by >50% [14]. In particular, studies of OSCE history-taking and free-text documentation have begun to demonstrate that LLM-based raters can apply communication-focused rubrics in virtual or simulated encounters with performance comparable to that of trained faculty, underscoring the relevance of AI-supported assessment in simulation-based learning contexts [2,6,8]. Consistent with ChatGPT’s passing performance on the US Medical Licensing Examination [15], this study suggests that foundation models possess clinically relevant semantic competence even in spoken communication tasks. Moreover, the 58% reduction in analysis time mirrors the 2025 Time for Class survey, where 36% of faculty who used generative AI daily reported a measurable workload decrease [16].

Beyond efficiency, such time savings could play a decisive role in addressing the growing problem of clinician educator burnout and faculty shortages, which are societal challenges that threaten the sustainability of CBME [17,18]. These concerns mirror prior reports that CBME assessment documentation requires several minutes per form and that the cumulative “assessment burden” is perceived as a threat to program sustainability [3,4]. By automating labor-intensive scoring, AI can free physicians to devote more time to high-value coaching and mentorship, thereby enhancing both educator well-being and learner support [17]. Furthermore, the superior scoring consistency observed with LLMs may help curb rater drift and cognitive biases such as leniency, halo, or contrast effects, which multicenter OSCE analyses have identified as long-recognized sources of unreliability and examiner-related variance in workplace-based assessments [19]. Improved fairness and reliability in assessment would advance equity in trainee progression and, ultimately, foster a more competent, patient-centered workforce.

From an educational perspective, 3 observations are noteworthy when framed across the 3 rater groups (HBA, ABA-o1, and ABA-5).

A key strength is the dual evaluation of accuracy (agreement) and efficiency (time), providing a more complete picture of implementation value than through accuracy alone. Nonetheless, several limitations warrant caution:

First, only 7 participants and a single thyrotoxic periodic paralysis scenario were tested, limiting generalizability across learner levels, languages, and clinical contexts. The small and homogeneous sample also restricts the psychometric interpretation of the findings; for example, differences in learner experience, case complexity, and language environment may influence both human and AI scoring behaviors. Therefore, these results should be viewed as preliminary and hypothesis generating rather than confirmatory. Second, convenience sampling and self-selection may have introduced bias toward technology-friendly participants. Third, model and prompt dependence was a limitation; the results correspond to GPT-o1 Pro and GPT-5 Pro with a specific rubric prompt; other LLMs or prompt engineering strategies could alter performance. Fourth, speech-to-text errors were not exhaustively audited and may have influenced ratings. In addition, the evaluation was limited to transcribed textual data; nonverbal cues, vocal tone, and conversational pauses present in the actual interviews could not be assessed. Fifth, there was potentially a systemic bias. High concordance does not preclude shared cognitive blind spots between AI and human raters; fairness audits across sex, accent, and cultural communication styles remain necessary. In practical implementations, this would entail periodic subgroup analyses of score distributions, qualitative review of discrepant cases, and predefined procedures for pausing or adjusting AI-based scoring if systematic disparities are detected.

Future studies should (1) evaluate multiple diverse clinical scenarios, including psychosocially complex cases; (2) compare real-time vs postencounter AI feedback; (3) examine learner outcomes such as skill acquisition and satisfaction; (4) conduct cost-effectiveness analyses at scale; and (5) develop and evaluate bias mitigation and explainability techniques—such as routine fairness dashboards, scheduled revalidation against human ratings, and faculty-led oversight processes—to satisfy accreditation requirements.

As this study was limited to transcript-based assessments of simulated encounters, future work is also needed to evaluate how well ABA scores correlate with actual clinical performance and whether AI can reduce assessor burden while maintaining fairness and reliability.

Within the constraints of this pilot, GPT-o1 Pro and GPT-5 Pro matched expert physicians in scoring simulated patient interviews; produced more reliable ratings; and delivered a substantial 58% and 67.6% reduction in analytical time, respectively. These preliminary results indicate that LLMs could serve as a complementary or alternative tool to human raters for transcript-based interview assessments. This approach warrants further investigation as a means to contribute to assessment efficiency in medical education. Careful curricular design and continuous human oversight will be essential to ensure that such tools enhance rather than compromise the validity and equity of learner evaluations.