Gray-zone decisions in subclinical hypothyroidism: a qualitative analysis of ChatGPT responses Authors: Emre Vuraloglu Published: 2026 ABSTRACT BackgroundThe management of subclinical hypothyroidism can at times present a gray area for clinicians due to heterogeneous clinical features, variable TSH (Thyroid-Stimulating Hormone) thresholds, and differences in guideline interpretations. Although large language models (LLMs) have the potential to support clinicians in such uncertainty-laden clinical situations, the existing literature has focused almost exclusively on quantitative outcomes. However, how an LLM such as Chat Generative Pretrained Transformer (ChatGPT) structures a clinical decision, which thematic components it draws upon, and how it organizes its reasoning has not yet been qualitatively examined in the context of subclinical hypothyroidism scenarios. This study aims to fill this important gap in the literature by thematically analyzing how ChatGPT generates decision-making patterns in subclinical hypothyroidism cases.MethodsFive standardized subclinical hypothyroidism scenarios, each detailing clinical features representative of the TSH 4–10 mIU/L range, were presented to ChatGPT. The responses were evaluated using a structured framework approach and reflexive thematic analysis. The resulting themes, their contents, and the corresponding clinical decision profiles were systematically compared.ResultsSeven main themes emerged: guideline-based framework, individualized decision-making, trial treatment approach, risk stratification, avoidance of overtreatment, shared decision-making, and referral to endocrinology. These themes converged into three broader decision profiles: a guideline-based conservative profile, a risk-focused proactive profile, and a symptom-balanced shared-decision profile.ConclusionsThis study is among the first to qualitatively examine the decision-making processes of ChatGPT in gray-zone clinical conditions such as subclinical hypothyroidism. The decision patterns identified here provide a methodological foundation for future large-scale, comparative studies incorporating real patient data to evaluate LLM-based clinical decision support systems.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12902-026-02356-1. Background Subclinical hypothyroidism is a biochemical condition in which serum thyroid stimulating hormone (TSH) levels are elevated above the reference range, while free thyroxine (FT4) levels remain within normal limits [1, 2]. Although this condition is often asymptomatic, some patients may experience nonspecific complaints such as fatigue, cold intolerance, weight gain, or cognitive slowing [2]. The fact that some cases of subclinical hypothyroidism are asymptomatic while others are symptomatic, the presence of additional risk factors, the possibility that the condition may be transient, and the wide variability in TSH levels can at times make the management of subclinical hypothyroidism challenging for clinicians [1, 3, 4]. These clinical variabilities can also prevent the literature and current guidelines from reaching a full consensus regarding management [1, 2, 5]. Moreover, because many clinicians do not always have access to an endocrinology specialist for consultation, these uncertainties can further complicate the clinical decision-making process. Gray-zone cases with TSH levels between 4 and 10 mIU/L constitute one of the groups in which disagreements are most frequently encountered [2, 5]. The American Thyroid Association (ATA) and the European Thyroid Association (ETA) recommend initiating treatment when TSH is ≥ 10 mIU/L; however, when TSH levels fall within the 4–10 mIU/L range, they advise that decisions be made based on individual factors such as age, the presence of symptoms, antibody positivity, cardiovascular risk profile, and pregnancy plans [2, 5]. However, several randomized controlled trials and meta-analyses in the literature have shown that levothyroxine therapy does not provide a significant benefit in terms of quality of life or symptom improvement in all patients with subclinical hypothyroidism [6, 7]. In the guideline published in 2019, routine initiation of thyroid hormone therapy was not recommended for all adults with subclinical hypothyroidism. The 2019 guideline re-evaluated earlier treatment thresholds and, in some interpretations, shifted the threshold upward while still allowing treatment to be considered in the presence of severe symptoms [8]. Nevertheless, there are also studies suggesting that early treatment may be beneficial in younger patients who are symptomatic, have autoimmune thyroiditis, or have cardiovascular risk factors or dyslipidemia [9]. However, these differing approaches and disagreements lead to inconsistent decisions among clinicians in the management of subclinical hypothyroidism and further deepen clinical uncertainty. In recent years, artificial intelligence-based large language models (LLMs) have attracted attention for their potential to support physicians in areas where additional decision assistance is needed, offering suggestions that complement rather than replace the clinician’s judgment [10–12]. However, the ways in which these models perform clinical reasoning, the thematic structures underlying their recommendations, and the qualitative nature of the responses they generate within clinical contexts have only recently begun to be explored [13]. In the current literature, the clinical performance of LLMs has been evaluated predominantly through quantitative metrics such as accuracy, safety, or guideline adherence; however, studies examining the cognitive frameworks, linguistic patterns, and decision shaping elements through which these models construct a clinical recommendation remain quite limited. Particularly in gray-zone conditions such as subclinical hypothyroidism, where interpretive variability is pronounced, it is not only the correctness of a model’s response but also the logical steps through which it constructs that response that hold critical importance for clinical practice. Therefore, studies that thematically analyze the structural content of LLM outputs fill an important gap in understanding the decision-generation architecture of these models. To our knowledge, no study in the literature has qualitatively examined the thematic structures underlying the clinical suggestions produced by LLMs in subclinical hypothyroidism cases. The aim of the present study was to conduct a qualitative analysis of the clinical recommendations generated by Chat Generative Pretrained Transformer (ChatGPT) in gray-zone scenarios of subclinical hypothyroidism, where clinicians often experience decision-making difficulty, and the literature provides limited definitive guidance. Importantly, rather than evaluating the correctness of ChatGPT’s recommendations, this study focuses on how the model structures clinical reasoning in uncertainty-laden scenarios. Materials and methods Study design and procedure This study was conducted using a qualitative thematic analysis design aimed at examining the clinical reasoning patterns of large language models. The analysis focused on how ChatGPT generated physician decision support suggestions in subclinical hypothyroidism cases, a domain in which the literature lacks clear consensus on management. The methodological process consisted of eight stages: developing five standardized clinical case scenarios related to subclinical hypothyroidism; presenting each scenario to ChatGPT-5.1 in separate sessions; recording the model’s responses in their raw form; applying thematic analysis based on the principles of Braun and Clarke [14]; extracting and categorizing themes; conducting cross-case comparison; mapping shared and divergent themes; and interpreting the findings. Development of case scenarios Five standardized case scenarios were developed to represent subclinical hypothyroidism presentations. The term “standardized” refers to the use of a consistent structural template and a fixed prompt across all cases. Each scenario was constructed using the same predefined components, including the patient’s age, gender, educational background, occupation, clinical complaints, medical history, family history, physical examination findings, and laboratory results, ensuring comparability across cases. Five scenarios were purposively developed to represent key clinical decision axes in the management of subclinical hypothyroidism. In this context, profiles were selected to include a younger low-risk patient, a middle-aged patient with cardiovascular risk, an anti-TPO–positive autoimmune patient, an older patient at risk of overtreatment, and a male patient with a high cardiovascular risk burden. The full scenarios and the responses generated by ChatGPT are provided in Appendix 1. Data collection procedure Data were obtained by presenting each case scenario to ChatGPT-5.1 using a single fixed prompt. Prompt: “Can you assist a family physician in managing the following subclinical hypothyroidism scenario?” Each scenario was presented only once, in a separate temporary chat session, to avoid any influence from prior interactions or contextual carryover between cases. The aim was to observe the model’s natural clinical reasoning structure under standardized conditions. Researcher reflexivity and analytical rigor The researcher is a family physician and academic with over five years of clinical experience, routinely managing thyroid related conditions. The researcher also serves as the Deputy Chief Physician of a tertiary-level training and research hospital. The analysis was conducted using an inductive and reflexive approach, in which themes were not pre-specified but derived directly from the data. In this context, the analysis was exploratory and not based on any prior assumptions. The responses were first read in their entirety to achieve familiarity with the data, after which meaningful units were identified and initial codes were generated. These codes were compared both within and across cases and subsequently grouped into candidate themes. Themes were then reviewed, refined, and named through an iterative process. Coding was primarily conducted at a semantic level, focusing on the explicit content of the responses, with selective interpretive (latent) extension to capture underlying decision patterns. Finally, higher-level clinical decision profiles were derived through the interpretation of recurring thematic patterns across cases. Ethical considerations This study was conducted entirely using ChatGPT generated data and involved no human participants; therefore, ethical approval was not deemed necessary. All content was based on fictional scenarios. Results In this study, the responses generated by ChatGPT-5.1 for five different standardized subclinical hypothyroidism scenarios were examined using a thematic analysis approach. Before proceeding with the analysis, differences among the cases in terms of demographic characteristics, biochemical profiles, accompanying risk factors, and clinical context were evaluated. To facilitate theme development, the features of the cases were systematically compared. Table 1 presents the similarities and notable distinctions across cases in relation to age, sex, clinical complaints, TSH pattern, anti-thyroid peroxidase antibody (anti-TPO) status, cardiovascular risk burden, and ChatGPT’s management tendencies. This comparative overview provided a foundation for understanding how each case’s clinical risk profile shaped ChatGPT’s pattern of recommendations and enabled a more transparent identification of the origins of the emerging themes. Table 1Summary of case scenariosFeatureCase 1Case 2Case 3Case 4Case 5Age / Sex28 / Female45 / Female32 / Female70 / Female55 / MaleMain ComplaintsFatigue, hair loss, general weaknessFatigue, hair loss, exertional tirednessFatigue, hair loss, cold intolerance, dry skinFatigue, cold intolerance, attention difficultyFatigue, exertional intolerance, cold intoleranceTSH (mIU/L)6.97.25.87.87.1FT4NormalNormalNormalNormalNormalAnti-TPONegativeNegativePositive (Autoimmune)NegativeNegativeComorbiditiesNoneHT, HLNoneNoneHT, HL, smokingPregnancy / PlanNoNoNoNot relevantNot relevantDistinctive FeaturesYoungMiddle age + CV riskHashimoto-related progressionElderly, overtreatment riskMale, high CV burdenChatGPT Management InclinationObservation first; trial therapy optionalObservation or trial therapyFavors therapy initiationObservation; avoid overtreatmentFavors therapy initiationReason for Management DirectionYoung age, low riskSymptoms + comorbiditiesAutoimmune thyroiditis + progression riskAdvanced age → AF/fracture riskHigh CV risk + TSH trend + symptomsSuggested LT4 Initiation Dose25–50 µg25 µg25–50 µgNot recommended, if needed 12.5–25 µg25–50 µgFollow-Up Strategy6–12 months TFT6–8 weeks (if treated)6–8 weeks + symptom reassessment6–12 months TFT6–8 weeks + lipid/CV follow-upHT, hypertension; HL, hyperlipidemia; CV, cardiovascular; TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody; TFT, thyroid function tests; LT4, levothyroxine; µg, microgram Table 1Summary of case scenariosFeatureCase 1Case 2Case 3Case 4Case 5Age / Sex28 / Female45 / Female32 / Female70 / Female55 / MaleMain ComplaintsFatigue, hair loss, general weaknessFatigue, hair loss, exertional tirednessFatigue, hair loss, cold intolerance, dry skinFatigue, cold intolerance, attention difficultyFatigue, exertional intolerance, cold intoleranceTSH (mIU/L)6.97.25.87.87.1FT4NormalNormalNormalNormalNormalAnti-TPONegativeNegativePositive (Autoimmune)NegativeNegativeComorbiditiesNoneHT, HLNoneNoneHT, HL, smokingPregnancy / PlanNoNoNoNot relevantNot relevantDistinctive FeaturesYoungMiddle age + CV riskHashimoto-related progressionElderly, overtreatment riskMale, high CV burdenChatGPT Management InclinationObservation first; trial therapy optionalObservation or trial therapyFavors therapy initiationObservation; avoid overtreatmentFavors therapy initiationReason for Management DirectionYoung age, low riskSymptoms + comorbiditiesAutoimmune thyroiditis + progression riskAdvanced age → AF/fracture riskHigh CV risk + TSH trend + symptomsSuggested LT4 Initiation Dose25–50 µg25 µg25–50 µgNot recommended, if needed 12.5–25 µg25–50 µgFollow-Up Strategy6–12 months TFT6–8 weeks (if treated)6–8 weeks + symptom reassessment6–12 months TFT6–8 weeks + lipid/CV follow-upHT, hypertension; HL, hyperlipidemia; CV, cardiovascular; TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody; TFT, thyroid function tests; LT4, levothyroxine; µg, microgram Summary of case scenarios HT, hypertension; HL, hyperlipidemia; CV, cardiovascular; TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody; TFT, thyroid function tests; LT4, levothyroxine; µg, microgram The analysis revealed that the model’s clinical reasoning patterns clustered into seven main themes (Table 2), as well as three explanatory clinical decision-making profiles across cases (Table 3). Table 2Thematic findingsThemeNoMain ThemeContent of the Theme (Subthemes)Cases1Guideline-based frameworkTSH thresholds, diagnostic alignment with FT4, emphasis on the gray zone1, 2, 3, 4, 52Individualized decision-makingAge, symptom burden, cardiovascular risk, autoimmunity, and systematic evaluation of whether symptoms may be non-thyroidal1, 2, 3, 4, 53Trial therapy approachLow-dose initiation, 6–8-week reassessment, discontinuation if not beneficial1, 2, 3, 54Risk stratificationAnti-TPO status, TSH trend, cardiovascular risk, dyslipidemia3, 55Avoiding overtreatmentAtrial fibrillation, osteoporosis, advanced-age safety considerations46Shared decision-makingPatient education, collaborative decision-making, presentation of therapeutic alternatives1, 2, 3, 4, 57Referral to endocrinologyRed flag indicators, referral criteria1, 2, 3, 4, 5TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody Table 2Thematic findingsThemeNoMain ThemeContent of the Theme (Subthemes)Cases1Guideline-based frameworkTSH thresholds, diagnostic alignment with FT4, emphasis on the gray zone1, 2, 3, 4, 52Individualized decision-makingAge, symptom burden, cardiovascular risk, autoimmunity, and systematic evaluation of whether symptoms may be non-thyroidal1, 2, 3, 4, 53Trial therapy approachLow-dose initiation, 6–8-week reassessment, discontinuation if not beneficial1, 2, 3, 54Risk stratificationAnti-TPO status, TSH trend, cardiovascular risk, dyslipidemia3, 55Avoiding overtreatmentAtrial fibrillation, osteoporosis, advanced-age safety considerations46Shared decision-makingPatient education, collaborative decision-making, presentation of therapeutic alternatives1, 2, 3, 4, 57Referral to endocrinologyRed flag indicators, referral criteria1, 2, 3, 4, 5TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody Thematic findings TSH, thyroid-stimulating hormone; FT4, free thyroxine; Anti-TPO, anti-thyroid peroxidase antibody Table 3Clinical decision profileClinical Decision ProfileDefinitionDecision StrategyCasesGuideline based conservative profileNo automatic treatment when TSH < 10; monitoring prioritized when overall risk is lowMonitoring first approach; low-dose treatment only if symptoms are prominent1,2,4Risk oriented proactive profileIncreased tendency toward treatment when autoimmunity or cardiovascular risk is elevatedStrong arguments in favor of treatment; 25–50 µg trial recommended3,5Symptom balanced shared decision profileFlexible choice between treatment and monitoring when symptoms are significantTrial therapy with symptom based reassessment1,2TSH, thyroid-stimulating hormone Table 3Clinical decision profileClinical Decision ProfileDefinitionDecision StrategyCasesGuideline based conservative profileNo automatic treatment when TSH < 10; monitoring prioritized when overall risk is lowMonitoring first approach; low-dose treatment only if symptoms are prominent1,2,4Risk oriented proactive profileIncreased tendency toward treatment when autoimmunity or cardiovascular risk is elevatedStrong arguments in favor of treatment; 25–50 µg trial recommended3,5Symptom balanced shared decision profileFlexible choice between treatment and monitoring when symptoms are significantTrial therapy with symptom based reassessment1,2TSH, thyroid-stimulating hormone Clinical decision profile TSH, thyroid-stimulating hormone First, all responses shared a common starting point rooted in a guideline-based clinical framework. In each scenario, the model characterized subclinical hypothyroidism according to TSH thresholds consistent with ATA and ETA guidelines; it generally recommended treatment when TSH was ≥ 10 mIU/L, while identifying the 4–10 mIU/L range as a gray zone requiring individualized decision-making. Across all five cases, normal FT4 levels were interpreted as supportive of the diagnosis of subclinical hypothyroidism, and in every response the model adopted a universal approach of risk-based decision-making rather than automatic treatment when TSH < 10 mIU/L. The second theme demonstrated that individualized decision-making was structured in a systematic manner. In each case, the model evaluated parameters such as age, symptom burden, autoimmune status, and cardiovascular risk profile, shaping recommendations for treatment or observation accordingly. It also consistently considered the possibility that symptoms might be explained by non-thyroidal factors (e.g., anemia, vitamin deficiencies, sleep disorders, or psychosocial stress), thereby situating the clinical assessment within a holistic framework. For example, in a young and otherwise healthy patient (Case 1), watchful waiting was proposed as the initial option, whereas the presence of hypertension and hyperlipidemia in a middle-aged patient (Case 2) modestly shifted the evaluation toward the possibility of treatment. Similarly, in the 70-year-old case (Case 4), advanced age was emphasized, and the potential harms of therapy were highlighted, leading to a cautious stance. This pattern indicates that the model does not simply apply guideline knowledge but incorporates patient-specific factors and alternative explanations for symptoms as an additional interpretive filter. The third theme was the model’s framing of levothyroxine as a reversible trial treatment option. In four cases (Cases 1, 2, 3, and 5), the model presented treatment not as a definitive or mandatory step but as an experimental intervention whose biochemical and symptomatic effect could be reassessed after a defined interval. It explicitly noted that treatment could be discontinued if no meaningful improvement in symptoms was observed. The fourth theme reflected a risk stratification-oriented mode of reasoning. Autoimmunity, particularly anti-TPO positivity with rising TSH trends in Case 3, was identified as a strong determinant favoring treatment. Similarly, cardiovascular risk factors in Case 5 were incorporated into the management plan, indicating that the model performed multidimensional risk assessment. The fifth theme centered on avoidance of overtreatment and safety-focused reasoning, especially in older adults. In Case 4, the model emphasized that levothyroxine therapy may increase the risk of atrial fibrillation, osteoporosis, and fractures, thereby adopting a strong preference against treatment. This aligns with contemporary literature cautioning against overtreatment in the elderly and shows that the LLM accounted for potential harms when formulating clinical decisions. The sixth theme was a consistent emphasis on shared decision-making and patient education across all scenarios. The model recommended ensuring that patients understand the treatment options, including the advantages and disadvantages of observation versus trial therapy, and encouraged explicit discussion of factors such as pregnancy planning, medication interactions, and follow-up intervals. This theme underscores the model’s capacity to replicate clinical communication patterns. The seventh and final theme was referral to endocrinology. Across all cases, the model identified specific thresholds and red flag situations for specialist consultation. These included TSH ≥ 10 mIU/L, progression to overt hypothyroidism reflected by a decline in FT4, development of thyroid nodules or asymmetric goiter, pregnancy or pregnancy planning, inability to stabilize TSH despite repeated dose adjustments, and the presence of complex cardiac comorbidities each consistently framed as triggers for referral. Based on these themes, three emergent clinical decision-making profiles were interpretively derived to summarize recurrent cross-case patterns. These profiles were not pre-specified and were not intended as rigid or mutually exclusive categories; rather, they represent higher-level interpretive groupings reflecting dominant decision orientations across cases. Accordingly, some cases showed partial conceptual overlap across profiles, although one profile was considered more prominent in each case for explanatory purposes. The guideline-based conservative profile characterized younger and lower-risk patients (Cases 1 and 2), emphasizing guideline-centered, observation-focused management. The risk-focused proactive profile emerged in cases with pronounced autoimmune activity (Case 3) or substantial cardiovascular risk factors (Case 5), generating stronger arguments in favor of treatment. The symptom-balanced shared-decision profile represented cases with notable symptoms but limited risk factors (Cases 1 and 2), balancing treatment and observation while prioritizing joint decision-making with the patient. Discussion This study qualitatively examined how ChatGPT-5.1 structures its clinical decision-making suggestions in gray-zone scenarios of subclinical hypothyroidism, where clinical uncertainty is particularly pronounced. By analyzing the model’s responses to standardized cases, we identified distinct thematic patterns including a guideline-based framework, individualized decision-making, a trial-treatment approach, risk stratification, avoidance of overtreatment, shared decision-making, and specialist referral and demonstrated how these themes varied across different clinical contexts and clustered into three overarching decision profiles. The findings suggest that the model organizes common variables in subclinical hypothyroidism management into recognizable linguistic and conceptual patterns and is capable of transforming gray-zone clinical presentations into a systematic decision schema. While these results are encouraging, prior studies have shown that LLMs may oversimplify complex material or, conversely, over pathologize clinical information; therefore, the use of LLMs should remain subordinate to human clinical judgment [15, 16]. In our study, ChatGPT consistently adopted a guideline-based framework across all scenarios and, in parallel, systematically applied the theme of individualized decision-making. The model largely reproduced the overarching approach of the ATA and ETA guidelines, most notably by framing the TSH range of 4–10 mIU/L as a gray zone requiring individualized assessment mirroring the guidance provided in these clinical recommendations. These findings suggest that ChatGPT can produce response patterns that appear structurally coherent and broadly consistent with selected elements of existing clinical guidelines [17]. The fact that ChatGPT responses appear to align with clinical guidelines does not mean that these responses are clinically accurate. This apparent alignment may largely reflect linguistic and conceptual patterns learned from the model’s training data. Therefore, the findings of this study should be interpreted not as an evaluation of the correctness of the recommendations, but as an attempt to understand how the model structures the clinical decision-making process. However, this appearance of guideline adherence and individualized reasoning should not be mistaken for genuine clinical judgment. The impression that LLMs engage in human-like inference is often misleading, as these models do not comprehend cognitive processes; rather, they generate responses by reorganizing linguistic patterns learned from their training data [18]. In our study, we also observed that ChatGPT consistently adopted a referral to endocrinology stance across all cases. As emphasized in the comprehensive review by Omiye and colleagues, although LLMs can provide supportive input in clinical reasoning, they cannot replace human expertise in decisions involving uncertainty, and therefore frequently generate safety-oriented messages recommending specialist evaluation [19]. Similarly, the comprehensive systematic review by Shool et al., which included 761 studies, emphasized that LLMs tend to adopt a cautious stance in clinical scenarios characterized by uncertainty and frequently produce specialist referral recommendations as a common output pattern [20]. Also, in the study conducted by Reis et al. in 2026 using four different LLM models on real patient queries, it was shown that as the urgency level of clinical questions increases, the models more frequently and more strongly recommend consulting a healthcare professional [21]. All these findings indicate that the referral behavior observed in our study is not only a reflection of the model’s internal consistency but also part of a broader safety-oriented decision support tendency. However, this pervasive referral tendency should be interpreted cautiously, as it may reflect a risk-averse response strategy embedded in the model rather than a context-specific clinical judgment. This raises important considerations regarding the balance between patient safety and the potential for over-referral in real-world clinical applications. In our study, although not present in every case, another notable theme was the trial treatment approach. This reversible trial therapy framework is consistent with randomized controlled trials and meta-analyses showing that levothyroxine provides only limited symptomatic benefit in patients with subclinical hypothyroidism. For example, in the TRUST trial involving 737 participants, no significant difference was found between levothyroxine and placebo in thyroid-specific quality of life measures or fatigue scores [6]. Similarly, the meta-analysis by Feller and colleagues, which included 21 randomized trials, demonstrated that thyroid hormone therapy does not provide clinically meaningful improvement in quality of life or thyroid-related symptoms in patients with subclinical hypothyroidism [7]. These findings help rationalize ChatGPT’s framing of treatment not as a permanent or obligatory intervention, but rather as a conditional option initiated with modest expectations and allowing for reassessment based on objective response. The literature similarly emphasizes the risks of levothyroxine overtreatment particularly in older adults regarding both cardiac and skeletal outcomes. In the large cohort study by Turner et al. involving individuals aged over 70 years, high dose levothyroxine use was associated with a significantly increased risk of fractures, demonstrating a clear dose response pattern [22]. In this context, the model’s avoidance-of-treatment theme and its tolerance for a wider TSH target range in the 70-year-old case (Case 4) reflect a reasoning pattern that aligns with current clinical risk warnings. The qualitative design of this study adds important methodological originality to the existing literature, as there are very few studies that examine the use of ChatGPT in the field of endocrinology through a qualitative lens [23]. Most prior studies evaluating LLM performance in thyroid disorders have relied on quantitative metrics such as accuracy, comprehensiveness, safety, readability, or communication quality. However, they have not examined in depth the underlying decision logic, thematic structure, or the ways in which these models guide clinicians through the reasoning process [10, 24–26]. Therefore, the existing literature has largely evaluated not how LLMs think, but rather how accurately they respond. However, in clinical gray zones particularly in interpretive contexts such as subclinical hypothyroidism accuracy metrics alone are insufficient. Analyzing the structural composition of responses, the decision-making elements they employ, and their linguistic patterns is more meaningful for clinical applicability. The primary contribution of this study is to provide a transferable coding framework for examining the structure of “gray-zone decision narratives” in LLM outputs. Subclinical hypothyroidism is positioned as a clinically relevant exemplar in which this framework is applied. By revealing these internal components at a thematic level, the study shifts the focus from performance evaluation toward understanding how LLMs structurally organize clinical decision-making processes. At the same time, this work can be viewed as an early contribution that may help lay the groundwork for future systematic reviews as the literature in this field expands. This study has several limitations. First, the analysis is based on only five case scenarios, which may not fully reflect the complexity of real-life clinical environments, patient physician interactions, or long-term follow-up processes. In addition, it is possible that the number of scenarios was insufficient to capture all possible thematic patterns that may emerge in subclinical hypothyroidism. The cases were constructed with variation in age, gender, symptoms, occupation, laboratory values, and comorbidity, which allowed exploration across clinically meaningful contexts; however, this limited the ability to directly compare the model’s responses under controlled conditions. A more controlled design, in which a single parameter is varied while other variables are held constant, could have provided more precise insights into how the model evaluates individual factors in clinical decision-making. Furthermore, because all laboratory parameters were systematically provided across scenarios, the present design did not allow evaluation of whether the model would independently request or identify missing clinically relevant information such as autoimmunity markers. Second, only a single ChatGPT model available at the time of data collection was evaluated; different LLM versions or models may produce different results. In addition, the model may change over time due to system-level updates, and model labels may not always fully reflect the same underlying technical configuration. This may limit the ability of readers to independently verify the model used. Therefore, reproducibility and model transparency constitute important limitations in studies involving LLMs. Third, thematic analysis is an interpretive methodology, and because it incorporates the researcher’s analytical judgments, its generalizability is limited. Another limitation of this study is that the thematic analysis was conducted by a single clinician-researcher. Future studies that compare multiple LLMs, use a broader range of case scenarios, and include clinicians from different countries through multicenter international designs would strengthen the validity of the findings. In addition, more controlled study designs in which individual clinical parameters are varied while others are held constant may provide more precise insights into how LLMs evaluate specific factors in clinical decision-making. Accumulating evidence over time may also create the foundation for the first comprehensive systematic reviews in this field. Conclusion This study presents a qualitative thematic analysis designed to make visible how ChatGPT-5.1 generates decisions in gray-zone clinical scenarios such as subclinical hypothyroidism, where management inherently involves uncertainty. In contemporary practice, it is increasingly recognized that providing merely accurate clinical answers is not sufficient for LLM-based tools; what is critically important is understanding the linguistic patterns, logical steps, and risk-prioritization strategies through which the model constructs its decisions. The seven overarching themes identified in our analysis guideline-based framework, individualized decision-making, trial-treatment approach, risk stratification, avoidance of overtreatment, shared decision-making, and specialist referral demonstrate that LLM outputs contain a structured decision architecture that extends beyond surface-level correctness. This qualitative perspective aligns with prior work suggesting that the risk benefit balance of artificial intelligence in clinical care can only be fully understood through content-level examination rather than accuracy metrics alone. By making these internal components of LLM outputs visible at a thematic level, our study provides a unique methodological framework for the future design of artificial intelligence-supported decision processes in subclinical hypothyroidism that may become more auditable, transparent, and improvable over time. In conclusion, this study offers one of the first qualitative examinations of how LLMs reason in the management of subclinical hypothyroidism and contributes to building a broader methodological foundation for the use of artificial intelligence in clinical gray-zone decision-making. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1