Authors: Sarper İçen, Arif Hüdai Köken
Categories: Research, Child and adolescent psychiatry, Ethical decision-making, Artificial intelligence, Medical ethics, Clinical decision support systems, Cultural sensitivity, Autonomy, Confidentiality, Explicability, Accountability, Large language models
Source: BMC Medical Ethics
Authors: Sarper İçen, Arif Hüdai Köken
Ethical decision-making in child and adolescent psychiatry (CAP) is inherently complex, shaped by developmental vulnerability, evolving autonomy, and competing responsibilities to patients, families, and the legal system. Clinicians often face moral dilemmas when navigating adolescent confidentiality, parental authority, and mandatory reporting duties, especially in high-stakes or culturally sensitive contexts. As large language models (LLMs) enter clinical settings, their potential to support ethical reasoning remains underexplored, particularly outside Western paradigms. This study qualitatively investigates how different LLMs provide ethical, legal, and emotional guidance to clinicians facing ethically challenging scenarios in CAP, situated within Turkiye’s sociocultural and legal landscape.
A scenario-based qualitative design was employed. Three expert-developed case vignettes reflecting ethically charged dilemmas, such as adolescent autonomy, parental conflict, and confidentiality, were submitted to the three LLMs (ChatGPT 4.0, Gemini 2.5 Flash, and GROK 3). Responses were analyzed using content and thematic analysis to identify key patterns of ethical-legal reasoning, alongside discourse analysis to examine tone, empathy, and cultural sensitivity. Two researchers, with backgrounds in CAP and medical ethics, conducted independent coding and reached consensus through a reflexive, interdisciplinary approach.
All LLMs addressed core ethical principles (autonomy, non-maleficence, beneficence, and justice) and referenced Turkish legal frameworks such as the Child Protection Law, Patient Rights Regulation, and mandatory reporting obligations, situating their guidance within the national regulatory context. They also differed in their engagement with sociocultural GROK 3 emphasized therapeutic communication and relational trust, Gemini 2.5 Flash applied a highly structured, rule-based style focused on procedural compliance, while ChatGPT 4.0 provided concise and practical suggestions. Despite thematic overlaps, these varying approaches shaped how effectively the models aligned with Turkiye’s clinical realities. Notably, LLMs frequently acted as “thinking companions,” offering ethical and legal justifications while leaving interpretive responsibility with clinicians.
LLMs in CAP hold promise not only as cognitive aids but also as emotionally attuned, context-sensitive companions in ethical decision-making processes. Their effectiveness depends not just on algorithmic precision but also on explainability, empathy, and cultural alignment. Rather than replacing clinician judgment, LLMs may serve to ease emotional burden, enhance therapeutic reflection, and foster ethically sound care in complex, high-pressure situations.
The online version contains supplementary material available at 10.1186/s12910-025-01323-0.
Child and adolescent psychiatry (CAP) includes specific ethical challenges due to developmental vulnerability, limited autonomy, and multi-actor communications in practice. Clinical decisions require protecting the best interests of the child in addition to encouraging developmentally appropriate participation and a trust-based therapeutic relationship [1]. The autonomy of a patient who has not completed cognitive and moral developmental stages often conflicts with parental preventive approaches. In such situations, it is not easy to balance between parents’ expectations and children’s needs for the clinician [2, 3]. For this reason, ethical evaluation and decision-making in CAP is not only an individual but also a contextual and relational process [4]. Personal sensitivity of the physician, as well as ethical intuition, and support of supervision, play a determining role in these decision-making processes [4]. When legal issues are addressed, ethical challenges become more complex by considering the best interests of children. One example of this complexity is sharing confidential information about children without their consent in mandated reporting [5]. In general, CAP practice requires preserving the privacy of adolescent patients and adhering to the duty of confidentiality of the physician. Specifically in medicolegal conditions such as suicidal risk, child abuse, or substance use, the extent of preserving patients’ confidentiality necessitates a fine balance between therapeutic alliance and ethical responsibility [6]. For these reasons, CAP practice should be understood as a specialty area requiring developmental sensitivity, contextual awareness, and ethical responsiveness rather than strict rules.
Ethically, the globally relevant respect for autonomy, beneficence, non-maleficence, and justice principles provide a framework for clinical decision-making processes in CAP. Besides, ethical theories that form a basis for ethical reasoning in clinical decisions strengthen this framework [7]. Each country’s own legal and cultural systems determine the process of evaluating ethical challenges. In Turkiye, CAP clinical practice is shaped not only by universal ethical principles but also by national legal frameworks and family-centered cultural values. Culturally, Turkiye is characterized by a multilayered value system in which family authority, interdependence, and social reputation play a central role. This often means that parental authority and collective family decision-making can outweigh individual adolescent autonomy, creating unique tensions for clinicians compared to more individualistic Western contexts. Issues such as secrecy, stigma (e.g., around mental illness or sexuality), and strong parental involvement in healthcare decisions shape how ethical dilemmas are experienced and resolved. In clinical encounters, physicians often face expectations not only from parents but also from extended family networks, where preserving family harmony and protecting social reputation (e.g., avoiding shame or dishonor) may become as influential as legal obligations.
Legally, the Constitution of the Republic of Turkiye encompasses the broadest framework for clinical decisions in CAP [8]. In accordance with the Constitution, the Turkish Civil Code shapes the CAP clinical practice by defining legal representation authority and the principle of protection of individuals who are not of full age, as well as the protection of personal rights, the definition and exercise of custody rights, and the protection of the child’s rights [9]. The Patient Rights Regulation clearly defines the rights of individuals to give consent, to receive information, and to have their privacy protected [10]. The Law on the Protection of Personal Data No. 6698 sets out the ethical and legal boundaries for the processing of health data and their use in digital systems [11]. The Turkish Penal Code defines sanctions for acts that constitute crimes related to clinical cases [12]. In addition to all this national legislation, the Convention on Human Rights and Biomedicine (Oviedo Convention), to which Turkiye is a party and which is also part of domestic law as it was voted and adopted by the Grand National Assembly of Turkiye, provides an international framework for clinical practices in child and adolescent psychiatry [13]. According to the United Nations Convention on the Rights of the Child (1989), to which Turkiye has also signed and is a State Party, the general principles that must also be ensured in clinical practice in child and adolescent psychiatry include the provision of the child’s inherent right to life and the maximum possible assurance of their development, adopting the best interests of the child as a primary consideration in all actions affecting children, ensuring that no discrimination is made against children for any reason, and granting children the right to freely express their views on all matters affecting them, with due weight given to those views in accordance with their age and maturity [14].
The Turkish Civil Code defines “full age” as 18 years, marking the age of majority. Minors under 18 generally cannot provide full legal consent, except in specific circumstances such as emancipation through marriage. Parental authority (custody) is exercised jointly by parents, who are considered legal representatives of the child until majority, unless limited by court order. Decision-making capacity, however, is evaluated adolescents may be encouraged to participate in health decisions proportionate to their age and maturity, even though their legal guardians hold formal consent authority. These definitions differ across jurisdictions; for example, in some European countries, adolescents can consent independently to certain treatments from age 14 or 16, underscoring the importance of contextual sensitivity in CAP ethical dilemmas. In Turkiye, clinicians must navigate this intersection of legal parental authority and the adolescent’s evolving capacity for autonomy.
In terms of ethics support structures, Turkiye differs from many Western countries. While most universities and hospitals host Scientific Research Ethics Committees that review research protocols, there are currently no routine clinical ethics consultation services available for day-to-day practice. Clinical decisions in ethically complex cases are therefore left mainly to the individual clinician’s judgment, occasionally supported by informal peer discussions or hospital administration. The Turkish Medical Association provides ethical guidelines, but these remain general and do not substitute for case-specific consultation. Consequently, in Child and Adolescent Psychiatry, where dilemmas often involve minors’ autonomy, parental authority, and mandated reporting, physicians must rely heavily on their own ethical judgment and legal obligations rather than institutionalized ethics consultation mechanisms.
Artificial intelligence (AI) has been increasingly used in medicine and health sciences, as well as in other scientific areas. Currently, AI has passed beyond offering information-based recommendations and reached the capacity of serving multidimensional clinical decision support. Progress in natural language processing, imaging analysis, and time series data has increased the potential of AI for individually tailored approaches in health services [15]. In terms of mental health research and practice, the use of AI shows promising advances in diagnosis, risk prediction, and treatment decision support processes [16]. AI-based clinical decision support systems can integrate a patient’s demographics, diagnosis, and treatment history to provide personalized guidance to physicians [17, 18]. Explainable AI models also clearly demonstrate how the systems generate their recommendations, thereby strengthening the accuracy of the decision-making process and enabling physicians to fulfill their responsibilities [19, 20]. In this study, we focus specifically on large language models (LLMs), a subset of AI designed for natural-language understanding and generation, to examine how they articulate ethical and legal guidance in CAP.
Current evidence supports AI-based systems’ accuracy in diagnosis or quickening the process; however, their role in guiding ethically challenging, complex decisions remains underexplored in the literature. Beyond the four basic principles of biomedical ethics, Floridi and Cowls (2022) suggested a fifth principle for AI explicability, which emphasizes intelligibility and accountability [21]. While explicability has mainly been discussed as a technical property of AI systems, its relevance is particularly salient in psychiatry and CAP, where cultural, relational, and emotional factors shape decisions. A recent review emphasized that the ethical support capacity of AI is theoretically framed around autonomy, beneficence, and explicability, but empirical and contextual research is lacking [22]. This gap highlights the need to investigate not only whether AI can provide accurate information but also whether its guidance is intelligible, accountable, and context-sensitive for clinicians in real-world practice.
In summary, the Turkish context requires balancing universal ethical principles with culturally embedded family dynamics and nationally specific legal frameworks. These intersecting layers make CAP practice particularly challenging and highlight the need to evaluate whether tools such as LLMs can adapt to both legal requirements and cultural sensitivities. However, currently available AI-based decision support systems are shaped around the Western approaches; cultural adaptation, justice principle, and human-machine interactions fall behind [22, 23]. This study analyzes the ethical and legal guidance of three different LLMs on ethically challenging case scenarios within the context of CAP clinical practice, aiming to qualitatively explore how AI offers an ethical approach based on the cultural and legal context of Turkiye. In this direction, the primary research question of this study How do LLMs offer guidance to physicians for the ethically challenging clinical decision process in the context of Turkiye?
In this study, we aimed to explore how LLMs guide the physician in ethically complex and clinically challenging situations, specifically in CAP. For this purpose, we adopted a qualitative research design based on LLM-generated responses and case scenarios.
The researchers developed three possible case scenarios for CAP clinical practice in Turkiye. These scenarios were structured to present ethical and legal conflicts specific to adolescence. Each of the three case scenarios represented a clinically challenging situation for the physician in a realistic manner. Table 1 provides concise summaries of the three vignettes along with the core ethical challenges embedded by the authors in each scenario, which were used to elicit LLM responses. The full texts of case narratives are provided in Supplementary Table S1 to maintain transparency and enable readers to review the verbatim vignettes used as prompts for LLM analyses, without interrupting the flow of the main text.
Table 1Summary of case scenarios prepared for this research by the authorsCase ScenarioSummary of CaseEthical Challenges1A 16-year-old under psychiatric follow-up reports improvement on sertraline + quetiapine, but is taking them secretly because the father rejects psychiatric treatment. The system’s electronic prescribing/notification process will inform parents if a prescription is issued, creating a conflict between the adolescent’s wish for confidentiality and parental/administrative transparency.Adolescent’s autonomy vs. parents’ intervening; confidentiality principle vs. health system’s principle of transparency; non-maleficence principle; dual obligations of the physician2A 17-year-old female from a rural setting presents accompanied by an older man and his father; she reports being “married off” and reports depressive symptoms. The clinician suspects possible coercion, abuse, or trafficking but faces legal and practical barriers to assessment and reporting given uncertain guardianship and possible reprisals.Mandated reporting vs. diagnostic uncertainty; autonomy vs. protectiveness; risk of injustice; wanting to help clinically vs. legal limitations3A 17-year-old male with recurrent non-suicidal self-harm discloses attraction to males and involvement in sexual relationships with older men found online, requests confidentiality, and reports associated guilt and ongoing self-harm, creating tensions between confidentiality, duty to protect, and cultural/legal sensitivities.Confidentiality and trusting relationship vs. reporting potential risks; non-maleficence vs. respect for privacy; societal norms vs. individual rights; self-harming behaviors vs. underlying emotional and social burden*Full texts of case narratives are available in Supplementary Table * S1
In preparation, the two researchers (a CAP specialist and a medical ethics/health law specialist) collaborated closely, drawing on their complementary expertise to ensure both clinical realism and ethical–legal accuracy. To enhance the credibility of the scenarios, the vignettes were cross-checked against typical cases encountered in clinical practice and aligned with established ethical dilemmas frequently discussed in the literature. Although the scenarios were not piloted with additional clinicians or ethicists prior to the study, they were deliberately designed to reflect recurrent ethical challenges (e.g., adolescent autonomy, parental intervention, confidentiality, mandated reporting, and sociocultural conflicts) that are highly representative of real-world CAP practice in Turkiye. Furthermore, the three cases were selected to capture a spectrum of dilemmas, ranging from conflicts between adolescent autonomy and parental authority to challenges shaped by cultural norms to sensitive disclosures requiring confidentiality, so that the LLMs’ responses could be assessed across diverse ethical situations. This approach allowed us to capture both clinical authenticity and contextual sensitivity while ensuring ethical dilemmas were grounded in practice realities rather than hypothetical extremes.
Developed case scenarios were introduced to the LLMs one at a time in separate, newly initiated conversations, so that each model processed the case independently without retaining memory of prior scenarios. Importantly, all queries were conducted anonymously without logging into the platforms, ensuring that no conversation history was stored or available to the models. For each case, the same standardized prompt was ‘Can you analyze the case scenario I am going to present to you about adolescent psychiatry in ethical and legal aspects,* and offer suggestions to the physician dealing with the case for supporting decision-making?*’
The full responses generated were then recorded verbatim as raw data for analysis. All case scenarios were entered into the LLMs in Turkish, and the models’ responses were generated in Turkish. The qualitative analyses (content, thematic, and discourse analyses) were conducted in Turkish by the research team. The analytic findings were then translated into English for reporting in this manuscript. To ensure transparency, we provide sample Turkish excerpts from the LLMs’ responses in Supplementary Table S2, with English translations. Full transcripts are available from the authors upon reasonable request.
The research team consisted of two Dr. Sarper İÇEN, M.D. (male, specialist in Child and Adolescent Psychiatry) and Dr. Arif Hüdai KÖKEN, M.D. (male, specialist in Medical Ethics, with a master’s in Health Law). Both researchers are trained in qualitative analysis and worked reflexively, aware of the influence of their professional backgrounds on the analytic process. While the psychiatry researcher approached the data with sensitivity to clinical decision-making and the patient–physician relationship, the ethics/law researcher evaluated responses through the lens of ethical principles, legislation, and legal responsibilities.
The responses from the three LLMs were systematically analyzed using a combination of content analysis as described by Krippendorff (2019) and thematic analysis as outlined by Braun & Clarke (2019) [24, 25]. In this process, each response text was examined carefully with open coding, and repeating themes were determined in ethical, legal, and clinical dimensions. In line with these themes, codes and higher-level codes were defined, and AI responses were categorized accordingly. We employed a hybrid thematic analysis approach, combining inductive coding of LLMs’ responses with a deductive focus on established ethical principles (e.g., autonomy, non-maleficence, beneficence, justice) and relevant Turkish legislation (e.g., Turkish Penal Code, Child Protection Law, Patient Rights Regulation). Themes were primarily analyzed at a semantic level, reflecting the explicit content of AI outputs rather than latent meanings. Both researchers coded independently and then engaged in iterative discussions to co-construct themes, ensuring a collaborative and reflexive analytic process.
In a second stage, linguistic features of the AI responses were analyzed to examine how recommendations were framed according to discourse analysis principles outlined by Gee (2014) [26]. This analysis focused on discursive aspects such as form of address, predicate use, empathy, cultural sensitivity, and references to the local context. The aim was to reveal how LLMs communicated ethical and legal guidance beyond the substantive content of their recommendations.
Several strategies were adopted to ensure analytic reliability. The development of themes was supported with representative excerpts from the LLMs’ responses. Triangulation between thematic and discourse analyses strengthened the interpretive depth, while a pattern-focused discourse analysis (e.g., consistency of communicative structures, repeated words, predicate strength, contextual sensitivity) enhanced rigor. Methodological integrity, transparency, and systematicity were maintained throughout all stages of analysis.
This study was conducted without seeking ethical approval since the study design was based on case scenarios and no real human data was used. The researchers declare that they cared about the research integrity principle, preserving scientific validity and reliability throughout the study.
In this part, findings from our qualitative analyses are presented. First, we present the findings related to the content analysis of the LLMs’ guidance, including an ethical and legal breakdown of the case, and suggestions for the physician related to decision-making. Second, findings related to the discourse analysis of the LLMs’ suggestions to physicians, including structural and contextual language use, are presented.
Before proceeding, we recommend that readers refer to Table 1 to view a summary of the case scenarios prepared by the authors for this study, along with the ethical challenges they have embedded in each of the scenarios specifically.
Responses generated by the three LLMs on three case scenarios were analyzed by classifying thematically in ethical and legal content and making comparisons across categories such as ethical principles (e.g. autonomy, non-maleficence, beneficence, and justice), legal obligations (e.g. obligation of reporting, patient rights, consent processes), referral to relevant supporting units (e.g. ethical committee, social services department, legal support unit), physician responsibilities, and patient confidentiality.
For Case Scenario 1, which concerned balancing adolescent autonomy with parental involvement, all three LLMs emphasized the importance of patient autonomy, confidentiality, and mandated reporting, situating their guidance within Turkish legal frameworks such as the Child Protection Law and electronic prescription system (Table 2). They also converged on the need for physicians to maintain a trusting relationship with the adolescent, communicate constructively with the family, and seek institutional supports when necessary. However, the LLMs differed in emphasis and structure. GROK 3 provided the most systematically organized analysis, separating ethical principles from legal obligations and highlighting both patient and physician responsibilities. Gemini 2.5 Flash integrated ethical and legal elements into a multidimensional framework, distinguishing itself through highly structured suggestions that extended to communication strategies and documentation duties. ChatGPT 4.0 provided a more concise overview, striking a balance between ethical principles and practical legal commentary in accessible language. In terms of recommendations, GROK 3 proposed a comprehensive set of steps, including strengthening trust, facilitating adolescent participation in decision-making, and thorough documentation. Gemini 2.5 Flash offered systematic, step-by-step guidance, covering communication with parents, transparency about treatment rationale, and referral for legal advice if needed. ChatGPT 4.0 presented shorter, directive suggestions, focusing on open communication, ethics committee consultation, and limited confidentiality protections.
Table 2Comparative synthesis of thematic analysis for case scenario 1DimensionGROK 3Gemini 2.5 FlashChatGPT 4.0Ethical focusAutonomy, confidentiality, beneficence, justiceAutonomy, informed consent, parental rights, confidentialityAutonomy, beneficence, non-maleficence, justiceLegal framingChild’s best interest, physician’s duty, e-prescription and confidentialityConsent capacity, parental refusal, mandatory documentationConfidentiality, competence to consent, mandated reportingRecommendationsMaintain trust, empower adolescent, shared decision-making, documentationTransparent communication with family, record-keeping, psychosocial supportCommunication, limited confidentiality, ethics/legal consultationOverallIntegrative and therapeutic approachSystematic and procedural frameworkConcise, directive guidance*Detailed thematic analysis of the three LLMs’ responses for Case Scenario 1 is provided in Supplementary Table * S2
For Case Scenario 2, which involved an adolescent bride presenting for health services, all three LLMs emphasized the physician’s duty of mandated reporting and referenced relevant Turkish legal frameworks, particularly the Child Protection Law (Table 3). Shared ethical concerns included protecting the adolescent’s safety, managing clinical care under conditions of coercion, and balancing autonomy with the principle of beneficence. However, the models differed in the way they structured these analyses. GROK 3 highlighted the risks of forced marriage and systematically addressed autonomy, non-maleficence, and justice, while questioning the limits of parental consent in such circumstances. Gemini 2.5 Flash placed the best interests of the child at the center, combining ethical sensitivity with concrete legal steps, including referral to the Ministry of Family and Social Services. ChatGPT 4.0 provided a more concise evaluation, focusing on confidentiality, beneficence vs. non-maleficence, and the age of marriage under the Turkish Civil Code. In their suggestions to the physician, all three models prioritized emergency management, legal reporting, and immediate clinical support. GROK 3 stood out for its multidimensional plan that included safety strategies for both adolescent and physician, communication with the family, and psychotherapeutic support. Gemini 2.5 Flash again offered a highly structured approach with systematic steps such as ensuring a safe environment, maintaining detailed records, and working closely with social services. ChatGPT 4.0 presented plain, directive recommendations, including cooperation with child protection units, continuation of follow-up care, and limiting risky contacts.
Table 3Comparative synthesis of thematic analysis for case scenario 2DimensionGROK 3Gemini 2.5 FlashChatGPT 4.0Ethical focusEmphasized beneficence, non-maleficence, and justice, with strong attention to risks of forced marriage.Centered on autonomy, confidentiality, and best interests of the child.Highlighted confidentiality and trust while balancing benefits vs. harms.Legal framingReferred to Child Protection Law and Penal Code provisions on forced marriage and abuse.Detailed references to Penal Code (Articles 103, 104, 109) and obligations under the Ministry of Family and Social Services.Focused on Civil Law (marriage age), Penal Code (103, 279), and Child Protection Law.RecommendationsEmergency safety planning, mandated reporting, multidisciplinary support, and strategies for physician safety.Emergency verification, safe environment, mandated reporting, meticulous documentation, and social service referral.Judicial notification, collaboration with Child Protection Unit, ongoing follow-up, limiting risky contacts.OverallMultidimensional and comprehensive.Highly structured, duty-focused, and systematic.Pragmatic and direct, prioritizing safety.*Detailed thematic analysis of the three LLMs’ responses for Case Scenario 2 is provided in Supplementary Table * S3
For Case Scenario 3, which involved disclosure of same-sex attraction during psychiatric follow-up, all three LLMs emphasized the adolescent’s right to privacy, autonomy, and safety, while recognizing the physician’s mandated reporting responsibilities under Turkish law (Table 4). Common concerns included balancing confidentiality with legal duties, managing family dynamics, and safeguarding the adolescent’s well-being. The models diverged in their emphasis. GROK 3 applied the four classical principles (autonomy, beneficence, non-maleficence, and justice) and provided a detailed discussion of minors’ participation in health decisions, the limits of sexual consent, and reporting obligations under the Civil Code and the Patient Rights Regulation. Gemini 2.5 Flash highlighted confidentiality, informed consent, and child well-being, providing a structured legal framing grounded in Turkish Penal Code 103 and 279, as well as social service regulations. ChatGPT 4.0 differed in that it explicitly emphasized cultural sensitivity and respect for sexual identity, while referencing Penal Code 103 and parental notification obligations. In terms of recommendations, all three models offered multilayered support strategies focused on maintaining trust, risk management, and communication with the family. GROK 3 proposed the most comprehensive plan, including psychotherapeutic support, medication re-evaluation, and multidisciplinary cooperation. Gemini 2.5 Flash emphasized risk-focused communication, encouraging adolescent participation in decisions and careful staged family involvement, alongside concrete reporting and documentation steps. ChatGPT 4.0 provided a more concise and directive approach, emphasizing secrecy and respect for identity, but calling for judicial reporting and gradual family involvement when risks were identified.
Table 4Comparative synthesis of thematic analysis for case scenario 3DimensionGROK 3Gemini 2.5 FlashChatGPT 4.0Ethical focusEmphasis on confidentiality, autonomy, beneficence, non-maleficence, and justiceAutonomy, informed consent, and parental rights; best interests of the childConfidentiality, beneficence, cultural sensitivity, and respect for sexual identityLegal framingCivil Code (consent of minors), Patient Rights Regulation (maturity-based consent), Penal Code (sexual relations 15–18), mandatory reporting (TPC 278–279)Patient Rights Regulation (confidentiality), Penal Code (sexual abuse, reporting), Social Services LawPenal Code (sexual abuse, reporting), Civil Code (parental notification), Child Protection LawRecommendationsProtect confidentiality, strengthen trust, conduct risk assessment, provide psychotherapeutic support, manage family involvement, comply with legal duties, adopt a multidisciplinary approachBuild trust, involve adolescent in decisions, carefully structure family communication, comply with reporting obligations, document thoroughlyMaintain confidentiality while fulfilling reporting obligations, provide psychosocial support, gradually involve family, document the entire processOverallCounsellor-like, therapeutic, and inclusiveProcedural, duty-focused and structuredColleague-like, concise and pragmatic*Detailed thematic analysis of the three LLMs’ responses for Case Scenario 3 is provided in Supplementary Table * S4
Across all three scenarios, the models displayed consistent discourse identities that shaped how they guided the physician (Table 5). GROK 3 employed an inclusive, consultative, and option-giving style, with softer auxiliary verbs that created space for physician discretion and shared decision-making. Gemini 2.5 Flash, by contrast, consistently used a structured, duty-oriented, and legalistic tone, with decisive predicates that framed responsibilities as obligations. ChatGPT 4.0 adopted a colleague-like, supportive, and empathetic manner, using moderate-strength predicates and auxiliary verbs to offer direction without coercion. The LLMs also demonstrated case-specific discourse applications, which aligned with the dilemmas highlighted in the thematic analysis. All three showed sensitivity to Turkiye’s local context, though in distinct ways. GROK 3 went beyond procedural compliance, weaving legal references into cultural and therapeutic paternal resistance to psychiatric care in Case 1, rural marriage practices in Case 2, and stigma linked to sexual orientation in Case 3. Gemini 2.5 Flash consistently adopted the most formal register, presenting duties in categorical terms and framing obligations as non-negotiable. ChatGPT 4.0 tended to reference systemic mechanisms more generally, in a supportive tone, without elaborating extensively. Collectively, these patterns suggest that while ChatGPT 4.0 emphasized systemic awareness, Gemini 2.5 Flash underscored duty, and GROK 3 contextualized obligations within cultural and relational dynamics, producing the most expansive framing of local context.
Table 5Cross-case comparative discourse analysis of LLMs’ suggestionsGemini 2.5 FlashGROK 3.0ChatGPT 4.0Form of AddressDuty- and responsibility-focused; reminds physician of obligations.Consultative, inclusive, advisory tone.Direct, collegial; speaks to physician as a peer with supportive authority.Tone & StyleSerious, structured, systematic.Empathetic, explanatory, detailed.Supportive, empathetic, motivational.Auxiliary Verbs & Predicate StrengthMandatory constructions (“is required,” “must”); very high strength.Option-giving (“may,” “could”); moderate strength, consultative.Necessity-focused (“should,” “is necessary”); moderate strength.Level of GuidanceHigh: detailed, structured, procedural.High: layered strategies, multidimensional.High: practical, actionable, stepwise recommendations.Managing Complex SituationsOutlines clear institutional/legal pathways.Encourages consultation, creativity, and flexible options.Suggests interim or supportive solutions, balancing patient needs.Physician–Adolescent–Parent BalanceStresses physician’s legal duty over relational balance.Seeks therapeutic compromise and shared responsibility.Emphasizes empathy and patient privacy.Local Context SensitivityExplicitly cites Turkish Penal Code and regulations.Integrates systemic/legal context and community supports.References national frameworks in general terms.Cultural SensitivityExplicit references to cultural practices (e.g., rural marriages).Strongest focus on stigma, prejudice, and family dynamics.Implicit, harmonious tone without overt references.Detailed discourse analysis of the three LLMs’ responses for each case scenario is provided in Supplementary Tables S6,* S7*,* and S8*
Sociocultural emphasis in LLMs’ suggestions varied across cases and models. Across cases, the models adapted their guidance to Turkiye’s family-centric norms, authority dynamics, and stigma-related concerns, though with varying depth. They consistently recognized the importance of protecting family harmony, avoiding loss of trust, and preserving consensus, reflecting cultural expectations alongside legal frameworks. All three acknowledged paternal authority and the adolescent’s fear of damaging trust in Case Scenario 1, recommending graduated disclosure and alliance-building to counter stigma toward psychiatric treatment. For Case Scenario 2, they engaged with risks arising from non-official unions, multigenerational male authority, and close-knit community dynamics. GROK 3 foregrounded power imbalances and safety concerns, Gemini 2.5 Flash stressed the rural context and practical safeguards, and ChatGPT 4.0 emphasized swift protective action with less cultural elaboration. The LLMs reflected anticipated stigma and rejection around sexual orientation for Case Scenario 3. GROK 3 linked shame to self-harm risk and advocated protective confidentiality; Gemini 2.5 Flash operationalized a staged, risk-focused engagement strategy; and ChatGPT 4.0 affirmed identity while cautioning against involuntary disclosure in conservative family settings. Taken together, these sociocultural emphases show that while GROK 3 contextualized cultural dynamics most comprehensively, Gemini 2.5 Flash balanced duty-focused communication with situational awareness, and ChatGPT 4.0 reinforced therapeutic trust and psychological safety, all three attempted to adapt ethical guidance to honor/reputation concerns, parental authority, and the limited infrastructure for psychosocial support in Turkiye.
The stylistic profiles of the three models remained stable across cases, yet their expression varied according to the specific ethical and cultural dilemmas. GROK 3 promoted consultative, culturally sensitive, and multidimensional guidance, Gemini 2.5 Flash stressed duty and legal compliance, and ChatGPT 4.0 consistently emphasized empathy and the therapeutic relationship. The interaction of these styles with case-specific challenges produced different consensus-building in Case Scenario 1, statutory compliance in Case Scenario 2, and stigma-sensitive confidentiality in Case Scenario 3. This shows that while the discourse styles of the LLMs are relatively fixed, the content emphasis adapts to the ethical, legal, and cultural dimensions of each case scenario.
All clinical decisions are intrinsically complex, including medical, ethical, and legal aspects. For the physician, making the right decisions and managing the case is an important responsibility. This brings forward the need to evaluate decision-making processes and make improvements. It is believed that the use of AI will revolutionize health services in various areas, including imaging, diagnosis, and treatment, as well as workflow optimization in clinics [27]. Recently, the use of AI-based clinical decision support systems offers a perspective to physicians, contributing to the decision-making process by analyzing and supporting clinical decisions [28]. Previous results suggest that AI can offer clinical guidance to physicians not only in medical aspects but also in ethical and legal aspects [29]. Our findings also clearly demonstrated that LLMs could contribute to clinical decision-making processes in addition to informing clinicians about the ethical and legal dimensions. All three LLMs frequently emphasized the basic ethical principles and drew legal implications in accordance with the current regulations in Turkiye (e.g., Turkish Penal Code 103, 104, 279; Child Protection Law; Patient Rights Regulation). Moreover, we observed remarkable stylistic differences among the models, including GROK 3’s systematic and normative approach in the tone of a counsellor, Gemini 2.5 Flash’s structured guidance with a more formal, duty-focused language, and ChatGPT 4.0’s practical guidance with high cultural sensitivity. These differences become more prominent in the context of scenarios, particularly in risk situations or socially sensitive cases. In summary, our findings suggest that AI-based clinical decision support systems have the potential to guide physicians in clinical situations with complex ethical and legal implications, providing multidimensional support with a focus on accountability.
The three LLMs frequently highlighted the foundational ethical principles, including patient autonomy, non-maleficence, and confidentiality, and offered implications compatible with the legal regulations in Turkiye. These findings align with previous studies, as reviewed by Benzinger et al., in that patient autonomy is central to the AI’s support in decision-making, demonstrating the potential of LLMs to support clinical understanding of autonomy to reflect choices made by the patients accurately [22]. Additionally, LLMs not only reference ethical principles but also integrate these into local legal regulations. These findings point out AI’s capability to make ethical and legal assessments in accordance with universal principles and legal regulations, and offer sensitive and knowledge-based support in clinical decision-making processes.
Ethical consultation plays a crucial role in addressing challenges encountered in clinical practice. Institutional ethical committees are multidisciplinary units that clinicians apply to for support in the management of challenging situations, usually with no possibility of obtaining quick responses [30]. LLMs may offer an alternative and quicker ethical and legal support to physicians, particularly in situations when a decision must be made as soon as possible [22].
The role of LLMs in the decision-making process can be described as a “decision support tool,” providing clarity and strengthening the reasoning in the process, rather than being a decision-maker [31]. Our findings suggest that LLMs provide a supportive rationale for ethical decisions, rather than making decisions themselves. For instance, prioritization of the values by ChatGPT 4.0 in presenting choices is a practical example of a “decision support tool”. In this respect, all three systems functioned as a practical guidance to ethical decisions in addition to presenting normative knowledge.
Integration of AI into the healthcare system offers a revolutionary potential due to AI’s unmatched ability to improve clinical case management while analyzing and taking multiple factors related to the case into account. The success of these programs lies not only in their technical perfection but also in their ability to confirm findings during the decision-support process. The future of AI’s use in healthcare appears promising, given its potential to ensure and safeguard the safety of both patients and healthcare professionals [32].
Our findings highlight that LLMs can support ethical decision-making in a conceptually and contextually valid way. However, it is worth noting that the specialist physicians are responsible for examining these suggestions with a critical stance and applying the scientifically valid aspects in practice. Previous work has established that the use of AI in healthcare should comply with predetermined consensus standards, such as evaluating different levels of risk in the implementation of AI, as described by the AI Act of the EU Commission, and guiding principles for AI use in healthcare defined by the World Health Organization [33].
Moreover, there is a need to develop a more nuanced understanding of the problems arising from the use of AI at the clinical level, for which qualitative research with practicing specialists will be helpful [34]. A previous study among nurses reported that, although AI suggestions may be perceived as time-saving, ethical principles were not detailed enough [35]. While our findings highlight that LLMs are able to make suggestions in accordance with the local context, some of the suggestions made by ChatGPT 4.0 were more concise and lacked depth compared to those made by GROK 3 and Gemini 2.5 Flash. We nonetheless believe that ChatGPT’s style of guidance may be more appropriate for fast-paced clinical practice in crowded places or peak times.
One critical consideration would be whether the use of AI as a support in clinical decisions would cause deskilling among physicians in terms of ethics [36]. Our findings suggest that LLMs can encourage physicians to participate in decision-making with appropriate use of language, thereby maintaining their ethical responsiveness alive. A particularly important implementation issue is whether AI’s reminders of ethical norms in an idealized manner are compatible with the realities of clinical practice. Some of the suggestions by LLMs may contradict current practice patterns, raising concerns about feasibility. For instance, previous work reported that pediatricians in Turkiye have limited practice in obtaining informed consent from their patients [37]. This highlights a potential if AI tools are to be helpful in guiding ethics, they must not only recall principles but also be contextually adapted to the actual clinical and cultural environment in which decisions are made. Ethical consultation requires not only cognitive accuracy but also emotional and contextual awareness. Since the LLMs lack real-life experiences, it would be impossible for them to achieve human-level insight into these factors. Hence, they cannot replace human-based support systems such as clinical ethics consultation, ethical committee guidance, or legal consultation. For these reasons, LLMs’ support to physicians in ethical decision-making should be viewed as thinking companions rather than decision makers, tools that provide explainability, reasoning, and ethical clarity.
Previously, the METHAD system, an algorithm developed for ethical decision making in clinical practice, emphasized ethical principles with little attention to cultural adjustment [38, 39]. In contrast, the LLMs’ responses in our study addressed justice alongside other ethical principles and included suggestions in accordance with local regulations. This capability of developing strategies fitting the local ethical and legal frameworks in Turkiye, especially by GROK 3 and ChatGPT 4.0, is remarkable for cultural adjustment. In this respect, our study is one of the few showing the cultural adaptability of AI support. However, AI’s support for clinical decision-making may not always express cultural and emotional awareness [22]. For instance, Gemini 2.5 Flash’s duty-focused approach left empathic contributions insufficient for some case scenarios in our study. This observation suggests that LLMs may struggle to capture cultural and emotional nuances, which are crucial for supporting clinicians.
Medical education places less emphasis on communication with patients compared to the diagnosis or treatment of their diseases. However, communication with the individuals we care for is of the utmost importance. Ideally, physician-patient communication reflects a much deeper interaction and represents a critical process in learning about patients’ personalities, families, and social backgrounds. The physician expected to find a solution for a serious illness should be experienced and ethically bound to values, in addition to having good communication with the patient, relatives, and the healthcare team [40]. Regarding the use of AI tools, communication should also be effective at the social level. Our findings related to providing a balance between the team, regulating patient-physician-family relationships, and reducing the clinical loneliness of the physician align with the previously reported empathetic and trust-based communication frameworks [41, 42].
Professionalism in the case of a physician includes a therapeutic relationship based on competent and compassionate care principles that meet the patient’s expectations and benefits. This relationship, based on ethical foundations, requires prioritizing the patient’s interests over the physician’s, determining and maintaining the standards of competence and integrity, and giving professional advice on health to the public [43]. A study in the area of education reported that AI tools exhibit deontological or pragmatic attitudes, while teachers adopted contextual and value-based ethical approaches [44]. While we did not observe such a difference in the clinical context, it was clear that GROK 3 had a more developed capacity for contextual and empathetic guidance.
A striking aspect of this study is that it establishes the emotional and contextual, in addition to conceptual, appropriateness of AI while guiding the clinician in ethically challenging cases. Specifically, GROK 3 employed a counsellor-style language in its suggestions to the physician, demonstrating that language and tone can make a difference in presenting ethical content in a more supportive manner. These discourse differences found in our analyses deepen the previously proposed “explicability principle” by showing its relevance at the level of human-machine interactions [21, 22]. Our findings extend the explicability principle beyond algorithmic transparency, showing that the tone and discourse style of LLMs also shape whether their guidance is intelligible and accountable to clinicians. In this sense, explicability emerges not only as a technical requirement but also as a relational principle in ethically charged CAP contexts. By offering empathetic and trust-building guidance, LLMs can provide physicians with meaningful support when they encounter clinical uncertainties in real-life practice.
Studies suggest that AI tools can offer an empathetic and supportive role in clinical communication in addition to providing medical knowledge [45]. This could make them valuable as communication support tools in addition to decision support tools. A previous work by Ngan et al. (2022) established that providing information and empathetic guidance can be presented together in a dual-layered support model [46]. In our study, all three LLMs made efforts to decrease the physician’s loneliness, to find a balance among the patient-physician-family triangle, and to offer suggestions sensitively to the local context. Our findings suggest that AI tools have the potential to serve as a guide in the future, supporting not only decision-making but also clinical communication and reducing professional burden.
The use of LLMs as clinical decision support systems can increase physicians’ efficiency in clinical decision-making, helping them make more conscious and correct choices. This support can only occur when patient information is processed and analyzed by the AI. In addition to decision-making efficiency for clinical practice, this process should take all ethical issues related to the patient into account, prioritizing ethical principles such as the patient’s benefit, wellness, dignity, confidentiality, and autonomy [47]. While AI-supported health services may be a practical solution for crowded and overburdened hospitals, they also pose ethical risks due to the need for private and sensitive patient information to function [48, 49].
Bioethics, law, and health policy play crucial roles in regulating the use of AI in healthcare. For this reason, different countries have made regulations, such as the recent revision of the National Bioethical Law of France, setting standards to address the rapid growth of digital technologies (e.g., AI, health data, telemedicine) in the health system. Accordingly, a principle of human supervision or human guarantee is needed on the use of AI in healthcare, which includes patient and clinician evaluations [50]. Standards of AI application in healthcare should always include a thorough analysis and management of the case at the ethical and legal levels [51].
Our findings also have practical implications for CAP clinical and policy contexts in Turkiye. While LLMs cannot replace human ethical consultation, they could be piloted in CAP clinics under institutional supervision to provide rapid ethical–legal guidance in urgent cases. Such pilot use would need to comply with the Law on the Protection of Personal Data (No. 6698), which sets strict boundaries on storing and processing health-related data. This means that any integration of LLMs into clinical workflows would require anonymization, robust data security, and institutional oversight to prevent breaches of confidentiality. Policymakers and professional bodies could consider structured frameworks where LLMs act as supplementary “thinking companions” while final decisions remain clinician-led.
Future research should evaluate real-world implementation, including (i) how LLM guidance interacts with clinicians’ ethical responsiveness, (ii) the acceptability of AI support among Turkish clinicians, patients, and families, and (iii) safeguards needed to ensure compliance with both legal requirements and cultural sensitivities. Comparative studies across jurisdictions would also clarify how varying definitions of adolescent autonomy and consent shape the applicability of AI decision-support tools.
Our analyses suggest that LLMs made efforts to adapt their ethical guidance to Turkiye’s sociocultural context, though with varying depth. Across scenarios, the models reflected local expectations such as protecting family harmony, preserving trust, and managing stigma, even when they did not name “culture” explicitly, as in the Case Scenario 1. Case Scenario 2 demonstrated the most pronounced divergence across LLMs, with some outputs acknowledging rural marriage practices, while others defaulted to more generic child-protection frames. Case Scenario 3 highlighted how all three models recognized stigma around sexuality and confidentiality as central concerns. Taken together, these patterns suggest that LLMs can approximate cultural adaptation indirectly, but their sensitivity remains uneven. While they sometimes captured nuanced family dynamics and stigma-related risks, they also produced overly general guidance. This underscores both the potential and the limits of current LLMs in addressing sociocultural realities and highlights the need for context-specific training and human oversight in ethically complex cases. However, the unevenness of these adaptations also raises concerns about whose experiences are represented or overlooked, a point we expand upon in the following subsection on algorithmic bias and justice.
Although our findings underscore the promise of LLMs in supporting ethical decision-making, it is essential to recognize their vulnerabilities to algorithmic bias. Large language models are trained on datasets shaped by prompts and patterns of digital text, which may not adequately represent the lived experiences of all populations. This creates a risk that AI guidance may reflect rigid, standardized frameworks, such as repeatedly invoking ethical principles and legal codes, while overlooking the contextual nuances required in complex clinical realities. Moreover, because the outputs in our study were elicited through specific clinical prompts, the results themselves inevitably reflect this prompt-dependence, highlighting how AI guidance can vary considerably with the way questions are framed.
Such limitations are particularly salient for marginalized adolescents. For instance, in scenarios involving same-sex attraction disclosed in conservative contexts, LLMs may risk pathologizing or minimizing adolescents’ experiences if different sexual orientation perspectives are underrepresented in training data. Similarly, rural youth or girls in forced marriage situations may be disadvantaged if LLMs’ outputs fail to capture the sociocultural and legal complexities shaping their vulnerabilities. Beyond these specific cases, our findings that some LLMs’ responses lacked empathy, cultural adjustment, or emotional nuance illustrate how algorithmic decision-making could unintentionally reinforce systemic inequities in clinical care. These risks underscore that algorithmic bias is not merely a technical limitation, but a central ethical concern directly tied to the principle of justice in bioethics.
To prevent AI guidance from inadvertently exacerbating inequities in clinical care, safeguards are necessary. These include ensuring greater diversity in training datasets, raising clinician awareness about the risks of algorithmic bias, and maintaining human oversight in ethically sensitive decision-making. Given that large language models have also been linked to harmful outputs, including responses related to self-harm and suicide, clinicians must approach AI guidance with caution. While AI can serve as a valuable “thinking companion,” it should never substitute for the complex judgment, empathy, and emotional nuance that only human clinicians can provide.
This study has several limitations, some of which stem from the authors’ conscious choices in conducting this research. First, the data analyzed comprises AI responses to case scenarios formed by the authors, rather than the actual case information. Due to this limitation, conveying the results to real-life practice may be partially difficult. However, considering the aim of the current study, the choice of using case scenarios was methodologically valid and meaningful. In addition, sharing sensitive information about real cases with LLMs would be ethically questionable. Instead, case scenarios were conceptualized with utmost similarity to real practice experiences.
The second limitation of the current research may be due to LLMs being continually updated; thus, the findings reflect the current versions of the programs available at the time of analysis. This may limit the reproducibility and generalizability of our findings over time. As ethical and clinical practices also change over time, future work should apply data generation with timely versions of LLMs. Although each case was presented in a newly initiated and anonymous session to prevent memory carry-over between scenarios, we cannot entirely rule out the possibility that model-level training biases shaped responses across sessions. While this precaution minimized contamination within our study, future research should continue to examine how platform-specific settings and underlying training data may influence outputs in ethically sensitive contexts.
Third, we did not conduct a frequency-based content analysis in the current study. Due to this, we did not quantitatively evaluate the frequencies of codes and themes. However, the purpose of the study was to systematically and in-depth evaluate LLMs’ ethical guidance potential, for which the qualitative analyses applied were appropriate and sufficient. In addition, while the analytic process was conducted reflectively to ensure reliability of the findings, there may be some level of subjectivity in the interpretation of findings based on the comments made by only two researchers. Lastly, including only three case scenarios limited the variation of themes addressed in the current study. Future work should incorporate a larger pool of case scenarios to include different ethical challenges.
With all limitations described, our study is novel in that it is one of the first qualitative explorations of AI guidance in ethically challenging case scenarios within the context of Turkiye. The strengths of this study include the combination of thematic and discourse analysis, a reflexive approach in the analytic process, and a multidisciplinary research team. In addition, analytic reliability was increased with strategies like independent coding. These methodological choices strengthen the validity of the findings and transferability to clinical practice, in addition to providing a holistic evaluation of AI’s potential to support clinicians in challenging situations.
The findings of this study indicate that ethical decision-support systems should be developed not solely on the basis of algorithmic accuracy but also in line with criteria such as explainability, contextual awareness, empathetic language use, and cultural adaptability. Analysis of three scenarios developed explicitly within the Turkish context demonstrated that LLMs provided support in ethical decision-making across four key (1) recalling and justifying core ethical principles, (2) offering legal guidance consistent with local regulations, (3) employing empathetic and therapeutic language, and (4) providing intellectual companionship that mitigates physicians’ sense of isolation.
The guidance styles of different systems varied considerably. While ChatGPT 4.0 placed greater emphasis on emotional support and cultural sensitivity, Gemini 2.5 Flash adopted a more structured and task-oriented approach. GROK 3, on the other hand, was characterized by a consultative role grounded in empathy and therapeutic orientation, while allowing decision-making space for the clinician. This diversity suggests that ethical decision support is shaped not only by technical accuracy but also by communication style, rhetorical strength, and normative awareness.
LLMs should therefore be positioned as a new generation of digital guides, entities that do not possess decision-making authority but think collaboratively with the clinician, offer justifications, and share emotional burden. Particularly in clinical contexts where physicians experience professional isolation, face time constraints, or lack access to formal clinical ethics consultation, the cognitive and emotional support provided by these systems may hold critical value. The present findings suggest that AI-based tools could be structured not only as decision-support mechanisms but also as resources that enhance clinical communication and contribute to reducing professional workload. However, the effectiveness of such support is directly linked to the LLMs’ capacity to foster empathy, openness, cultural sensitivity, and critical thinking.
In conclusion, ethical decision support attains true meaning not through standardized algorithms, but through context-sensitive, explainable, trustworthy, and flexible structures. Our findings suggest that implementation of LLMs in clinical practice must account for the gap between idealized ethical norms and the realities of healthcare delivery; therefore, systems should be designed to align ethical reminders with cultural, legal, and practical contexts. This study highlights the potential for LLMs to evolve into digital companions capable of providing ethically and legally informed guidance, not only in healthcare but across a wide range of professional domains in the future.
Supplementary Material 1.
Supplementary Material 2.
Supplementary Material 3.
Supplementary Material 4.
Supplementary Material 5.
Supplementary Material 6.
Supplementary Material 7.
Supplementary Material 8.