Can ChatGPT Help General Practitioners Become Acquainted with Conversations About Dying? A Simulated Single-Case Study

TL;DR

This study explores if ChatGPT can help train general practitioners to have meaningful conversations about dying with patients.

Contribution

It demonstrates that ChatGPT-4o can simulate a structured conversation about dying, offering a novel training tool for GPs.

Findings

ChatGPT-4o successfully created a seven-step scenario aligned with evidence-based communication models.

The AI avoided explicit terms like 'dying' or 'death', despite the topic being central to the conversation.

Cultural and religious aspects of dying were not addressed in the generated scenario.

Abstract

Background/Objectives: General practitioners (GPs) should be able to initiate open conversations about death with their patients. It is hypothesized that a change in attitude regarding discussions of death with patients may be accomplished through doctors’ training, particularly with the use of artificial intelligence (AI). This study aimed to evaluate whether OpenAI’s ChatGPT can simulate a medical communication scenario involving a GP consulting a patient who is dying at home. Methods: ChatGPT-4o was prompted to generate a medical communication scenario in which a GP consults with a patient dying at home. ChatGPT-4o was instructed to follow seven predefined steps from an evidence-based model for discussing dying with patients and their family caregivers. The output was assessed by comparing each step of the conversation to the model’s recommendations. Results: ChatGPT-4o created a seven-step scenario based on the initial prompt and addressed almost all intended recommendations. However, two points were not addressed: ChatGPT-4o did not use terms like “dying”, “passing away”, or “death”, although the concept was present from the beginning of the conversation with the patient. Additionally, cultural and religious backgrounds related to dying and death were not discussed. Conclusions: ChatGPT-4o can be used as a supportive tool for introducing GPs to the language and sequencing of speech acts that form a successful foundation for meaningful, sensitive conversations about dying, without requiring advanced technical resources and without placing any burden on real patients.

Full text

1. Introduction

Nyatanga [1] wrote in 2012 that while general practitioners (GPs) may wish to discuss death with patients, time constraints often make it difficult, and since palliative care represents only a small part of their clinical care duties, they may lack the necessary knowledge and skills in this area. Thus, not all GPs initiate open conversations about death since some may feel unprepared [1]. Another reason GPs avoid telling terminally ill patients about their approaching death is that they see themselves primarily as healers, and this role conflicts with the harsh reality of the patient’s condition, which can feel hopeless [2]. Additionally, ethnic differences between doctors and patients have been identified as barriers to discussing end-of-life issues [3]. Since dying patients want physicians to be truthful regarding their disease and treatment options [4], developing the right knowledge and communication skills is vital to improve this aspect of care and ensure a dignified death [1]. Improved communication between patients and physicians is linked to better outcomes, including less physical distress at the end of life, fewer high-intensity interventions, a lower likelihood of dying in the ICU, earlier and more frequent hospice referrals, and reduced medical costs [5].

Virtual patients provide a controlled and risk-free setting for learning, enabling repeated practice that is often challenging to achieve in real clinical training [6]. Thanks to recent advances in artificial intelligence (AI) and language technology, virtual patients powered by large language models are now available for educational use (e.g., the SOPHIE—Standardized Online Patient for Health Interaction Education—platform [7]).

Since most GPs receive limited training in serious illness communication, a conversational AI chatbot assistant could offer a valuable, risk-free environment to help them develop and refine their communication skills [8]. It is hypothesized that a change in GPs’ attitudes about discussing death with patients may be accomplished through doctors’ training, in particular with AI.

Launched on 30 November 2022, OpenAI’s Chat Generative Pre-trained Transformer (ChatGPT) has provided everyone with easy access to generative AI. Since then, it has garnered significant research attention, with 8731 entries on PubMed as of March 2025, and this number is expected to continue rising. In medical education, AI and generative language models can surpass medical candidates in exam performance [9] and can enhance learning by providing realistic simulations, virtual patients, personalized feedback, and improved evaluation methods and by overcoming language barriers [10].

Previous research has shown that ChatGPT’s language model can have a positive role in palliative care communication training [11] and can successfully assist in role-playing scenarios for breaking bad news in emergency medicine [12]. However, its use in GP medical education for discussing end-of-life conversations with patients has not been studied before.

Therefore, this study aimed to evaluate whether OpenAI’s ChatGPT-4o can be used to simulate a medical communication scenario in which a GP consults a patient dying at home.

2. Materials and Methods

OpenAI’s ChatGPT-4o was prompted to generate the following medical scenario: You are a family physician visiting a patient you know very well at his home. You are there to discuss the patient’s approaching death due to pancreatic cancer with him and his family caregiver, who is his daughter.

ChatGPT-4o was also informed that this conversation must follow seven steps: (i) conversation opening, (ii) situation assessment, (iii) needs analysis, (iv) goal setting, (v) action plan, (vi) review and summary, and (vii) conversation closing. These predefined steps were derived from a recently published evidence-based model by Felber et al. [13] for discussing dying with patients and their family caregivers (FCs).

The output was assessed by directly comparing each step of the conversation to the recommendations in Felber et al.’s [13] model for discussing dying, which served as the quality assessment for the simulated medical communication scenario.

The seven-step recommendations for end-of-life conversations by Felber et al. [13] have been fully described elsewhere [13]. Briefly, they ensure compassionate and structured communication between healthcare providers, patients, and FCs. It begins with a conversation opening, where contact is established and key concerns are understood. Situation assessment follows, evaluating the patient’s condition and acknowledging the likelihood of death in the coming days. Next, a needs analysis identifies the patients’ and FCs’ concerns regarding dying and death. Goal setting involves collaborative decision-making on treatment and care, ideally aiming for a dignified death while considering both hopeful and worst-case scenarios. Based on these goals, an action plan is developed, detailing care priorities, contingency planning, and key contacts. The conversation then moves to review and summary, ensuring information is clearly communicated and feedback is gathered. Finally, conversation closing allows space for questions, ensuring a respectful and supportive conclusion [13].

3. Results

The output data from ChatGPT-4o, along with its corresponding analysis, is presented in Table 1. ChatGPT-4o created a seven-step scenario based on the initial prompt. The comparison of each step of the conversation to the recommendations in the Felber et al. [13] model for discussing dying shows that the conversational AI covered almost all intended recommendations. However, two points were not addressed: ChatGPT-4o did not use words like “dying”, “passing away”, or “death”, though the concept was present from the start of the conversation with the patient, and cultural and religious backgrounds related to dying and death were not discussed.

4. Discussion

This single-case study successfully demonstrated that OpenAI’s ChatGPT-4o can effectively simulate a medical communication scenario in which a GP consults with a patient dying at home, addressing the needs of both the patient and his family caregiver by using a simple prompt and without the requirement of complex prompt engineering. Nonetheless, skilled prompt engineering may be useful in medical education [14], as it could facilitate more detailed conversations; however, this aspect was beyond the scope of the current study.

These findings are important because they demonstrate the potential of ChatGPT-4o as a supportive tool for introducing GPs to the language and sequencing of speech acts that form a successful foundation for meaningful, sensitive conversations about dying without requiring advanced technical means or prompt engineering knowledge. These results align with previous research conducted in the emergency setting that indicated that ChatGPT can be used to train physicians on breaking bad news to patients [12].

In end-of-life care, empathetic communication and responsiveness to the needs of both patients and caregivers are essential. ChatGPT has previously shown promising results by generating quality, empathetic responses to patient questions [15]. As a result, these AI tools can provide a safe place for GPs to practice and improve their end-of-life discussions without placing any burden on real patients.

This study had several limitations. First, only one interaction with ChatGPT was evaluated, and it is uncertain whether the success would be maintained with repeated interactions. Second, while following the seven steps from Felber et al. [9] helped guide ChatGPT in structuring the conversation, strict adherence may have missed opportunities to adapt fluidly to the scenario based on the patient’s needs. Third, ChatGPT was not instructed to consider cultural, religious, or personal beliefs about death and dying, which are highly individualized and may have influenced the final conversation output. Lastly, no comparison with different ChatGPT versions or other conversational AI chatbots was performed.

Nonetheless, this single-case study not only serves as an early-stage exploration before larger studies but also provides an in-depth analysis of a sensitive conversation. Additionally, it demonstrates the ease of creating simulated cases for GPs and aims to help them overcome the discomfort of speaking openly with their dying patients by gaining exposure to this topic and encouraging reflection.

In future studies, researchers could involve real patients and caregivers to provide feedback on simulated consultations, thereby evaluating the realism of AI interactions rather than simply comparing the responses to recommendations, as was performed in the present study. Additionally, it is important to examine whether the use of ChatGPT and other conversational AI chatbots in training correlates with measurable improvements in GPs’ communication skills over time in actual healthcare settings.

5. Conclusions

ChatGPT-4o can serve as a valuable tool to help GPs become acquainted with conversations about dying. It can provide structured support for role-playing sensitive medical discussions without requiring advanced technical resources or placing any burden on real patients. While ChatGPT-4o can assist in simulating these scenarios, it cannot replace human interaction, as only a GP familiar with the patient’s background can effectively integrate emotional cues, cultural sensitivities, and ethical responsibility into his/her interactions.