Introduction

Research involving human subjects must undergo ethical review before initiation to protect their welfare and rights. Each country has established guidelines, regulations, and ethics review committees to ensure compliance [1, 2]. In the medical field, international ethical guidelines such as the Belmont Report [3] and the Declaration of Helsinki [4] provide foundational concepts. These related ethical guidelines and regulations are important from the viewpoint of international and universal review, but the most crucial aspect is the quality of review by ethical review committees. In an effort to ensure the quality of ethics committee reviews, the World Health Organization and the Council for International Organizations of Medical Sciences have issued guidelines on review quality and practices [5]. In recent years, there have been accreditation systems for ethics committees, wherein a third-party organization evaluates their quality. For instance, in the United States, the Association for the Accreditation of Human Research Protection Programs (AAHRPP) oversees the Human Research Protection Programs [6]. In the Western Asia-Pacific region, the Forum for Ethical Review Committees in the Asian and Western Pacific Region plays a similar role [7]. In Japan, the Ministry of Health, Labor, and Welfare has introduced a certification system without exception [8]. Establishing an organizational system for ethical review and receiving certification from an external organization is highly beneficial for ensuring the protection of research subjects. Meanwhile, to address the quality issues of actual ethics committees, it is considered most important for the members of each committee to “provide a comprehensive opinion of ethical considerations for the protection of human subjects” [9–11]. This stems from a recent survey conducted by AAHRPP accrediting bodies, which raises questions about whether improvements to the system alone can sufficiently address the quality of the ethics review [12]. For instance, even when an ethical review takes place, it remains a formal process, and the absence of ethical opinions regarding human protection is a significant concern. Additionally, in some countries, including Japan, collaborative research involving multiple institutions is typically subjected to review by a single ethics committee [13]. Although a single ethics review streamlines procedures and consolidates opinions, ongoing debates question the quality of such reviews [11, 14]. To enhance opinions on ethics reviews, efforts are being made to elevate the quality of reviews by providing advanced education to ethics committee members and fostering opinion exchanges among committees at the national level. Nevertheless, in the current Japanese context, there are suspicions that the quality of reviews is declining in the humanities and social sciences due to disparities in opinions and review resources resulting from the excessive establishment of ethics committees, and the medical field is no exception [15]. Addressing this disparity in opinions among ethics committees is crucial for unifying the quality of ethical reviews. Therefore, as a pioneering effort to extract specific opinions, we conceived the idea of utilizing natural language processing in artificial intelligence (AI) to extract ethical opinions on research issues [16]. An AI capable of detecting such errors could be highly valuable, particularly in thoroughly discussing ethical considerations within a research protocol. Generative Pre-trained Transformer (ChatGPT), an interactive AI that underwent extensive language learning, has received attention in recent years. ChatGPT, developed by OpenAI, admits mistakes based on the results of a huge database and supervised machine learning. Although its accuracy is continuously reviewed, it indicates the potential to challenge incorrect assumptions [17]. Additionally, it can follow the content of a previous dialogue, answering questions without the need for repeated back-and-forth exchanges within a single dialogue [18]. Released in November 2022, ChatGPT’s AI dialogue system gained global recognition and was quickly employed in medical research. It facilitated professional discussions with experts on the biological role of computational systems in stem cell research [19]. ChatGPT also explored the use of rapamycin, a drug preventing organ transplant rejection, showcasing substantial expertise in understanding its effects and side effects and responding appropriately [20]. These observations suggest the potential use of generative AI in discussing ethical considerations in research. However, there are currently no reports that have queried ChatGPT about the ethical viewpoint of research and disclosed the results. Consequently, it remains intriguing to explore whether opinions from the standpoint of research ethics can be extracted.

To this end, the present study aimed to input summaries of various medical studies into ChatGPT and examine whether interactive AI can identify ethical issues in research.

Methods

In this study, we used ChatGPT (January 30 Version, OpenAI LP, https://chat.openai.com/chat) [21]. We chose this version as it was the latest available when the research commenced and was easily accessible with a simple account registration, making it widely applicable. Additionally, we select ChatGPT over other generative AI tools due to its extensive testing and numerous reports. We structured questions simulating research scenarios with ethical considerations and directed them to ChatGPT, adopting its responses as our results. When designing a research scenario, we encountered a processing limit of approximately 2000 words with ChatGPT-3.5, making it challenging to process documents of the same length as our actual research plan. Therefore, we decided to create a concise English version of ≤100 words, considering ChatGPT’s performance degradation with increasing word count and the need for Japanese language support. Consequently, for this study, we used ChatGPT for interactive exploration of ethical considerations in hypothetical research queries instead of adapting lengthy documents such as research plans. The questions asked to ChatGPT were divided into five main categories: “Q1. Learning status of ethical principles in medical research ethics by ChatGPT”; “Q2. Opinions regarding ethical considerations for patients with dementia and depression”; “Q3. Opinions regarding ethical considerations for invasiveness and risks to subjects”; “Q4. Opinions regarding ethical considerations for the requirement of informed consent”; “Q5. Regarding research ethics judgment by ChatGPT.” Detailed contents are provided below (questions entered into ChatGPT). To ensure the reproducibility of ChatGPT’s responses, we repeated the same questions multiple times and confirmed consistency in the answers. As this study is exploratory in nature and focuses on the content rather than external evaluation indicators for assessing the ethics and accuracy of ChatGPT responses, the results output from ChatGPT are tabulated. Two research ethics experts then engaged in open-ended discussions about the content, comparing and discussing it in light of previously published studies. These discussions included assessing whether ChatGPT responses were comparable to the human subject protection content reported in previously published research ethics papers. Furthermore, because this study did not involve personal information or human subjects, it falls outside the scope of Japan’s ethical guidelines. Consequently, the Kyoto University Graduate School and Faculty of Medicine Ethics Committee determined that ethical review was unnecessary.

Results

ChatGPT provided relevant answers to all 11 questions (Q1-1 to Q5-1). In response to Q1, ChatGPT indicated that it had studied the Declaration of Helsinki, the Belmont Report, and Japan’s unique ethical guidelines for medical research involving humans. It asserted its ability to propose ethical considerations based on these guidelines (Table 1). Furthermore, when interpreting Japan’s unique ethical guidelines for medical research involving human subjects, ChatGPT demonstrated an understanding that these guidelines encompass the Belmont Report, Declaration of Helsinki, and Council for International Organizations of Medical Sciences (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Learning status of ethical principles in medical research ethics by ChatGPT.

https://doi.org/10.1371/journal.pone.0311148.t001

In response to Q2-1 and Q2-2, ChatGPT emphasized the importance of informed consent, confidentiality and privacy, risk minimization, benefits, justice, deception, and data management in research involving patients with dementia and depression (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Opinions regarding ethical considerations for patients with dementia and depression.

https://doi.org/10.1371/journal.pone.0311148.t002

Additionally, ChatGPT emphasized the importance of avoiding the exclusive recruitment of participants from vulnerable populations in studies involving individuals with dementia, especially those who may have difficulty comprehending the associated risks (Table 2). In the context of research on patients with depression, ChatGPT recommended measures for mental health protection, screening tests during the informed consent process, and underscored the necessity for suitable support both during and after the study (Table 2).

In response to Q3-1, ChatGPT highlighted the importance of being attentive to the vulnerability of patients with dementia during lumbar puncture procedures. Furthermore, it emphasized the need to establish a suitable medical system to mitigate potential complications, such as headaches and infections (Table 3). Regarding Q3-2, risk minimization was addressed, with ChatGPT stating that the off-label use of anticancer drugs should be implemented carefully and that a monitoring system for side effects should be established (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Opinions regarding ethical considerations for invasiveness and risks to subjects.

https://doi.org/10.1371/journal.pone.0311148.t003

In response to Q4-1, as a standard practice, written consent was recommended, and ChatGPT provided responses to 14 items in line with international ethical principles regarding precautions for informed consent explanations (Table 4). For Q4-2, ChatGPT emphasized the importance of considering the subject’s comprehension and, in certain situations, obtaining informed consent from an authorized legal representative or family member (Table 4). Addressing Q4-3, ChatGPT explicitly listed headache, infection, bleeding, and allergic reactions as the most prevalent risks. Moreover, acknowledging the potential for other complications, it underscored the significance of informed consent (Table 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Opinion on ethical considerations for the requirement of informed consent.

https://doi.org/10.1371/journal.pone.0311148.t004

In response to Q5, ChatGPT was found to understand relevant laws and regulations related to research ethics and ethical review, allowing it to make proposals in accordance with research ethics principles; however, ChatGPT could not make judgments (Table 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Regarding research ethics judgment by ChatGPT.

https://doi.org/10.1371/journal.pone.0311148.t005

Discussion

In this study, we verified the learning status of ChatGPT regarding research ethics and tasked it with addressing questions about ethical considerations related to subject protection. Initially, when inquiring about ChatGPT’s understanding of the Belmont Report and the Declaration of Helsinki, it accurately provided the adoption years and history of revisions. Moreover, ChatGPT demonstrated familiarity with informed consent, privacy protection, protection of vulnerable groups, and Japan’s unique ethical guidelines. Although ChatGPT could answer questions based on these principles without error, it remains unclear whether it can apply ethical principles to diverse research scenarios and respond flexibly to subject protection opinions. Subsequently, when asked about ethical considerations for an observational study involving 50 patients with dementia, ChatGPT emphasized the importance of informed consent, confidentiality and privacy, risk minimization, beneficence, justice, deception, and data management. It also highlighted the need to minimize infection risks from blood collection and stressed the voluntary participation of vulnerable subjects. These responses were based on formal ethical principles, outlining essential elements for any research [3, 4]. Some answers delved into the risks of blood sampling and the imperative to maximize the benefits of diagnosing and treating patients with dementia. Notably, the varying levels of detail and abstraction in ChatGPT’s responses may result from machine learning, possibly reducing opinions to the minimum required [22]. Another study reported a phenomenon wherein ChatGPT responses were consistently repeated in a fixed form. It is plausible that it provided answers to questions related to this research with a certain degree of prepared responses [18]. Next, we assessed the responses to a questionnaire comprising more than 50 questions directed at 500 patients with depression, along with ChatGPT’s perspectives on research involving the analysis of brain function through MRI. An explicit viewpoint emerged, emphasizing the vulnerability of participants with depression and the necessity for studies to minimize potential harm to their mental health. Consequently, there was an identified need for participants to undergo screening, ensuring their ability to participate without causing serious harm, with adequate support during and after the research. Although there was no commentary on the sample size, we interpret these responses as establishing a link between the mental burden from the questionnaire and depression. Reports on ethical considerations for treating depression highlight the importance of assessing and supporting potential problems on a case-by-case basis [23, 24]. The significance of informed consent and confidentiality in research on mental disorders has been previously reported [25]. The role of informed consent in addressing psychiatric issues has also been a subject of questioning, particularly in the context of editorial responsibilities for psychiatric journals [26]. These findings align with the responses provided by ChatGPT. Furthermore, ChatGPT addressed considerations for a study involving lumbar puncture for dementia patients, expressing concern about the potential risks of headaches and infections. It emphasized the necessity of establishing a reliable medical system to mitigate these risks, addressing typical complications associated with lumbar puncture [27, 28], and we believe that these responses demonstrate a professional understanding of the risks involved. Additionally, in the virtual study on the off-label use of anticancer drugs in patients with pancreatic cancer, ChatGPT asserted the importance of monitoring side effects and striving to reduce associated risks. In essence, opinions focused on the establishment of a system capable of handling side effects and ensuring the thoroughness of free consent regarding the off-label use of pharmaceuticals. Interestingly, ChatGPT provided insights into the immunosuppressive drug rapamycin and discussed its range of indications and concerns in vivo when used off-label [20]. By subdividing questions within ChatGPT, it could offer expert insights into the drug’s side effects. Next, when inquired about the informed consent required for research on patients with dementia, ChatGPT asserted that, as a principle, written consent was recommended. It emphasized the establishment of a process for voluntary participation based on ethical principles. ChatGPT stressed that the content of informed consent should not only outline the risks and benefits of the study but also include the disclosure of conflicts of interest. This response indicates that ChatGPT likely had an accurate understanding of ethical review-related laws and regulations. However, the initial response (Q3-1) maintained a similar level of detail, lacking specific considerations for patients with dementia. Conversely, when directly inquired about special considerations for dementia patients, they suggested the necessity of explaining the research to legal representatives and family members for obtaining consent. These responses align with reported key factors in the clinical challenges of dementia [29]. The results were interactive, revealing more detailed answers with an increased number of questions. In essence, it was confirmed that ChatGPT can provide more accurate and comprehensive responses through detailed inquiries.

Collectively, these findings indicate that ChatGPT possesses an understanding of ethical principles and can provide fundamental ethical responses to a variety of questions. However, the impact of machine learning was evident, with some responses being extreme, overly focused on specific ethical considerations, or generally superficial. ChatGPT itself acknowledged the limitation, stating that it cannot comprehensively address ethical considerations but can offer insights following ethical principles. Additionally, the phenomenon of repeated answers was observed, which was prominently confirmed in the results of this study. This suggests that machine learning-driven streamlining of responses may lead to specialization or extreme viewpoints on ethical matters. Moreover, this study’s results depend on the specific version of the large-scale language AI model used. Therefore, if the language model’s training data or algorithm is modified in the future, responses could vary, posing challenges to reproducibility. Additionally, while providing accurate responses is crucial in research ethics, large-scale language models can inherit biases, necessitating caution to prevent AI from perpetuating existing biases or introducing new ones. It is essential to consider generative AI while understanding these concerns. On the other hand, in recent years, AI has been actively used to address medical ethics issues through natural language processing [30] and to extract genetic medical problems using large-scale language models [31], highlighting the importance of ethical consultation [32, 33]. Based on these existing papers and the results of this study, there is a strong need to explore research ethics consultation using generative AI, and we must consider its future applications carefully. However, the ongoing debate on AI’s role in decision-making emphasizes the ultimate need for human experts [34]. Moreover, AI-driven research and development demand thorough ethical considerations. In response to Q5, ChatGPT clarified its role, stating “We can propose in line with ethical principles, but we cannot make judgments as we are AI.” This underscores that ChatGPT is an AI tool, highlighting the necessity for human decision-making and judgment. Based on these outcomes, it is apparent that existing generative AI can address basic ethical considerations in research; however, caution should be exercised in interpreting such responses. In the future, we believe that AI could serve as a valuable tool for ensuring subject protection by accurately identifying ethical issues. However, this study has limitations as the ethical validity of ChatGPT was not thoroughly examined using an evaluation tool, making the current results only indicative of ChatGPT’s responses. Additionally, the questions in this study were posed in a hypothetical research situation rather than an actual research plan. Consequently, with questions limited to 50 words or about 3–4 lines of text, it remains uncertain whether ethical considerations can be extrapolated from a comprehensive text such as a genuine research plan. The current study also lacked a robust framework to assess the ethical depth and relevance of the AI-generated responses. For example, conducting a comparative analysis between responses from the Human Ethics Committee and those generated by AI and exploring how AI-derived insights are integrated with human judgment, are areas for future study. Therefore, as demonstrated in this study, relying on AI for ethical considerations in medical research may have operational, legal, and ethical implications and must be approached with caution. Nevertheless, in exploring specific ethical aspects of research, there is potential to interactively confirm ethical points using generative AI such as ChatGPT. In recent years, reports have emerged on using AI trained in ethical principles for clinical ethics decision-making. While AI can provide efficient decision-making advice, it is sometimes inferior to human judgment [35]. Therefore, careful consideration will be necessary for future applications. However, we aim to explore further uses of AI to protect human subjects by considering the learning capabilities of generative AI and rigorously verifying its performance.

Conclusion

In this study, we employed ChatGPT, a generative AI, to assess ethical issues within a specific research context. The results yielded basic responses generally aligned with international ethical guidelines; however, machine learning optimization revealed certain biased opinions. Although the use of generative AI for enumerating ethical issues is a novel approach, it suggests potential applications in protecting human subjects in the future. Future research should focus on evaluating the ethical validity of generative AI and exploring effective ways to integrate AI into the realm of research ethics.

Acknowledgments

The authors thank Nanoka Ueda for expert assistance with the preparation of this manuscript and research.