Introduction

The rise of large language models (LLM) is transforming the landscape of digital health, paving the way for innovative applications in medical research, diagnosis, and treatment1. In less than half a year, three major Artificial intelligence (AI)-based chatbots have been successfully released. It is known that Open AI's ChatGPT (hereafter referred to as ChatGPT-3.5) initially debuted in November 20222. Early February 2023, Microsoft officially launched the Bing Chat or Bing AI chatbot (hereafter referred to as Bing). Then, in March 2023, Google released Google Bard (which is currently known as "Gemini", hereafter referred to as Bard)3,4. These chatbots immediately gained a lot of attention for their ability to understand and generate human-like text based on large amounts of training data5,6,7. The transformational power of AI has been reported in various medical features, ranging from diagnosis capabilities to therapy planning and patient care8,9,10.

Meanwhile, equality in AI access and the quality of technological development should not be based on a country's location or economic level. The advancement of AI-based chatbots should be enjoyed fairly throughout the world. Technological advances caused uneven access in various parts of the world due to existing patterns of discrimination, inequality, and bias, which should be avoided11.

Identifying information inequities is crucial for public health as it ensures equal access to reliable health information, reduces disparities, and allows policymakers to support disadvantaged populations effectively. Overcoming these inequities will lead to improved health literacy, informed decisions, and health outcomes12,13. Several studies have found that unequal access to medical information is associated with poor health outcomes14 and unsatisfied patient experiences15,16.

Gynecologic oncology, relies heavily on accurate and timely information, as inaccuracies can be harmful. Geographical inaccuracies can exacerbate global health disparities. This study examines whether AI-LLM provides consistent, unbiased medical information across different countries/regions. The goal is to identify potential discrepancies in AI-generated medical advice, ensuring fair and reliable health information access worldwide. The insights gained could guide the improvement of AI systems, fostering global health equity and trust in AI technology. Understanding these differences is vital for advancing and adopting AI-based technologies in medical practice, ultimately improving patient outcomes and experiences globally.

Methods

Ethics

Approval from the ethics committee was not required since no patients were involved in our study.

Materials

We used AI-based chatbots: Bing (https://www.bing.com/), Bard (https://gemini.google.com/app), and ChatGPT3.5 (https://chatgpt.com/) in this study. The patient scenarios were based on published case reports17,18,19,20,21,22. The VPN we used was the SoftEther VPN client downloaded from https://softether-vpn-client.en.softonic.com/.

Study design

Three freely accessible AI-based chatbots (Bing, Bard, and ChatGPT-3.5) were investigated. Each chatbot in Indonesia, Nigeria, Taiwan, and the United States was presented with six patient cases (from now on referred to as “patient”) (Fig. 1). The patient case scenarios were publicly available published data and then adapted into vignettes and prompts. This step is processed into a question by adding a command sentence and a structured writing system. The selected patients are those with various ages, clinical stages, and histopathology results. We used The National Comprehensive Cancer Network (NCCN) guideline as the standard of care for endometrial cancer23, which is widely accepted as treatment guidelines in many countries and provides treatment principles for endometrial cancer. Then we provide the patient vignette to the chatbot. After that, the identical question was asked “What is the most appropriate adjuvant therapy in this case based on consensus from NCCN?” (Supp. 1).

Figure 1
figure 1

The research protocols. Based on public data, six endometrial cancer patient cases were examined as patient scenarios (1). Six patient cases were tested in Indonesia, Nigeria, Taiwan, and the United States using three AI-based chatbots (Bing, Bard, and ChatGPT-3.5) (2). Answers from each chatbot in the 4 locations were coded and randomized (3). A panel of ten gynecologic oncologists analysed and scored the responses from each chatbot (4). Data analysis and statistical computations were used to determine the importance and compare the results of each chatbot's responses in different places (5). Use diagrams and tables to present study findings and conclusions (6).

Full size image

The vignette of each patient was input in each chatbot and location only once. All questions were tested on 3 chatbots in 4 countries on the same date (Sunday, November 26th, 2023). All the input across four locations was completed within a 24-h period to minimize the potential variation from ever-evolving AI chatbots. The replies were immediately entered into a database for review by a rater team, which included ten well-qualified gynecologic oncologists. To avoid bias, the responses from the AI chatbots were coded and randomized before being transferred to the raters for scoring. The raters rated the responses without knowing which response was from which chatbot and from what location. To evaluate the output from the chatbots, we used 5 parameters including "relevance", "clarity", "depth", "focus", and "coherence"9,24,25,26 with a 5-point Likert scale (Table 1)27,28.

Table 1 Assessment parameters and scoring.
Full size table

Statistical analysis

Homogeneity Index of the raters: Pearson and Spearman correlation coefficient analyses were performed to study the homogeneity among scores assigned by the raters to the responses. The interpretation of correlation coefficients is as follows: values greater than or equal to 0.4 but less than 0.7 are considered moderate, values greater than or equal to 0.1 but less than 0.4 are regarded as weak, and values greater than or equal to 0.7 but less than 1.0 are indicative of a strong correlation29,30,31. Furthermore, we performed a one-way ANOVA test with Scheffe's post hoc analysis to evaluate the homogeneity of raters.

We investigated the performance of three AI chatbots in four countries/regions in providing medical recommendations for six patients. The medical advice was evaluated by ten experienced doctors in five parameters: (1) relevance; (2) clarity; (3) depth; (4) focus; and (5) coherence. The assessment parameters were scored by physicians on a Likert scale ranging from 1 to 5, where higher scores denoted superior performance. Subsequently, these evaluations were categorized into "poor," "fair," and "good" based on scores of 1–2, 3, and 4–5, respectively. Additionally, based on prior studies and recommendations, items responses were linearly converted from a 5-point Likert scale to a 0–100 scale (higher scores representing better performance). The total score was calculated by averaging the transformed values of the five parameters for subsequent analysis32,33,34,35.

In statistical analyses, to examine the differences in the performance of chatbots across four locations, we employed a Chi-square test to assess variations in the proportions of three categories (i.e., good, fair, and poor) across the four regions. Additionally, Pearson and Spearman correlation coefficients, as well as a one-way ANOVA test with Scheffe’s post hoc analysis were used to assess the homogeneity among scores assigned by each rater to all responses. To assess disparities in both the overall performance and each of the five parameters for the medical recommendations among the four regions, a one-way ANOVA test with Scheffe’s post hoc analysis was employed. Recognizing potential confounding due to patient condition heterogeneity and the subjectivity of doctor assessments, multiple linear regression was utilized to explore variations in the total score across the four regions. The regression model featured the total score as the dependent variable and the four regions as independent variables, with chatbot, patient, and rater as covariates. All statistical analyses were conducted using SAS (Version 9.4, SAS, Cary, NC, USA), with a significance level set at 0.05.

Results

Our preliminary findings suggest that AI chatbots may provide different responses to general inquiries based on the user's location. When a query about domestic violence was sent to the chatbot Bing from different countries, Bing's responses varied significantly (Supp. 2). This prompted us to raise concerns about potential disparities or inequalities in accessing medical recommendations across different countries/regions. To investigate this, we initiated a systematic, double-blind study to determine if chatbots respond differently to identical questions posed from various geographical locations.

We relied on the expertise of ten gynecologic oncologists to rate all the responses provided by the chatbots. To validate our results, we analyzed the rater’s scoring homogeneity. We found that most of the raters (8 out of 10) showed a correlation coefficient between 0.47 and 0.93 using Pearson's test (Fig. 2A), and a value of 0.39–0.91 using Spearman's test (Fig. 2B). Furthermore, we did a one-way ANOVA test with Scheffe's post hoc analysis to evaluate the homogeneity of raters (Fig. 2C). Except for raters 8 and 10 (p < 0.05), all other 8 raters showed no significant variation. These results indicate a moderate to high level of homogeneity among the raters, further validating our scoring. To examine if there is a disparity among the four locations, we analyzed the overall performance of the chatbots. Nigeria ranked first based on the combined performance of the three chatbots, with a significant first place (p < 0.001) over Indonesia, Taiwan, and the United States (Fig. 3A). Bing Nigeria performed significantly higher (p < 0.05) than Bing Indonesia, Taiwan, and USA (Fig. 3B). Similarly, Bard Nigeria performed significantly higher (p < 0.05) than Bard Indonesia and Taiwan, but not Bard USA. In addition, Bard in the USA performed significantly higher than Bard in Indonesia (Fig. 3C). The performance of ChatGPT-3.5 in Taiwan showed the highest score but was not significantly different from the other three locations (Fig. 3D). Taken together, these results indicate that the treatment recommendations provided by AI chatbots vary significantly by location.

Figure 2
figure 2

(A and B) Homogeneity of raters. (C) Distribution of overall performance among ten raters, with the p-value indicating the results of the one-way ANOVA test with Scheffe's post hoc analysis.

Full size image
Figure 3
figure 3

In (A), it displays the results of the multiple linear regression model (adjusted for chatbot, patient, and rater), where bars represent the regionally adjusted mean total scores, with error bars representing 95% confidence intervals. The numbers on the chart indicate the adjusted mean differences. Distribution of overall performance for AI chatbots by region: (B) Bing, (C) Bard, and (D) ChatGPT-3.5. In (B)–(D), bars denote the average total scores for the four regions, with error bars representing standard errors. The p-value in the figure indicates the results of the one-way ANOVA test with Scheffe's post hoc analysis.

Full size image

To further confirm our findings, we employed a VPN to examine if there are disparities in responses based on questioning location. The question was asked in Taiwan, without a VPN or with a VPN to different countries. ChatGPT-3.5’s responses from South Korea and Thailand with VPN differed dramatically from no VPN Taiwan results in terms of completeness and structured systematics (Supp. 3). These results indicate that by using a VPN to change location, one can obtain different responses provided by the chatbot, further supporting our results that response provided by AI chatbots vary by location.

To evaluate which chatbot’s response achieves the highest score, we analyzed the scores of the chatbots in the five individual parameters. Overall, among the four locations, Bard scored higher than ChatGPT-3.5 and Bing (Fig. 4A) by a significant gap (p < 0.001), suggesting that Bard provides better recommendations according to NCCN guidelines. The results showed that Bing in Nigeria had the highest ratings on all parameters substantially (all p < 0.001) compared to Indonesia, Taiwan, and the United States (Fig. 4B). A similar performance was seen for Bard Nigeria, which outperformed Bard in Indonesia (Fig. 4C). ChatGPT-3.5 in Taiwan achieved the highest score in the "depth", where it outperformed the other three locations significantly (p < 0.001) (Fig. 4D). A more detailed information table about the chatbot performance across five parameters by locations are shown in Fig. 5. These results indicate that the scores of the responses provided by AI chatbots vary by platform and Bard outperformed the other two chatbots in this specific study.

Figure 4
figure 4

In (A), it presents the results of the multiple linear regression model (adjusted for region, patient, and rater), where bars represent the adjusted mean total scores for the three AI chatbots, with error bars representing 95% confidence intervals. The numbers on the chart indicate the adjusted mean differences. Distribution of AI chatbot performance across five parameters by region: (B) Bing, (C) Bard, and (D) ChatGPT-3.5. In (B)–(D), radar charts summarize the 100-point scale scores for each parameter, with the p-value for each parameter indicating the results of the one-way ANOVA test.

Full size image
Figure 5
figure 5

Distribution of AI-based chatbot performance across five parameters by region: (A) Bing, (B) Bard, and (C) ChatGPT-3.5. The bar denotes the percentage of good, fair, and poor for the four regions. The p-value in the figures indicates the results of the Chi-square test.

Full size image

Discussion

This study shows that AI chatbots such as Bing, Bard, and ChatGPT-3.5 give different responses to the same question depending on the user's location. We believe the variations in responses observed in this study are not due to time differences36, as all questions were tested on the same date, within a 24-h window. The use of VPN in the study further supports that the different responses are indeed depending on geographical locations.

The variable performance of chatbots such as Bing, Bard, and ChatGPT-3.5 in various regions highlights the importance of considering local geographical and cultural factors in AI development. Our findings in this study will have several implications for future medical AI development:

Inequality of health conditions

These findings suggest that AI chatbots provide different endometrial cancer treatment recommendations based on geographic location. The observed differences in chatbot responses may exacerbate existing health inequalities. In regions where chatbot responses are less accurate or comprehensive, users may not receive important health information needed to make informed decisions. For example, if a chatbot provides inadequate advice regarding domestic violence or reproductive health in one region, this could lead to delayed or inappropriate action, negatively impacting health outcomes. The study's results highlight the necessity for standardizing AI chatbot responses to provide fair access to reliable health information globally.

Trust in AI to deliver health information

Disparities in the quality of responses from AI chatbots may undermine public trust in this technology. If users consistently receive inaccurate or incomplete information, they may become skeptical about the reliability of AI for health advice. This distrust may discourage people from using AI health tools, which are intended to improve access to medical information. Lack of trust in AI systems may hinder their adoption and effectiveness, ultimately limiting their potential to improve health literacy and outcomes. Addressing these inconsistencies is critical for building and maintaining trust in AI-based health solutions.

Challenges in implementing policies and regulations

The observed variability in AI chatbot responses poses challenges for policymakers and regulators. Ensuring that AI health technologies provide consistent and accurate information across different regions requires strong policies and regulations. However, developing and enforcing such standards can be difficult, especially in countries with diverse cultural and regulatory circumstances. Policymakers must acknowledge these complexities to establish regulations that support fairness and dependability in AI medical devices. Successfully addressing these challenges is critical to harnessing AI for equitable health access and improving public health outcomes globally. Public health stakeholders must be aware of these differences to reduce the potential for bias in AI health technologies.

In this study we also applied rigorous validation methods, including correlation and ANOVA tests, to ensure homogeneity of rater scores. The moderate to high agreement among the raters adds credibility to these findings. Validating the consistency of chatbot responses is critical to building trust in these AI systems.

We found that Bard outperformed ChatGPT-3.5 and Bing in our study in providing endometrial cancer adjuvant therapy recommendations. These advantages may be related to the algorithms used, the data train, or the way Bard processes and presents information. Chatbot performance is largely influenced by Natural Language Processing (NLP), Machine Learning (ML), and User experience (UX) considerations. In recent years, chatbot technology has advanced, making it more accessible and widely employed across a variety of businesses. NLP and ML are critical for making chatbots more conversational and human-like, while UX is required to ensure that users have a favorable engagement with the chatbot26.

Nevertheless, some limitations of AI-based chatbots are still need to be addressed. According to a recent study, a chatbot does not adequately represent the demographic diversity of medical disorders when creating clinical vignettes, resulting in stereotypical presentations37. AI can also generate inaccurate or nonsensical content, a condition known as hallucinations or confabulation38, implying that they might provide wrong answers, which is especially problematic for persons with low education or the capacity to judge the accuracy of information.

Several studies have compared Bing, Bard, and ChatGPT in a medical context, yielding various results about their performance. In a study on methotrexate use, ChatGPT-3.5, Bard, and Bing had correct answer rates of 100%, 73.91%, and 73.91%, respectively39. A comparative study in endodontics also found that ChatGPT-3.5 offered more reliable information than Bard and Bing40. Dhanvijay et al.'s comparative analysis, as well as a number of other studies with similar outcomes26,41,42. In other studies, Bard outperforms ChatGPT and Bing7,43,44. Interestingly, a Yemeni study found that Bing had the highest accuracy and specificity, outperforming Bard and ChatGPT-3.545.

Our study focuses on cancer, which is similar to Rahsepar et al.'s work9 and Sensoy's43 by using case scenarios in the form of vignettes such as Dhanvijay's et al.41. Our findings showed that Bard performed Bing and ChatGPT3.5 on the topic of endometrial cancer adjuvant therapy. The diversity in study results above suggests that each chatbot may have a distinct advantage in specific areas of the medical domain. In addition, since the chatbots are constantly evolving, testing at different times may produce different results.

AI-based chatbots are adapting and evolving constantly in response to user feedback and iterative training set upgrades. Although we limited our test to different locations within 24 h, it was very difficult to test the chatbots at the same time. This time difference may also cause some variance.

Our study was initiated when we accidentally asked two general questions (about domestic violence and a river incident, Suppl. 2) to Bing from different locations. We observed that the responses differed depending on where the questions were asked. This prompted us to consider the potential impact of such disparities when seeking medical information. The findings of our study suggest that such disparities in responses from AI-based chatbots extend beyond medical information to disparities in AI responses to general information related to people’s daily lives.

Based on the findings of this study, we urge AI developers to take on more social responsibilities and put in extra effort to mitigate disparities when creating AI products.We also urge AI industry regulators and policymakers to establish stricter policies for regulating the AI industry, in order to prevent AI products from exacerbating existing disparities. Proactive identification of disparities, targeted mitigation strategies, and increased transparency in model training and data sourcing are essential to ensure that AI benefits everyone, rather than just a few.