https://orcid.org/0000-0001-7010-5308
Figures
Abstract
Background
The digitalisation of healthcare has provided new ways to address disparities in sexual health outcomes that particularly affect ethnic and sexual minorities. Conversational artificial intelligence (AI) chatbots can provide personalised health education and refer users for appropriate medical consultations. We aimed to explore design principles of a chatbot-assisted culturally sensitive self-assessment intervention based on the disclosure of health-related information.
Methods
In 2022, an online survey was conducted among an ethnically diverse UK sample (N = 1,287) to identify the level and type of health-related information disclosure to sexual health chatbots, and reactions to chatbots’ risk appraisal. Follow-up interviews (N = 41) further explored perceptions of chatbot-led health assessment to identify aspects related to acceptability and utilisation. Datasets were analysed using one-way ANOVAs, linear regression, and thematic analysis.
Results
Participants had neutral-to-positive attitudes towards chatbots and were comfortable disclosing demographic and sensitive health information. Chatbot awareness, previous experience and positive attitudes towards chatbots predicted information disclosure. Qualitatively, four main themes were identified: “Chatbot as an artificial health advisor”, “Disclosing information to a chatbot”, “Ways to facilitate trust and disclosure”, and “Acting on self-assessment”.
Conclusion
Chatbots were acceptable for health self-assessment among this sample of ethnically diverse individuals. Most users reported being comfortable disclosing sensitive and personal information, but user anonymity is key to engagement with chatbots. As this technology becomes more advanced and widely available, chatbots could potentially become supplementary tools for health education and screening eligibility assessment. Future research is needed to establish their impact on screening uptake and access to health services among minoritised communities.
Author summary
This study explores how chatbots can help people assess their health and access medical advice, particularly focusing on ethnically diverse communities. Researchers surveyed 1,287 participants and conducted 41 follow-up interviews to understand attitudes towards using chatbots for health-related conversations. Most participants felt comfortable sharing sensitive and demographic information with chatbots, especially when anonymity was guaranteed. The study found that people’s willingness to share information depended more on their previous positive experiences with chatbots than on factors such as age, gender, or ethnicity. Participants valued chatbots for being accessible, non-judgmental, and private, making them especially useful for discussing sensitive topics like sexual health. However, concerns were raised about chatbots lacking human empathy and being unable to handle complex emotional issues. To improve trust, users recommended that chatbots provide clear explanations about data security, use simple and inclusive language, and offer translations in multiple languages. They also suggested that chatbots should provide clear steps for follow-up actions, such as booking appointments or ordering home testing kits. The findings suggest chatbots could supplement healthcare services by promoting education and screening, particularly for marginalised groups. Future research is needed to assess their long-term impact on improving health outcomes and addressing health inequalities.
Citation: Nadarzynski T, Knights N, Husbands D, Graham C, Llewellyn CD, Buchanan T, et al. (2025) Chatbot -assisted self-assessment (CASA): Co-designing an AI -powered behaviour change intervention for ethnic minorities. PLOS Digit Health 4(2): e0000724. https://doi.org/10.1371/journal.pdig.0000724
Editor: Yifan Peng, Weill Cornell Medicine, UNITED STATES OF AMERICA
Received: June 7, 2024; Accepted: December 12, 2024; Published: February 13, 2025
Copyright: © 2025 Nadarzynski et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Quantitative dataset available in repository: https://doi.org/10.34737/ww606 Qualitative data cannot be shared beyond individual quotations in the manuscript because participants did not give consent in terms of publication of their full transcript. The interviews contain sensitive data. This decision was made in accordance with the Westminster University ethics committee. Please contact the administrative coordinator of the research authority in Westminster University for requests: research-data@westminster.ac.uk.
Funding: This report is independent research funded by NHS AI Lab and The Health Foundation and it is managed by the National Institute for Health Research (AI_HI200028) to TN, DH, CH, CDL, TB, DR and IM. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Advancements in healthcare automation and artificial intelligence (AI) offer significant potential for improving public health outcomes. The use of AI in medicine has grown rapidly and now includes applications in drug development, disease diagnosis, health monitoring, digital consultations, and personalised treatment [1]. With the help of machine learning, deep learning, and natural language processing, medical algorithms have emerged, potentially allowing for more effective disease control using complex mathematical models [2–4]. AI can also improve disease surveillance, detect abnormalities in screening tests such as mammography or cervical cytology, and provide personalised health advice based on individual risk profiles and behavioural patterns [5].
Patient-facing AI systems such as chatbots or virtual agents that imitate human-to-human conversations have been shown particularly useful in health promotion and education [6]. Chatbots are capable of using natural language processing to recognise questions and respond with pre-determined clinically validated answers. They can also use behavioural algorithms as decision-making tools, symptom checkers or online triage systems. Unlike static websites, often rich in visual content, chatbots act as a message exchange platform, with their language, terminology and phraseology being key to their successful design and implementation. Although their effectiveness and safety need to be established through high-quality evaluations, chatbots could potentially contribute to improved knowledge about diseases, increased awareness of specialist healthcare services, such as vaccination or screening, and influence health beliefs, attitudes and perceptions, having an impact on motivation, health behaviours and subsequent healthcare utilisation [7]. To date, public health chatbots have been used for treatment and screening support, counselling and mental health support, behaviour change, self-management and improved health literacy [8]. Recently, more modern chatbots based on large language models, such as ChatGPT, have the potential to transform healthcare delivery and influence individual knowledge and behaviours. The role of healthcare and psychologist researchers involved in the design of health chatbots is to understand their mechanisms of action, optimise their articulacy and fluency as well as determine their influence on behaviours and cognition.
Chatbots could benefit people from marginalised social groups including gender and sexual minorities, as well as ethnic minorities that are disproportionally affected by poor health [9,10]. For example, in the context of sexual health, stigma, discrimination and prejudice have a substantial impact on the awareness and knowledge of sexually transmitted infections (STIs), sexual health screening and subsequent access to relevant healthcare services (e.g., HIV pre-exposure prophylaxis) [11–13]. Owing to embarrassment or shame, ethnic minorities are less likely to disclose sexual health-related information such as sexual orientation to healthcare professionals [14]. Consequently, they might not be offered STI/HIV testing, resulting in a missed opportunity for diagnosis, treatment and infection control. Chatbot-assisted interventions based on self-assessment may reduce these barriers and promote STI self-testing, potentially contributing to a reduction in health inequalities. Previous research has indicated a higher frequency of chatbot-led consultations about STIs, compared with hospital consultations [15] suggesting an appetite among the public for efficient, safe and reliable chatbot intervention in sexual health.
There are several potential psychological processes related to chatbot-assisted interventions. Multiple behaviour change techniques [16] could be applied within chatbots to facilitate the adoption of healthy behaviours and habits. For example, chatbots could help with goal setting by allowing users to set clear and measurable goals and work towards achieving them [17]. They can offer self-monitoring, encouraging users to track their health behaviours, increase their awareness of health-impairing behaviours and help them identify areas for improvement [18]. Chatbots could offer personalised feedback based on individual self-assessment, helping users better understand their health and motivating them to make positive changes [19]. They could also provide social support by offering encouragement, sharing success stories and connecting users with others working on similar health goals [20]. There is also a potential for chatbots to tailor messages to an individual’s unique needs and preferences, as well as engage in problem-solving to overcome barriers to healthy behaviours [21]. In addition to these techniques, the psychological effect of self-disclosure to chatbots per se may motivate users to engage in healthy behaviours by sharing personal information about themselves [22]. Studies have shown that self-disclosure to health chatbots has positively affected user attitudes towards these types of interventions [23]. Two meta-analyses also demonstrated ‘a question-behaviour effect’ in which the act of asking personal and health-related questions has a small but significant impact on behaviour change [24,25]. Therefore, conversational AI systems such as health chatbots that ask health-related questions and incorporate behaviour change techniques may result in desired health outcomes.
In this study, we aimed to identify how to design an effective and culturally sensitive self-assessment intervention based on the disclosure of health-related information to a chatbot. In particular, we wanted to identify the level and type of information that people from ethnic minorities were willing to disclose to health chatbots. Our objective was to understand what chatbot features were acceptable to a diverse group of users and to identify factors that should be considered in the design of chatbot-assisted self-assessments (CASA), particularly in the context of sexual health screening behaviours.
Method
Design
We employed a mixed methods design, combining a cross-sectional anonymous online survey and follow-up interviews with a subset of participants. We assessed the comfort of disclosing health-related information to chatbots to inform the development of future CASA interventions. The study was approved by the University of Westminster Research Ethics Committee (ref: ETH2122-0524/0561).
Participants
We aimed to recruit a diverse and representative sample, including individuals from religious and ethnic minorities, sexual and gender minority groups, and those with disabilities or chronic health conditions. The focus was on capturing the perspectives of “seldom heard” populations that are underrepresented in research. The target sample size was approximately 1,500 individuals over the age of 18 who reside in the UK, in order to align the results with the context of the UK National Health Service. Participants under the age of 18 and those residing outside the UK were excluded from the analysis.
Recruitment
Recruitment took place from January to August 2022 using a combination of community-based and online strategies to obtain a diverse sample. A community engagement approach involved contacting over 30 UK-based third-sector organisations with a request to circulate the study advert. We invited them to express views about automated online platforms for health advice, containing a URL and QR code that directed them to the online survey page on Qualtrics. Leafleting and poster-based approaches were also used within community settings, such as regions of London with higher representation of people from ethnic minorities (i.e., shops, community centres, and religious groups). A professional leafleting agency was also employed to distribute the study adverts in Manchester, especially in postcodes with higher ethnic diversity. Our partner organisation Positive East distributed leaflets amongst their networks across East London and via their sexual health and HIV testing outreach activities.
The advert was also distributed in university and higher education settings (i.e., university campuses, student union groups, and student association social media accounts). Similarly, paid advertisements via Facebook, Instagram and Prolific, and free adverts on the X Platform (formerly Twitter), were utilised to reach various groups such as those living in non-urban areas, sexual minority groups or specific social groups (i.e., trans and gender diverse). Face-to-face meetings with interested organisations were arranged. The study was also promoted via radio broadcasts and in publications directed towards specific ethnic minorities, such as Latin Americans living in the UK. Lastly, a snowball sampling approach was utilised with community champions and university students representative of different ethnic minorities to further distribute the study advert within their networks. Due to such varied recruitment methods, it was not possible to record the number of eligible participants or responses related to each recruitment approach. Three optional prize draws of £100 each were offered as an incentive for the completion of the survey.
Procedure
Participants who clicked on the study advertisement link were taken to an online survey on Qualtrics, which presented an information page and an electronic consent form. The survey consisted of demographic and attitudinal questions, such as opinions on health and sexual health chatbots, comfort level in disclosing information to chatbots, and expected reactions to chatbot recommendations for health screening. Additionally, to better understand participants’ risk of STIs, the survey included behavioural questions related to sexual health, such as the number of sexual partners and condom usage. For those unfamiliar with chatbots, a description of a health chatbot was provided in the supplementary material (S1 File). Upon completion of the survey, participants were redirected to a debriefing page that included contact information for sexual health organisations and a separate online form where they could opt-in to participate in the optional prize draw, as well as follow-up interviews by providing their email addresses.
The sample for the interview was purposefully selected to ensure diversity in terms of age, gender, ethnicity, and sexual orientation. Participants were selected based on a maximum variation sampling framework, based on their ethnicity, and were scheduled on a first-come, first-served basis. The interviews were conducted either in person at the University of Westminster or online through MS Teams, according to participant choice. The semi-structured interviews explored participants’ views on health chatbots using a topic schedule ( S1 File). During the interview, a demonstration of an existing sexual health chatbot – Pat (www.positiveeast.org.uk/chattopat) was conducted to allow them to interact with the technology. The participants were then asked about the different features of the intervention and their willingness to disclose health-related information. The 60-minute interviews were recorded and transcribed verbatim. Participants received a voucher worth £45 as a token of appreciation for their time.
Measures
The questionnaire was designed by the research team with an expertise in chatbot acceptability. The survey was consulted with our public and patient involvement group to validate its readability and comprehension. The questionnaire was also piloted on 10 members of ethnic minority groups where ‘think aloud’ technique was used to validate each question. The final survey consisted of six sections. The first section collected demographic information, including age, language spoken at home, gender identity, assigned sex at birth, presence of disability or chronic illness, education, and ethnicity [26]. These questions were taken from a set recommended by the UK Office for National Statistics [27]. Participants were then informed about health chatbots and asked about their experience with this technology, their awareness of chatbots and the likelihood of using one if available, to gauge acceptability. Next, a 10-item scale was presented to assess participants’ attitudes towards health chatbots. Sample items included statements such as “I would use a health chatbot if my doctor recommended it” and “Talking to a health chatbot would be frustrating”. Participants were asked to rate their agreement with these statements using a 5-point scale, ranging from “Strongly agree” to “Strongly disagree”. Afterwards, participants were asked about the frequency of their internet searches for sexual health information and their comfort level in discussing sexual health with healthcare professionals, using a 5-point scale ranging from “Very comfortable” to “Very uncomfortable.” They were then asked to rate their risk of getting STI compared to others of the same age, using a 5-point scale from “Much below average” to “Much above average.” The survey also included questions about the gender and number of sexual partners, to determine participants’ risk of STIs. A set of 21 items was used to gauge the comfort level of participants in disclosing specific information such as their name, ethnicity of sexual partners, and symptoms of a sexually transmitted infection. The five response options ranged from “Very comfortable” to “Very uncomfortable”. A higher score indicated a greater discomfort in disclosing information to the chatbot. Participants were then presented with four scenarios in which a chatbot made testing recommendations based on the answers they provided, such as “Based on your answers, you are at a higher/average/lower risk of STIs compared to people your age”. Participants were asked to choose one of seven anticipated responses, such as visiting a sexual health clinic, ordering a test, or taking no action. The final part of the survey included a 14-item scale that assessed participants’ attitudes towards sexual health chatbots. The items, such as “A sexual health chatbot could help me find information about STIs and HIV” and “I would not trust information provided by a chatbot about my sexual health”, were presented with five response options ranging from “Strongly agree” to “Strongly disagree”. The last question in the survey gauged the acceptability of a sexual health chatbot by asking participants, “If a sexual health chatbot was available today, how likely would you be to use it?”
The follow-up interviews searched deeper into participants’ reasons for sharing information and their attitudes towards health chatbots. Participants were asked to discuss ways to better establish trust in the technology, and increase engagement with the CASA intervention. They were also asked to voice any concerns they had about chatbots and their reactions to personalised health recommendations. The purpose of these discussions was to deepen our understanding of how to improve CASA as a potential psychological support tool for users making health-related decisions.
Data analysis
The research team conducted a thorough examination of the quantitative data to ensure accuracy and reliability. Inclusion criteria were checked and cases with no variation in item responses were removed. Descriptive statistics were used to understand the distribution of data and to determine if parametric assumptions were met. The 21 items related to participants’ comfort in disclosing information to chatbots were explored using Principal Components Analysis and Varimax rotation, which revealed three distinct sets of items: “sensitive health information”, “demographic information”, and “personally identifiable information”. Mean scores for each set were calculated as dependent variables for further analysis. A one-way ANOVA was performed to identify differences in disclosure between White British participants and those from minoritised ethnic groups. Correlation coefficients were calculated to understand associations between variables and linear regression models were used to identify predictors of disclosure for the three sets of items. Categorical predictors were analysed using t-tests or non-parametric equivalents, and only predictors with a significance level of p = .01 or higher were included in the models. ANOVA models were used to compare categorical groups or their interactions as appropriate.
The qualitative data were analysed by three researchers (TN, NK and DR) using Braun and Clarke’s thematic analysis approach [28]. The process involved familiarising with the interview recordings and data, and then transcribing all interviews into NVivo. The transcripts were coded using a pre-determined coding schedule and individually reviewed. Coding reports were generated containing relevant quotes grouped according to specific codes. The codes were then grouped and regrouped by two researchers to identify themes and subthemes, which were then discussed and organised into clear and consistent categories by the three researchers. Finally, the final themes and quotes were reviewed and agreed upon to ensure clarity and uniformity.
Public and patient involvement
This research involved engaging a public and patient involvement group at critical points, especially before finalising the research results. This involvement guaranteed that the analysis reflected real-world viewpoints and addressed issues related to the co-design of an AI intervention for people from disadvantaged social groups. It consisted of six members from minority ethnic groups, who actively assessed and endorsed the study’s discoveries during seven organized gatherings. They contributed to various aspects of the research, from analysing the data to enhancing the final concept of CASA. Their perspectives played a vital role in refining the proposed chatbot intervention, so that it can be optimised for people from ethnic minorities.
Results
Sample characteristics
The study included a sample of 1,287 participants (Table 1), of which 57.7% were women, and 59.7% were between the ages of 20 and 29. The sample was diverse in terms of ethnicity, with 28% identifying as White, 28.3% as Asian, 25.7% as Black, 8.5% as Mixed race, and 9.5% as other ethnicities. 78% of the participants were native English speakers, and 64% had a university degree or higher. 20.3% reported having a physical or mental health condition lasting more than a year, 86% identified as heterosexual and 21% had not been sexually active for the past 12 months. 72.6% considered themselves to have a low risk of STIs, and 65.2% were comfortable discussing sexual health with healthcare professionals. 76% were aware of chatbots, and 70% had previous experience with chatbots, with 26% having used them for healthcare. Participants had mostly neutral to positive attitudes towards health chatbots and sexual health chatbots, and over half of the participants expressed willingness to use chatbots for healthcare and sexual health.
Disclosure of information to chatbots
Most participants were “somewhat comfortable” disclosing information to chatbots (S1 Table). A within-subjects one-way ANOVA indicated that participants were more comfortable disclosing sensitive health information and demographic information than personally identifiable information F = 66.51, p = <.001. There was no significant difference across ethnic groups (S2 Table) in the level of discomfort disclosing sensitive private information F = (4, 384.62) = 1.47, p = .21; demographic information F = (4, 384.03) = 1.56, p = .18; or personally identifiable information F = (4, 376.58) = 2.07, p = .08).
The overall model for predictors of discomfort disclosing sensitive health information was significant R² = 0.43, F (9, 1,112) = 94.19, p < .001. Awareness of chatbots (β = .157, p = .04), previous experience with chatbots (β = .176, p = .01), comfort discussing sexual health with HCPs (β = .166, p < 001), sexual health chatbot acceptability (β = −.06, p = .02), and attitudes towards sexual health chatbots (β = .869, p < 001) predicted degree of discomfort disclosing sensitive health information.
The overall model for predictors of discomfort disclosing demographic information was significant R² = .26, F (5, 1,109) = 78.55, p < .001. Male gender (β = −.153, p < 001), previous experience with chatbots (β = .179, p = .01), attitudes towards health chatbot (β = .133, p = .01), and attitudes towards sexual health chatbot (β = .638, p < 001) predicted degree of discomfort disclosing demographic information.
The overall model for predictors of discomfort disclosing personally identifiable information was significant R² = .157, F (5, 1,121) = 41.73, p < .001. Attitudes towards sexual health chatbots (β = .738, p < 001) were the only predictor of discomfort in disclosing personally identifiable information.
Reaction to chatbot risk advice
There were various reactions to chatbot risk estimations (Table 2). In the ‘lower risk of STIs’ scenario, most participants (42%) would do nothing or order a home testing kit (20%). In the ‘average risk of STIs’ scenario, most participants would order a home testing kit (26%), call a sexual health clinic (20%), or visit a sexual health clinic (19%). In the ‘higher risk of STIs’ scenario, most participants would visit a sexual health clinic (51%), call a sexual health clinic (17%), or order a home testing kit (14%). In the ‘too early to test’ scenario, most participants would visit a sexual health clinic (32%), call a sexual health clinic (28%), or order a home testing kit (13%). Less than 5% of participants would either talk to a chatbot again or a friend/family member across all scenarios.
Qualitative results
In total, 41 survey participants took part in follow-up interviews (medium age = 27, age range: 19-60; 63% women; 70% heterosexual; 53% Black, 23% Asian, 17% Latin American/Middle Eastern/Mixed race and 7% White other). Four main themes (Table 3) were identified: “Chatbot as an artificial health advisor”, “Disclosing information to a chatbot”, “Ways to facilitate trust and disclosure”, and “Acting on self-assessment”. Key features of Chatbot-Assisted Self-Assessment (CASA) intervention for behaviour change are presented in Table 4.
Chatbot as an artificial health advisor.
Overall, the participants had a positive view of health chatbots. They perceived them as a convenient and quick source of information for their medical questions, as well as an engaging and interactive tool that could improve their understanding of illnesses, infections, health risks, and healthcare services. The participants agreed that chatbots could raise awareness about various health issues and improve knowledge about available clinical interventions and services. However, some participants were concerned that chatbots offered limited content, and were not comprehensive sources of information on all health-related topics. There was a divide in opinions regarding the ideal number of chatbots, with some participants preferring a single chatbot that covered a wide range of sexual and reproductive health topics, while others believed that multiple, topic-specific chatbots would be more accurate and precise. The latter group felt that a specialised chatbot for each topic, such as STI screening, vaccinations, or contraception, would provide more in-depth information.
The capabilities of health chatbots were compared to those of healthcare professionals. Limitations, such as the robotic, unnatural feel of chatbots were acknowledged. Emphasis was placed on the need for sensitive, empathic and compassionate language despite being a computer program. The framing of questions and response options was deemed crucial for inclusiveness and avoiding bias. Specifically, chatbots require plain language accessible to all users, including non-native English speakers, unlike healthcare professionals, who can tailor language to each patient. Participants were interested in using chatbots for sensitive topics such as STI symptoms and other taboo issues. Some saw chatbots’ lack of cultural awareness as a limitation, while others saw it as a fair aspect for all users.
Participants viewed chatbots as a suitable option for private conversations about sexual and reproductive health, which they considered difficult to have face-to-face with a healthcare provider. The advice from chatbots was perceived as non-judgmental and impartial due to their lack of human preconceptions. Chatbots were believed to be best used for anonymous interactions and personalised advice, where users could ask intimate questions and receive a response or recommendation for specific actions. Some compared chatbots to virtual ‘risk calculators’ or personal risk assessment tools in the form of online questionnaires or health surveys. Chatbots that asked a series of questions imitating a patient-doctor conversation were considered more personally relevant than those that only provided information. One participant framed the interaction with the chatbot as a “let’s talk about me” type of conversation where health-related personal concerns could be explored. Although chatbots were not seen as capable of diagnosing diseases, they were viewed as useful as a symptom checker or for assessing eligibility for preventive measures such as screening or vaccination. The majority believed that chatbots that asked personal questions for self-assessment were as accurate as the users’ responses to those questions. Thus, participants emphasised the need for chatbots to explain why each question is being asked to understand the relationship between the chatbot questions and their health status.
The main obstacles to the use of chatbots in healthcare were their comparison to human healthcare professionals. Although they were deemed suitable for sensitive and embarrassing topics, many participants felt that chatbots could not provide the reassurance and expertise necessary for a reliable and full consultation. As a result, participants viewed chatbots as most useful for health education and promoting adequate healthcare services, allowing users to make informed decisions about accessing healthcare. There were also concerns that people were unfamiliar with the technology and therefore unwilling to use it, particularly older adults or ethnic minorities who were more accustomed to in-person interactions. A general agreement was reached that most people lacked knowledge of how chatbots work and who, if anyone, was behind them.
Disclosing information to a chatbot.
Participants believed that the ideal number of questions asked by a chatbot should be between 10 and 20. Chatbots that asked only a few questions were viewed as too generic and less relevant to the user, resulting in inaccurate recommendations. If the chatbot involved too many questions, participants feared it would lead to drop-out and non-disclosure. If the chatbot was anonymous, participants were more willing to share personal health information, often feeling just as comfortable answering questions as they would with a healthcare professional. Some participants also felt more at ease discussing sensitive sexual and reproductive health topics with a chatbot compared to a human healthcare professional, particularly if they felt uncomfortable with the professional’s gender or religious background.
Participants expressed mixed feelings about revealing personally identifiable information to chatbots. While they were generally willing to discuss sensitive health topics anonymously, many felt uneasy about sharing personal details like their name or phone number. However, they felt that chatbots that asked personalised questions about health risks, such as the number of sexual partners or condom use, were more relevant and provided more accurate recommendations. Disclosing demographic information like age or sex was generally acceptable, but there was concern that questions about minority status could lead to distrust if not properly explained. Most participants agreed that chatbots should focus on behaviour related to health risks rather than demographic information to avoid further marginalization. Participants were comfortable discussing symptoms with chatbots, but acknowledged that users may not fully understand their symptoms. Finally, participants felt that sensitive topics requiring high levels of empathy, such as sexual abuse or sexual problems, were not suitable for discussion with a chatbot.
Ways to facilitate trust and disclosure.
The participants stressed the importance of certain design features in chatbots to improve their trust in the intervention and encourage more truthful disclosure of health information. They suggested that chatbots should present information about their creators, how they work, and the reliability and accuracy of their information, based on research or evaluation. A clear introduction or information page was seen as favourable, explaining the chatbot’s purpose, how it engages in conversation, what aspects of health it is designed for, and its limitations, such as the inability to diagnose diseases. The participants believed that setting clear expectations for chatbot usage would increase their confidence in the technology. This introduction should also include information about confidentiality, how user data is used and stored, and who has access to it.
Participants emphasised the importance of chatbots’ conversational ability and speed of response as key factors that set them apart from other health interventions such as static websites. They believed that a conversational approach makes the experience more engaging, and rapid responses were seen as more satisfying. However, a high frequency of messages from the chatbot was seen as overwhelming and not helpful for users with lower health literacy. Participants suggested that chatbots should have a polite and respectful personality, and should be able to initiate and end conversations in a manner that mirrors human interaction, using expressions like ‘hello’, ‘thanks’, ‘my next question is’, and ‘goodbye’.
The importance of using neutral and medically accurate language in the chatbot design was emphasised throughout the interviews. The chatbot should be appealing to individuals who may be uncomfortable talking to healthcare professionals by using positive, destigmatising, and normalising language when discussing behavioural health risks, such as sexual activity or regular health check-ups. The chatbot should show empathy towards users, offering contextual prompts for potentially difficult or embarrassing questions giving users the option to skip questions. Rather than mentioning the personal risk of disease, it was recommended the chatbot present information about “eligibility” for screenings, vaccinations, or doctor consultations. Participants strongly discouraged that chatbots inform users about disease risks. Additionally, to create a positive experience, chatbots should recognise the pleasure and enjoyment in various activities related to health and well-being rather than only discussing health risks. Chatbots should provide clear recommendations for action based on self-assessments, informing users what steps to take next, when to seek healthcare, and how to access it. Many participants preferred the chatbot to have an appointment booking feature if it recommends a professional consultation or health screening.
Acting on self-assessment.
There were mixed opinions and reactions about the chatbot’s recommendations for further action. Some were worried that users might disregard or reject the need for healthcare services, despite the screening recommendations, due to cultural and personal factors, and stigma perceptions. However, most participants believed that the chatbot could be a starting point for accessing healthcare services. Many said they would seek clarification from healthcare professionals if the chatbot recommended tests or treatments. Some participants felt that having a conversation with the chatbot, discussing personal and medical questions, could help prepare users for better consultations with healthcare professionals by promoting familiarity with potential questions. Some believed the suggestions from health chatbots would motivate them to schedule sexual health check-ups or utilise at-home STI screening options. Many felt they would respond positively to the chatbot’s suggestions but would also verify their consistency by re-engaging with the chatbot. There was a concern that the chatbot’s recommendations should not create any additional barriers for individuals already seeking screening or vaccinations. Thus, the suggestions should be presented positively and encouragingly.
Discussion
This study is the first to investigate the design features of a health chatbot for marginalised populations, specifically racialised minorities. Results show moderate acceptability and neutral-to-positive attitudes toward health chatbots, and a willingness to share demographic and health information as part of self-assessment. There was no significant difference in comfort levels among ethnic groups, indicating the potential usefulness of this intervention format in underserved populations. However, participants were less comfortable disclosing personally identifiable information, highlighting the importance of anonymity in chatbot designs. Previous chatbot experience, overall acceptability, and positive attitudes were the most significant predictors of information disclosure, suggesting that hesitancy to interact with chatbots could be due to the lack of familiarity, previous adverse experiences or negative perceptions of this technology. Design principles for chatbots in include privacy, transparency, conversationality, and multilingual translation for overall acceptability and engagement in diverse populations.
Our findings suggest that people’s attitudes, beliefs, and perceptions are more important than demographic factors such as gender, age, or ethnicity in determining their comfort level with disclosing health-related information to chatbots. This contradicts previous findings that younger people are more accepting of chatbots [29]. However, the COVID-19 pandemic may have had an impact on people’s attitudes towards digital health services, particularly conversational channels such as webchats [30]. Additionally, since this study focused on sexual health matters, the participants may have had previous exposure to digital interventions, representing a generally younger and more technologically literate population, which could explain relatively higher levels of chatbot awareness and previous use in our sample as compared with health conditions that affect older populations. Our studies support the findings that AI is initially viewed with caution and reservations, particularly by those with a limited understanding of health chatbots [31,32]. This aligns with the Diffusion of Innovation theory, which states that the spread of new technologies varies through cultures and depends on perceptions of relative advantages, compatibility, complexity, trialability, and observability [33]. Thus, moderate acceptability and neutral attitudes to chatbots are expected, especially amongst those with no previous experience with them.
Design principles for chatbot-assisted self-assessment
To promote positive attitudes and increase engagement with health chatbots among ethnic minorities, it is essential to clearly state their purpose and capabilities in accessible language. Chatbots must address specific problems and individual concerns related to trust and confidence before engagement is initiated. We found that transparency in how chatbots work and who is behind them is crucial for building trust, which is supported by other studies on AI [34,35]. Attitudes towards health chatbots may be influenced by concerns of bias or discrimination towards certain ethnic groups. Our study found that participants emphasised the importance of training AI algorithms using diverse and unbiased data. This is especially important for people from ethnic minority backgrounds who may need clear assurances that health chatbots are non-discriminatory. Positive experiences with health chatbots can foster trust among individuals who fear discrimination from traditional healthcare providers and subsequently increase their engagement with healthcare services. Therefore, developing an evidence base and guidelines for the ethical use of health chatbots for various minoritised communities is likely to increase trust, influence the initial experience with chatbots and promote positive attitudes. We emphasise that the co-production of Al-led health interventions, such as chatbots, involving diverse communities in terms of age, gender identity, ethnicity, socioeconomic status, religion and disability, needs to become a standard practice to eliminate any potential bias and optimise overall acceptability and engagement [36–38].
Participants in our study expressed concerns about the confidentiality of their data when using health chatbots. These concerns are well-documented and are likely to affect trust in this technology [39]. An explicit reassurance about data security could improve users’ willingness to disclose health-related information. Participants also expressed concerns about the disclosing demographic information, especially when it comes to ethnicity, as it could reinforce negative stereotypes. Previous research has highlighted that individuals who identify or have characteristics dissimilar to the majority population may experience marginalisation and a perceived social distance, which may affect their willingness to disclose information [40]. Thus, health chatbots should explain why they are requesting such information; users will be more likely to disclose this information if it leads to more tailored and specific information. Both anonymity and confidentiality must be considered, particularly for stigmatised conditions such as HIV or drug use, where users may be hesitant to provide personally identifiable information. These concerns may be higher in those from ethnic minorities who are more likely to distrust healthcare professionals [41]. Due to the risk that poorly designed chatbots may introduce an additional barrier to healthcare if users experience an adverse reaction, chatbot designers need to consider all aspects of marginalisation and intersectionality when sensitive questions are being asked.
Our research found a strong preference for the conversationality of health chatbots which could be understood as the ability to hold a natural-sounding conversation. This could help to create a more engaging and personalised experience for users and to build trust and rapport, making them more likely to engage with the chatbot and disclose personal information. Conversational chatbots can help to improve the user experience by providing more relevant and tailored information and support. There is a potential that large language models, such as ChatGPT, could improve the conversationality of chatbots while answering personalised questions about health and well-being [42]. However, more research is needed to evaluate these emerging technologies and ensure that they are ethically designed and implemented.
Strengths and limitations
We employed diverse recruitment methods to include underrepresented communities in AI research, which were represented in our sample. Our mixed-methods approach facilitated data triangulation and provided a more comprehensive understanding of issues related to CASA interventions [43]. Participants were asked to interact with a sexual health chatbot during interviews to ensure that opinions were based on experience rather than hypothetical scenarios. However, there were limitations to the study. Ethnic minorities are a heterogeneous and diverse population with a wide range of views on health chatbots due to cultural, linguistic, and other aspects of marginalisation. Thus, there were diverse views of both sexual health and AI. People of various cultures may formulate distinct and unique views on health chatbots, which may also differ according to the type of disease or condition as well as their perceptions of severity and stigma [44]. Future research should further explore attitudes among individual ethnic groups to identify those differences and explore the willingness to disclose health-related information to chatbots across several health conditions. In addition, individuals with limited access to technology and healthcare services may have been under-represented in our sample, despite our use of diverse recruitment methods. While we collected data on educational attainment, self-reported chronic conditions, and language spoken at home as potential indicators of marginalisation, future research should also consider measuring socioeconomic status and access to technology, such as smartphone ownership, to better understand those who may be most vulnerable to poor health outcomes. Our study highlights the need to establish methods for AI chatbots to serve those from lower socioeconomic backgrounds. In addition, the data were collected before the widespread advancements in large language models, and it is possible that since then, more people have become familiar with chatbots, such as ChatGPT, and their attitudes towards the use of this technology in health are different. As such, future research should monitor acceptability and hesitancy towards conversational AI in ethnic minorities to predict their engagement and uptake.
This research shows that CASA intervention is acceptable in ethnic minorities and can be used to engage in private conversations about sensitive and health-related issues. It highlights a strong preference for education around available services, such as screening or vaccination, based on self-assessment, rather than estimation of disease risk. People from ethnic minorities may be less likely to accept AI-led ‘risk calculators’ or risk prediction tools, and participants in our sample reacted differently to the chatbot risk appraisal of STIs, indicating that presenting ‘risk information’ may not lead to the desired action. To be effective, CASA interventions should provide simple but informative explanations of the role and function of health chatbots, their confidentiality and privacy, and accurate multilingual translations. Previous research demonstrated that ‘sensitivity signalling’ had a positive impact on user disclosure of personal information to chatbots [45]. Thus, to promote transparency and trust, chatbots for self-assessment should justify each health-related question and engage in ‘positive conversations’, offering encouraging and empowering responses and linking users with organisations that provide further support. However, more research is needed to understand the impact of health chatbots, using the CASA approach, on individual behaviours and healthcare services for each minoritised community to systematically evaluate this type of intervention. Guidelines for patient-facing AI-led health chatbots should consider using behaviour change theory within their designs, as outlined for other AI systems [46], so that the mechanisms of action for chatbots are sufficiently understood. Our study also highlights the importance of addressing concerns about bias and discrimination in developing health chatbots for ethnic minorities, as positive experiences with chatbots can foster trust and increase engagement with healthcare services.
In conclusion, our study offers evidence for a conversational intervention in which users assess their health via relevant questions posted by a chatbot. CASA interventions could reduce stigma and break the taboo around topics that require privacy. They can also allow users to prepare for medical consultations by explaining questions that healthcare professionals frequently ask. However, this type of intervention is unlikely to benefit people with poor access to technology or the internet. Therefore, we emphasise the need for outreach services to reach disadvantaged groups, ‘seldom heard’ individuals and those with unmet needs. Such services could use multilingual chatbots to reduce language barriers and increase awareness of healthcare services. Although more evidence is needed, we anticipate that a conversational intervention carefully co-designed with members of ethnic minorities may reduce some barriers to healthcare utilisation that are typically associated with how people perceive their health and whether they are aware of services that are available to them.
Supporting information
S1 Table. Means and standard deviations for comfort disclosing sensitive, demographic and personally identifiable information to chatbots.
https://doi.org/10.1371/journal.pdig.0000724.s002
(DOCX)
S2 Table. Means and standard deviations for comfort disclosing information to chatbots by ethnicity.
https://doi.org/10.1371/journal.pdig.0000724.s003
(DOCX)
References
- 1. Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Family Med Prim Care. 2019 Jul 1;8(7):2328–31. pmid:31463251
- 2. Jiang F, Jiang Y, Zhi H, Dong Y, Li H, Ma S, et al. Artificial intelligence in healthcare: past, present and future. Stroke Vasc Neurol. 2017 Dec 1;2(4).
- 3. Martinez-Millana A, Saez-Saez A, Tornero-Costa R, Azzopardi-Muscat N, Traver V, Novillo-Ortiz D. Artificial intelligence and its impact on the domains of universal health coverage, health emergencies and health promotion: an overview of systematic reviews. Int J Med Inform. 2022 Oct 1;166:104855.
- 4. von Groote T, Ghoreishi N, Björklund M, Porschen C, Puljak L. Exponential growth of systematic reviews assessing artificial intelligence studies in medicine: challenges and opportunities. Syst Rev. 2022 Jun 28;11(1):132. pmid:35761303
- 5. Panch T, Pearson-Stuttard J, Greaves F, Atun R. Artificial intelligence: opportunities and risks for public health. Lancet Digit Health. 2019 May 1;1(1):e13–4. pmid:33323236
- 6. Aggarwal A, Tam C, Wu D, Li X, Qiao S. Artificial intelligence–based chatbots for promoting health behavioral changes: systematic review. J Med Internet Res. 2023 Feb 24;25:e40789.
- 7. Pereira J, Díaz Ó. Using health chatbots for behavior change: a mapping study. J Med Syst. 2019 May;43(5):135. pmid:30949846
- 8. Wilson L, Marasoiu M. The development and use of chatbots in public health: scoping review. JMIR Hum Factors. 2022 Oct 5;9(4):e35882. pmid:36197708
- 9. Hayanga B, Stafford M, Bécares L. Ethnic inequalities in multiple long-term health conditions in the United Kingdom: a systematic review and narrative synthesis. BMC Public Health. 2023 Jan 27;23(1):178. pmid:36703163
- 10. Wayal S, Hughes G, Sonnenberg P, Mohammed H, Copas AJ, Gerressu M, et al. Ethnic variations in sexual behaviours and sexual health markers: findings from the third British National Survey of Sexual Attitudes and Lifestyles (Natsal-3). Lancet Public Health. 2017 Oct 1;2(10):e458–72.
- 11. Bonacci RA, Smith DK, Ojikutu BO. Toward greater pre-exposure prophylaxis equity: increasing provision and uptake for Black and Hispanic/Latino individuals in the U.S. Am J Prev Med. 2021 Nov 1;61(5 Suppl 1):S60–72. pmid:34686293
- 12. Small L, Beltran RM, Cordero L, Lau C, Shanur S, Miyashita Ochoa A. The invisibility of Black and Latina women in sexual health care: shifting from biological individualism to intersectionality. Cult Health Sex. 2023 Aug 1;25(8):1084–100.
- 13. Quinn K, Bowleg L, Dickson-Gomez J. The fear of being Black plus the fear of being gay: the effects of intersectional stigma on PrEP use among young Black gay, bisexual, and other men who have sex with men. Soc Sci Med. 2019 Jul 1;232:86–93.
- 14. Brooks H, Llewellyn CD, Nadarzynski T, Pelloso FC, De Souza Guilherme F, Pollard A, et al. Sexual orientation disclosure in health care: a systematic review. Br J Gen Pract. 2018 Mar 1;68(668):e187–96. pmid:29378698
- 15. Fan X, Chao D, Zhang Z, Wang D, Li X, Tian F. Utilization of self-diagnosis health chatbots in real-world settings: case study. J Med Internet Res. 2021 Jan 6;23(1):e19928. pmid:33404508
- 16. Michie S, van Stralen MM, West R. The behaviour change wheel: a new method for characterising and designing behaviour change interventions. Implement Sci. 2011 Dec;6:42. pmid:21513547
- 17. Prasetyo PK, Achananuparp P, Lim EP. Foodbot: a goal-oriented just-in-time healthy eating interventions chatbot. Proceedings of the 14th EAI International Conference on Pervasive Computing Technologies for Healthcare. 2020 May 18. p. 436–9.
- 18. Echeazarra L, Pereira J, Saracho R. TensioBot: a chatbot assistant for self-managed in-house blood pressure checking. J Med Syst. 2021 Apr;45(4):54. pmid:33723721
- 19. Hungerbuehler I, Daley K, Cavanagh K, Garcia Claro H, Kapps M. Chatbot-based assessment of employees’ mental health: design process and pilot implementation. JMIR Form Res. 2021 Apr 21;5(4):e21678. pmid:33881403
- 20. Skjuve M, Følstad A, Fostervold K, Brandtzaeg P. My chatbot companion-a study of human-chatbot relationships. Int J Hum-Comput Stud. 2021 May 1;149:102601.
- 21. Bendig E, Erb B, Schulze-Thuesing L, Baumeister H. The next generation: chatbots in clinical psychology and psychotherapy to foster mental health – a scoping review. Verhaltenstherapie. 2022 Dec 28;32(Suppl 1):64–76.
- 22.
Berg JH. Responsiveness and self-disclosure. Self-disclosure: theory, research, and therapy. Boston, MA: Springer US; 1987. p. 101–30.
- 23.
Neef C, Mai V, Richert A. “I Am Scared of Viruses, Too” - studying the impact of self-disclosure in chatbots for health-related applications. International Conference on Human-Computer Interaction. Cham: Springer International Publishing; 2022 Jun 16. p. 515–30.
- 24. Miles L, Rodrigues A, Sniehotta F, French D. Asking questions changes health-related behavior: an updated systematic review and meta-analysis. J Clin Epidemiol. 2020 Jul 1;123:59–68.
- 25. Wood C, Conner M, Miles E, Sandberg T, Taylor N, Godin G, et al. The impact of asking intention or self-prediction questions on subsequent behavior: a meta-analysis. Pers Soc Psychol Rev. 2016 Aug;20(3):245–68. pmid:26162771
- 26. Iqbal G, Gumber A, Johnson MR, Szczepura A, Wilson S, Dunn JA. Improving ethnicity data collection for health statistics in the UK. Divers Equal Health Care. 2009;6(4):267–85.
- 27. Office for National Statistics. Primary set of harmonised concepts and questions. 2021. Available from: https://webarchive.nationalarchives.gov.uk/ukgwa/20120504111926/http://www.ons.gov.uk/ons/guide-method/harmonisation/primary-set-of-harmonised-concepts-and-questions/index.html
- 28. Braun V, Clarke V. Using thematic analysis in psychology. Qual Res Psychol. 2006 Jan 1;3(2):77–101.
- 29. Young AT, Amara D, Bhattacharya A, Wei ML. Patient and general public attitudes towards clinical artificial intelligence: a mixed methods systematic review. Lancet Digit Health. 2021 Sep 1;3(9):e599–611. pmid:34446266
- 30.
Shankar N, Moid S, Beena F, Shukla VK. Pandemic-induced behavioral change in mobile banking adoption: an opportunity for Indian banking industry for embracing artificial intelligence. Computational intelligence: select proceedings of InCITe 2022. Singapore: Springer Nature Singapore; 2023 Feb 16. p. 167–78.
- 31. Nadarzynski T, Miles O, Cowie A, Ridge D. Acceptability of artificial intelligence (AI)-led chatbot services in healthcare: a mixed-methods study. Digit Health. 2019 Aug;5:2055207619871808. pmid:31467682
- 32. Nadarzynski T, Lunt A, Knights N, Bayley J, Llewellyn C. “But can chatbots understand sex?” Attitudes towards artificial intelligence chatbots amongst sexual and reproductive health professionals: an exploratory mixed-methods study. Int J STD AIDS. 2023 Oct;34(11):809–16. pmid:37269292
- 33.
Rogers EM, Singhal A, Quinlan MM. Diffusion of innovations. An integrated approach to communication theory and research. Routledge; 2014 Apr 8. p. 432–48.
- 34. Felzmann H, Fosch-Villaronga E, Lutz C, Tamò-Larrieux A. Towards transparency by design for artificial intelligence. Sci Eng Ethics. 2020;26(6):3333–61. pmid:33196975
- 35. Silva GR, Canedo ED. Towards user-centric guidelines for chatbot conversational design. Int J Hum–Comput Interact. 2024 Jan 17;40(2):98–120.
- 36. Raza MM, Venkatesh KP, Kvedar JC. Promoting racial equity in digital health: applying a cross-disciplinary equity framework. NPJ Digit Med. 2023 Jan 11;6(1):3. pmid:36631555
- 37.
Belen Saglam R, Nurse JR, Hodges D. Privacy concerns in chatbot interactions: when to trust and when to worry. HCI International 2021-Posters: 23rd HCI International Conference, HCII 2021, Virtual Event, July 24–29, 2021, Proceedings, Part II 23. Springer International Publishing; 2021. p. 391–9.
- 38. Barony Sanchez RH, Bergeron-Drolet L-A, Sasseville M, Gagnon M-P. Engaging patients and citizens in digital health technology development through the virtual space. Front Med Technol. 2022 Nov 25;4:958571. pmid:36506474
- 39. Pomey M-P, Brouillard P, Ganache I, Lambert L, Boothroyd L, Collette C, et al. Co-construction of health technology assessment recommendations with patients: an example with cardiac defibrillator replacement. Health Expect. 2020;23(1):182–92. pmid:31691439
- 40. Phillips KW, Rothbard NP, Dumas TL. To disclose or not to disclose? Status distance and self-disclosure in diverse environments. Acad Manage Rev. 2009;34(4):710–32.
- 41. Armstrong K, Ravenell KL, McMurphy S, Putt M. Racial/ethnic differences in physician distrust in the United States. Am J Public Health. 2007 Jul;97(7):1283–9. pmid:17538069
- 42. D’Amico RS, White TG, Shah HA, Langer DJ. I asked a ChatGPT to write an editorial about how we can incorporate chatbots into neurosurgical research and patient care. Neurosurgery. 2023 Apr 1;92(4):663–4.
- 43. Fielding NG. Triangulation and mixed methods designs: data integration with new research technologies. J Mix Methods Res. 2012 Apr;6(2):124–36.
- 44. Miles O, West R, Nadarzynski T. Health chatbots acceptability moderated by perceived stigma and severity: a cross-sectional survey. Digit Health. 2021;7:20552076211063012. pmid:34917391
- 45. Carmichael L, Poirier S-M, Coursaris CK, Léger P-M, Sénécal S. Users’ information disclosure behaviors during interactions with chatbots: the effect of information disclosure nudges. Appl Sci. 2022;12(24):12660.
- 46. de Hond AAH, Leeuwenberg AM, Hooft L, Kant IMJ, Nijman SWJ, van Os HJA, et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review. NPJ Digit Med. 2022 Jan 10;5(1):2. pmid:35013569