Chatbots are becoming more commonplace in our daily lives, especially in fields such as consumer marketing, customer support, education, and health care, and have significantly increased in recent years [1,2]. Chatbots enable users to have humanlike conversations on various topics. They can vary widely in complexity and functionality, ranging from simple information-giving chatbots to those using artificial intelligence to understand human language input [3]. There is a great diversity of chatbots available today, ranging in abilities, features, and complexity. A hierarchy of chatbots has been proposed [4,5], which classifies chatbots according to their ability to use algorithms or artificial intelligence or both to recognize the context of language as it is written in a real-time conversation, called natural language processing (NLP), and respond with greater specificity. Chatbots higher on the classification scale have more advanced NLP, promoting a greater likelihood of mimicking an actual conversation. The more advanced the chatbot, the greater the need for specialized technical expertise to build and maintain it.

Chatbot development platforms have come a long way, enabling those without technical expertise to use visually interactive approaches to develop chatbots and provide a simplified deployment approach that allows easy integration into social media platforms [6]. As technology advances, the opportunity for low- and middle-resource environments to move into the chatbot space increases. In their study describing emerging research needs in chatbots, Følstad et al [7] stated that a priority area is democratizing chatbots to all. Democratizing chatbots means removing barriers to entry, such as financial resources, technical resources, or specialized human resources, to help make chatbots a possibility for the wider global population. The goal of democratizing chatbots is to improve access to information, help reduce the digital divide between nations, and improve areas of public good [6,7].

An area that has seen expanded activity in recent years is chatbots for health care and health communication. Chatbots are increasingly used in health care to address various concerns, from simple to complex. Chatbots in the health space are typically domain specific, deployed for a particular area of focus [2]. Common uses for chatbots include simple tasks, such as providing tracking and reminders to support medication and appointment adherence [8,9]. However, more advanced chatbots are implemented to support and promote more complex health concerns, such as mental health support [10-12], smoking cessation [13], and promoting physical health and nutrition [9].

Chatbots that address these more complex health concerns often integrate a behavior change model into the conversation. This ensures that the chatbot does not simply deliver information, the least efficient way to impact health behaviors [14], but converses with the participant to maximize the opportunity for behavior change. These chatbots are often more advanced in functionality, such as NLP. The more advanced the chatbot, the greater the need for specialized technical and human resources, creating additional costs, which restrict chatbots to only affluent nations that can fund such projects. However, with the advent of simplified and low-cost development platforms enabling even nonprogrammers to build chatbots, there is an opportunity to democratize chatbots to all nations, including those with low resources. In these nations, health communication chatbots may make the most impact, potentially alleviating some of the burden on health care providers and health care systems to be the sole voices of outreach to public health. For example, studies have reported building a chatbot that asks diagnosis questions to help rule out or detect possible COVID-19 cases, thereby reducing the number of patients coming into primary care [15], or chatbots that can diagnose a disease and provide some information about it before consulting a physician, thereby reducing health care costs and providing medical information from a credible source [16,17]. Other chatbots aim to promote healthy lifestyles and public health education by delivering nutritional education [18] and continued care at home for geriatric patients after hospital discharge [19]. All these chatbots and more serve as virtual assistants ensuring patient care and education without burdening health care systems.

Our domain of focus is COVID-19 vaccine misinformation. The public’s ability to receive information, communicate their needs, connect with others, and mobilize community engagement are all factors that can impact the success of health communication initiatives [20]. The COVID-19 pandemic is an example of a worldwide impact in which social media was used to propagate misinformation regarding the virus and the vaccine [21]. Inaccurate and false information severely impacts public health, delaying individual health choices to take preventative measures, such as mask wearing or social distancing, and having broader impacts on vaccine uptake [22]. Even considering multiple personal and business accounts, a significant number of the population use and interact on social media [23]. Social media users increasingly use these platforms as information sources, rapidly consuming and sharing information [24]. Misinformation about COVID-19 has been prevalent on social media since the start of the pandemic [25], negatively impacting public trust in new COVID-19 vaccines and delaying, and even denying, uptake in various communities [22].

This study explores the feasibility of developing a chatbot using approaches that are accessible in low- and middle-resource settings. These approaches include using technology that is low cost, can be developed by nonprogrammers, and can be deployed over social media platforms to reach the broadest-possible audience. In addition, the technology used does not require a specialized technical team, uses freely available and accurate knowledge bases, and is developed using evidence-based practices to create a conversational model that integrates the potential for a change in health behaviors.

The paper answers the following research questions:

The Mohammed Bin Rashid University (MBRU) Institutional Review Board (IRB) approved this study (approval no. MBRU IRB-2021-67). Consent was obtained from the participants.

Low-resource environments face a myriad of challenges that can prevent them from entering the chatbot space. These include lacking available human, technical, or specialized resources. Another barrier of entry is the lack of exposed details about how to design a chatbot, an area not often elaborated on in published studies. This section reports on the considerations for the technical environment, the project team, the process of conversation design, and implementation on the technology platform.

We carefully reviewed several technology platforms using the following requirements: (1) low recurring monetary costs, (2) a simplified development interface that could be easily used by an individual with little to no technical expertise, and (3) cloud hosting to enable easy deployment and avoid the need for specialized hardware. As a result, we selected ManyChat [26], a cloud-based platform with an easy-to-use interface, simple and direct integration into social media platforms, and low and predictable recurring costs.

ManyChat simplifies conversation development using interactive visual displays of conversational decision trees, enabling users to drag actions and responses (see Figure 1). ManyChat also seamlessly integrates into major social media platforms (ie, Instagram, Telegram, WhatsApp, Facebook Messenger). We selected Facebook Messenger as our deployment platform because it is freely available and widely used worldwide and Google Sheets as a means to collect data from each conversation.

The core project team for this study consisted of 2 educational experts and 1 research assistant, with consultation advice from 2 health professionals with subject matter expertise in vaccinations and 1 researcher with a focus on computer science.

A conversation implemented in VWise consists of multiple-choice logic trees with preplanned answers and conversational branches that allow for some personalization of the conversation without the need for NLP. Figure 7 represents the different stages of a conversation. The first 2 phases (welcome and personalize) use rapport building to learn demographic details about the participant, including vaccination status, and share details about VWise as a character. The participant then identifies the area of concern to be addressed using buttons to represent each concern. Next, VWise addresses the concern by exchanging information, which includes gathering additional details about the participant, such as their perceptions about vaccination and the methods of how they typically receive information. Once the concern is addressed, VWise offers an opportunity to address another concern (Figure 8), and the conversation loops back until the participant responds “no.” VWise then asks whether to share further information in the form of a free online course, for which the participant can select “yes” or “no.” Finally, the conversation ends with understanding the participant’s perception of the chatbot and the influence of the conversation itself. Data from the conversation are stored and then written to a Google Sheet at the end of each stage.

Participant perception was investigated using the Chatbot Usability Questionnaire (CUQ) [43], a validated tool that assesses different aspects of usability: chatbot’s personality, onboarding, user experience, and error handling. The questionnaire consists of 16 questions, 8 (50%) related to the positive aspects and 8 (50%) to the negative aspects of chatbot usability. The CUQ was embedded in the chat, and scores were calculated using a Microsoft Excel spreadsheet provided by the questionnaire developers. The CUQ score is calculated out of 100. The CUQ developers have designed it such that the scores are comparable to the System Usability Scale (SUS) [44], where scores >68 are considered above average. The process data collected are further described in the Case Study section.

Next, we present a case study detailing the pilot for VWise.

This section describes how participants engaged with VWise during the pilot, including their journey through the conversation, concerns selected, and points of attrition.

This study sought to address the possibility of democratizing chatbot access to all by exploring the feasibility of developing a chatbot for public health communication using readily available resources and technology that would be accessible in low- and middle-resource settings. We explored this through a detailed description of the design and development considerations for our chatbot and by presenting a case study describing our initial pilot with 33 people who engaged in a conversation about COVID-19 vaccine misinformation.

Our study highlights that even in low-resource environments, the ability to develop a functioning and low-cost chatbot is feasible. A high level of engagement with the chatbot was demonstrated by the large number of participants who stayed with the conversation to its natural end and requested to see the free online resource, who selected to view all information about a given concern, and who returned to have a dialogue about a second concern. More than half of the participants (52%) continued the conversation till the end, and around 36% went for a second chat. These numbers are promising and suggest that with further improvement, the retention rate could be high.

Participants’ responses to the CUQ were positive. Emailed comments from participants revealed a need to work on the conversational flow as well as the chatbot’s personality (ie, extra friendly, a funny character). The discrepancy in the reception to the chatbot’s personality could be attributed to the cultural diversity of the participants. Nißen et al [46] recently noted that demographic differences, especially age, may be a determiner of receptivity to a chatbot, helping determine whether a bond between participant and chatbot might be formed. An emerging area of exploration is the development of personality-adaptive chatbots, in which the chatbot’s personality is tailored to an individual user based on key characteristics [47]. Future work on VWise could investigate whether the ability to create personality-adaptive chatbots is conducive to environments in which specialized resources are scarce.

Although behavior change was outside the scope of this study, we observed participants’ responses that could be construed as willingness or resistance to change. These observations could not be processed in real time (ie, only through post hoc manual coding), so we believe that the results, at the minimum, demonstrate that our conversation flow is promising and stimulates thought processes. This could be a precursor to behavior change. Future iterations should seek to take advantage of the advanced features of ManyChat, such as keyword detection, designed to help simulate an NLP experience. Another avenue for exploration is to use a hybrid approach, in which non-NLP chatbots serve to conduct an initial consultant, helping identify candidates for further intervention. For example, Lee et al [48] followed a hybrid approach in which a low-tech chatbot was used as a mediator for patients to self-disclose mental health needs before approaching a mental health professional. In our case study, a wealth of information about participants was collected, making branching scenarios possible, in which specific participants are pushed toward professionals or educational interventions.

Finally, to help further the aim of democratizing chatbot access for all, we recommend that future studies expose and labor the design and development processes and technology choices for their chatbots to enable others to reproduce their work. When consulting the literature to guide this project, only a few studies elaborated on the process of designing a conversational model or included recommendations for a smooth user experience [10,49,50]. Industry is further ahead in this area and was used more as a guide for this study [33,51-53]. Research should seek to catch up to industry by sharing the best practices and processes through published, peer-reviewed work.

This early study explored the feasibility of an approach to developing chatbots in low-resource environments. However, this study has many limitations.

This pilot was not a controlled study. Convenience sampling was used, and our sample size was quite small, consisting primarily of fully vaccinated people and with fewer respondents to the CUQ. Therefore, due to the limited scope of this paper (ie, feasibility), our results are largely descriptive, with the CUQ results serving only to aid the research team in areas for improvement. Future iterations should include a larger and more diverse sample to help us obtain a better understanding of the effectiveness and of the improvements needed to the content and conversation design.

Concerning the subject matter and our approach to branched conversations based on vaccination status, our approach for fully vaccinated participants centered on information delivery. A by-product of addressing vaccine misinformation is to increase vaccination uptake, so fully vaccinated participants do not fall into this category. Chatbots for health communication deployed over social media cannot know ahead of time the characteristics of those who will use them. As such, it is important to undertake persona exercises to understand who your participants might be and include an intended outcome for each. Future studies should use an engagement strategy in which the goal is to empower fully vaccinated participants to share their experiences and information with others via social media, as well as exploring different social media strategies that might attract a greater diversity of individuals to engage with the chatbot.

Since behavior change was not in the scope of this study, integrating MI as a behavior change model needs to be explored in future iterations. MI was largely selected due to it being a well-published model in the literature about chatbots [10,54,55], but there are also those who use an eclectic approach [27] or develop and implement their own models [28-30]. Future work should include the exploration of other models of behavior change based on a larger and more diverse participant population. Additionally, a behavior change expert was not consulted in this study, and future studies would benefit from having this skill set on the development team.

We were also limited by not being able to conclude that any behavior changed; only indications of change and resistance could be detected, albeit post hoc. Previous research, such as Altay et al [56], has shown that behavior change can be detected using a chatbot; however, as the feasibility of democratizing development was the aim of this paper, future work is needed on how this might be achieved using low-cost technology solutions.

This study explored the feasibility of and design and development considerations for VWise, a chatbot created to enable a greater diversity of environments to enter the chatbot space by using readily available human and technical resources. Although our findings are descriptive in nature, our pilot of VWise shows promise with regard to whether low-resource environments can enter the health communication chatbot space. The early conversational model also shows promise as many participants followed the conversation to its natural end and many extended the conversation through selection of a second concern. However, improvements to the chatbot’s personality and conversation flow are needed and further pilots with a larger sample size and diversity of vaccination status are necessary. This study represents early work of a chatbot in its virtual infancy. We hope this study will help provide those who feel chatbot access may be out of reach with a useful guide to entering this space, enabling more democratized access to chatbots for all.