Health care interventions involving written communication play a pivotal role in disseminating critical information to patients and promoting positive health outcomes. However, the process of crafting effective health care content has historically been labor-intensive, time-consuming, and often lacks the necessary uniformity and transparency required for rigorous research and development.

We propose the application of generative artificial intelligence (AI) technologies to address the pressing need for efficient and transparent content creation in health care interventions. In particular, we focus on harnessing the capabilities of pretrained large language models (LLMs), which are sophisticated AI systems designed to understand and generate human-like text (refer to subsection Generative AI With LLMs). By using these rapidly growing technologies, we aim to assist researchers in the content creation process, making it more accessible, systematic, and adaptable. As a tangible example, we introduce the first publicly available data set of AI-generated brief messages tailored for individuals with type 2 diabetes, specifically targeting medication adherence, a critical aspect of diabetes management. Notably, our data set of 1150 messages also stands as the current largest data set of health care intervention messages publicly available. Furthermore, we make our source code replicable and accessible to the research community while providing a comprehensive breakdown of our design process. In doing so, we seek to use generative AI to pave the way for a new era of health care intervention content development, one characterized by transparency, efficiency, and scientific rigor. Our main contributions are as follows:

In the ever-growing landscape of health care, effective communication is essential to enhancing preventive measures and developing intervention strategies that improve public health outcomes. Each year, a great number of new intervention studies are added to the health care literature [1,2]. However, with the growth in the quantity of interventions often comes an increase in their technical complexity, especially in the area of mobile health (mHealth). Many of these interventions are delivered through proprietary apps or other nonstandardized platforms, which complicates their integration into future programs and often causes their results to be obfuscated by the unique specifics of their deployment. This not only makes them challenging to apply elsewhere but also ensures that their development is both time-consuming and resource-intensive [3]. In addition, while research into mHealth has boomed in the last decade, studies suggest that the overall success rate of most mHealth interventions is not exactly clear, despite the strong interest and obvious potential such interventions have [4,5]. Ensuring that these interventions are feasible, effective, and sustainable, is vital for preventing unnecessary research waste.

Within the sphere of mHealth interventions, there is growing evidence supporting the success of text-message–based programs (also known as SMS) in modifying health behaviors [6-9]. With >97% of Americans currently owning some type of cell phone and the prevalence of smartphone ownership having increased from 35% to 85% in the last 10 years [10], text messaging has become a staple mode of communication for most people in the modern world. Using a platform already embedded in most individuals’ routines, text messages eliminate the need for additional equipment or substantial behavior change. This universal reach and familiarity not only enhances patient engagement but also bridges the gap for underserved communities, thus playing a pivotal role in reducing health disparities [11-13]. Their omnipresent nature, readership and engagement advantage, and the ability to mirror the conversational tone of in-person counseling all underscore the unique value of text messaging in contemporary health interventions [14,15].

Content has been described as “the central driver of behavior change” in interventions [16], and its thoughtful incorporation through modalities like text [17,18], imagery [19,20], and other media [21,22], is key to effective intervention design. For brief message interventions in particular, textual content serves not only as a vessel for information, but also as the critical and emotional linchpin motivating behavior change. When we consider interventions designed to induce change, clarity in the content creation process becomes indispensable. It provides a coherent road map for both practitioners and researchers, ensuring that the outcomes of the intervention—successful or not—can be understood, dissected, and refined. Furthermore, a transparent process of content creation not only bolsters the effectiveness of an intervention but also builds trust within the broader scientific community, allowing for constructive critiques, replication of studies, and meaningful advancements in the field.

The conventional SMS intervention development pipeline, shown in Figure 1, consists of the following parts: 1) formative research into the problem setting, behavior, and target population; 2) the establishment of the chosen theoretical framework and development of content; 3) a necessary review of the created content for quality assurance, safety, and research standards, as well as a pretest to gauge initial user feedback on the messages; and finally, a revision of messages based on accumulated feedback [23]. In this study, we address the second step of intervention design—the creation of content within a scientific and theoretical framework—due to its complexity and implication for specialists outside the traditional research team. Content designers can be used to greatly enhance the quality and efficacy of content for health care interventions; however, their involvement is often limited due to monetary, time, and resource constraints on the research team. Consequently, researchers are frequently tasked with taking on the roles of content designers themselves.

However, despite the involvement of the research team in content design and the critical importance of content, there exists a conspicuous opacity surrounding the creation of content in health care interventions. Numerous published works fall short in delineating the intricacies of their content creation processes while simultaneously withholding disclosure of their final message data sets, leaving a void in our understanding of how preliminary findings or formative research become translated into the finalized intervention—an omission that has led to the development process being described as a black box [25]. This lack of transparency is especially concerning given the tendency of some researchers to view text messaging as the intervention itself rather than just the means of delivery [26]. In those instances where the message creation process is disclosed, it often reveals a narrow involvement, typically limited to a few individuals within the intervention team [23,27]. Their varied levels of expertise in content design and differing perceptions of what constitutes “good content” can, as a result, lead to vast inconsistencies in outcomes that could be mistakenly attributed to other metrics like participant demographics, study duration, message volume, or the theoretical techniques used instead of the more crucial variable: the nature and quality of the content itself. In addition, this limited participation in content drafting tends to perpetuate familiar methodologies, sidelining innovative approaches that could potentially address persistent challenges like medication adherence [28]. Such exclusions not only hinder academic progress but could also inadvertently reduce the efficacy of interventions. When the foundation—content creation—is not soundly built with a clear and shared understanding of its underpinnings, it runs the risk of diluting the potential positive outcomes of the intervention. As brief message interventions continue to increase in number, complexity, and scope, the need for innovative and transparent approaches to content creation grows even greater.

Faced with the complexities and opacity of content creation, generative AI offers a promising solution to unveiling this enigmatic “black box.” Recent advancements in the development of LLMs using the transformer architecture [29] have brought about a revolutionary change in natural language processing. Unlike earlier models that process text sequentially, transformer models use a technique known as “self-attention” to analyze and draw connections between different parts of input data simultaneously. By converting text into corresponding numerical representations called embeddings, transformers can process language data with exceptional accuracy and speed. Furthermore, being pretrained on vast corpora of web-text data, these LLMs are not only equipped to simulate human conversations but also excel as versatile tools across a spectrum of nuanced tasks, such as question answering, writing support, translation, coding, and more [30-37].

Though the concept of data generation using LLMs is not novel in itself [38-40], the accessibility and enhanced generative capabilities of contemporary large-scale pretrained models like those in the GPT series have magnified their impact and broadened their potential applications [41,42]. With up to hundreds of billions of parameters [32,43], these models excel at rapidly generating vast quantities of contextually appropriate content, streamlining the traditionally painstaking process of manual drafting while simultaneously enhancing adaptability across diverse domains and sectors.

An integral aspect of effective LLM use lies in the art and science of prompt engineering. A prompt is any input given to an LLM that influences the nature of the LLM’s output [44]. Prompts are often given as sets of instructions or requests that establish the rules and guidelines of the conversation. Through prompt engineering, the context of the conversation can be strategically structured to direct the LLM to process relevant information and shape the desired form and content of its output [45]. This process is pivotal in refining and enhancing the capabilities of generative models and allows for the generation of more precise and relevant responses, which is especially imperative in complex fields like health care where the accuracy of information is essential.

Prompt engineering for LLMs is appropriate for the preliminary design of health care interventions for many reasons. First, LLMs can rapidly generate vast amounts of content, effectively reducing both the time and costs typically required for intervention development. This efficiency may allow researchers to allocate resources more appropriately, diverting their energies toward other critical aspects of the project while enabling the exploration of diverse content approaches that were previously considered daunting or impractical. Moreover, LLMs serve as a vital aid to researchers who may not have an extensive background in content design. By providing large amounts of well-written, contextually tailored content, these models offer a structured foundation that researchers can then build upon and further customize during the content review process while avoiding the overwhelm of “blank page paralysis” commonly inherent to creative tasks [46].

Perhaps most significantly for health care intervention research, the application of generative AI introduces a revolutionary level of transparency into the content creation process. By leveraging generative AI models as configurable tools, researchers gain access to a more standardized and reproducible approach for content design. This is primarily enabled through the adjustment of key parameters, such as the “temperature” setting, which are essential for tailoring the models’ outputs to specific needs [47,48]. A lower temperature results in more predictable and conservative outputs, while a higher temperature allows for increased variability and creativity in responses. Such configurability not only ensures reproducibility and accessibility but also allows for the establishment of standardized writing styles for health care interventions by minimizing the influence of tone, style, and other confounding variables. With a clearer understanding of the content generation process, researchers are better able to create content at scale, refine content with confidence, and make informed decisions that ultimately enhance the overall efficacy and impact of health care interventions.

We have chosen the setting of medication adherence for people with type 2 diabetes for our case study on the use of generative AI in health care interventions. Diabetes mellitus currently affects more than 415 million individuals worldwide, with an overwhelming 90% of these instances being attributed to type 2 diabetes [49,50]. Type 2 diabetes is often managed through a combination of dietary modifications, increased physical activity, and the consistent use of oral glucose-lowering medications. However, while oral antidiabetic medications are often critical to the management of type 2 diabetes, poor adherence to these medications is alarmingly common, with studies suggesting an average adherence rate of only 58% [51,52]. Recent attempts to address this issue have produced mixed results. Notably, a comprehensive review [53] of 182 randomized controlled trials focusing on interventions to improve medication adherence revealed that the evidence supporting their efficacy is largely unconvincing, despite many randomized controlled trials included in the review being extremely time- and resource-intensive. Consequently, such methods are challenging to scale and integrate into routine clinical settings. The paradoxical observation is that the increased complexity and costs of in-person, counseling-style intervention design might not directly lead to better adherence rates, resulting in a pressing need for more innovative, cost-effective, and scalable strategies.

In light of these concerns, SMS-based interventions have emerged as a promising avenue. These brief messaging interventions have previously demonstrated efficacy in promoting various health care behaviors [9,54-57]. Specifically in the domain of type 2 diabetes, interventions based exclusively on messaging [58-60] have shown encouraging results in enhancing medication adherence, though these findings are drawn from a limited number of trials and are not uniformly conclusive [61]. Furthermore, a limitation echoed in these studies is the notable absence of explicit theoretical frameworks guiding the interventions. For these SMS-based interventions to realize their full potential, it is paramount that they are founded on solid theoretical and technical bases, as adopting such an approach ensures that the behavioral mechanisms driving adherence are addressed effectively.

Described as the “active ingredients” of an intervention, behavior change techniques (BCTs) epitomize the most fundamental, replicable, and observable elements designed to modify the processes that regulate behavior [62,63]. To translate these strategies into a unified language, a taxonomy encompassing 93 BCTs organized into 16 groups was developed to guide behavior change interventions [64]. This standardization not only aids in replicating and optimizing strategies across various health behaviors but also enhances the comparability of research outcomes. By establishing which techniques are most effective under specific conditions, the taxonomy serves as a valuable resource for researchers and practitioners to select evidence-based approaches tailored to improving behavioral outcomes specific to their patient populations.

However, despite the taxonomy’s pivotal role in unifying terminology and subsequently facilitating more comprehensive correlations across behavior change interventions, the application of these BCTs in the realm of message-focused diabetes self-management research remains limited. Among the 93 BCTs outlined in the version 1 taxonomy, only a fraction has been used in published reports for this particular setting [65,66], despite evidence suggesting that interventions using more BCTs typically exert more substantial behavioral effects than those with fewer BCTs [67].

In light of this discrepancy, a comprehensive systematic review of systematic reviews was undertaken to quantitatively pinpoint various BCTs associated with medication adherence across chronic physical health conditions and qualitatively assess them in the context of type 2 diabetes [68]. Overall, the systematic review identified 46 BCTs pertinent to medication adherence in type 2 diabetes that can be used to develop direct messages for mobile devices to improve adherence among patients while simultaneously breaking down the various theoretical constructs (ie, variables from theories targeted by interventions) and mechanisms underlying specific behavioral strategies (ie, techniques not exclusively anchored to one theory but incorporated in interventions due to their predictive value in behavior). Therefore, from this systematic review, there emerges a robust theoretical foundation ripe for practical applications and explicitly suitable for crafting a bank of messages tailored for medication adherence among patients with type 2 diabetes.

This paper describes the use of generative AI to develop messages for patients with type 2 diabetes. When using generative AI for nuanced content creation tasks, understanding the context, requirements, and restrictions of the desired content becomes pivotal before initiating the development of a prompt.

To generate textual content, a pretrained LLM is needed. While a variety of options currently exist, both proprietary (eg, GPT [32] and LaMDA [76]) and open-source (eg, Orca [77] and Llama 2 [78]), we use GPT for this particular project. As one of the most advanced and widely recognized models in the field of AI-driven language generation [41,42,79,80], GPT benefits from an extensive body of research and a thriving community of developers.

In this work, we use the gpt-3.5-turbo-0301 model through OpenAI application programming interface (API) [47] calls to generate health care messages and use the chat completion API to communicate with the model. While the use of the completion functionality might appear to be more suitable for a single-prompt interaction, we observed superior results through the chat function during initial testing, and therefore continued development in a chat setting. Moreover, this choice aligns with the practical recommendations provided by OpenAI [48].

As our experiments are of an illustrative nature, we mostly use default parameter values for the API calls. In more nuanced use cases, these values could be tweaked by the prompt engineer to further tailor the output of the model to comply with the application, but adjustments were unnecessary for our use case. However, to ensure reproducibility of the presented results, we globally set temperature equal to 0, even though in practice one may obtain better outcomes by setting a positive temperature and rerunning the same query until a more satisfactory result is achieved. For instance, in ChatGPT, the value of the temperature is set to 0.7, which allows for more varied, human-like responses.

Message generation and analysis are performed in a Jupyter notebook using Python 3.8 on a consumer-grade laptop. The source code is included in Multimedia Appendix 1.

For this work, we consider single-prompt chat completion where the messages parameter contains 2 roles—“system” and “user”—and their corresponding “content.” The conversation begins with an initial system prompt, followed by a prompt from the user. The interaction concludes with a response from GPT, which provides the generated output. To enhance the performance of the model for a task with many constraints (in our case, these included length, complexity, style, and BCT incorporation), attributed prompt design [81] has been used for both the system and user roles.

The system role provides general context and behavior instructions to the assistant. It is used to explain the setting, rules, parameters, and personas of each participant in the conversation.

The content of our system prompt is given in Textbox 1 and consists of four main components:

The user role begins the conversation by providing the first interaction to which the assistant role can respond. In our setting, the user role has been defined through the system prompt as “behavioral scientist,” and reflects a templatized version of the BCT database to provide the assistant with the selected BCT and its corresponding information.

The structure of our user prompt is given in Textbox 2, where the tokenized attributes are replaced with their corresponding values from the BCT database for each query. The five attributes used in our prompts are as follows:

Results are delivered through the “assistant” role, which is the content generated by the chosen GPT model in response to each particular combination of system and user inputs. To maintain the consistency of the presented results, we postprocess the model output by stripping quotation marks and standardizing the message separation to a single line break.

This research focused on the development of content for SMS interventions using generative AI, which did not require the collection or analysis of personal data or direct engagement with individuals. According to federal guidelines from the US Department of Health and Human Services, known as the “Common Rule” (45 CFR 46) [82], ethics board approval is required only for research on human subjects that entails obtaining data through interaction with individuals or the use of identifiable private information. Since this study did not meet these criteria, an ethics review was not applicable. The generated content and associated processes were evaluated for adherence to ethical standards in the context of AI-generated material and were designed to be transparent, reproducible, and free of harm.

Using the attributed system and user prompts, 25 messages are generated for each of the 46 selected BCTs, resulting in 1150 BCT-informed messages. A total of 45,905 tokens costing US $0.07 were used for the entire data set generation, with an average of 998 tokens used and US $0.0015 expenditure per BCT. Approximately US $10 was spent for the entire prompt development process, which required over 300 iterations of the prompt. The first 5 generated messages for the first 6 BCTs are presented in Table 1, with the full data set made available in Multimedia Appendix 1.

We analyze the generated messages to evaluate their overall adherence to the required metrics. Specifically, we consider the following parameters: message length, message readability, and diversity of messages and their BCT alignments compared with an existing message data set.

Messages should be 160 characters or less. In total, 1034 of the 1150 (89.91%) total generated messages satisfy this requirement (Figure 3). The average length of all generated messages is 119 characters.

Messages should fall within or below an 8th-grade level on the Flesch-Kinkaid Grade Level Test. In total, 688 of the 1150 (59.83%) total generated messages satisfy this requirement with no alterations for our setting (Figure 3). The average grade level of all generated messages is 8.4.

While this is the initial score for all messages, it is critical to note that the nonnegotiable and unsubstitutable word “medication” is considered complex due to its character length (10 characters) and number of syllables (4 syllables). However, for our setting, it is assumed that a population prescribed diabetes-management medications will be cognizant of the word “medication,” and it will thus not pose the same complexity barrier in our context as it might in other applications. Therefore, when the word “medication” is ignored during the readability calculation, 928 out of the 1150 (80.7%) total messages satisfy the readability requirement, with an average grade of 6.5 (Figure 3)—a closer, more accurate metric of complexity for our particular use case.

In this paper, we propose a novel approach to creating behaviorally informed content for brief message interventions. Using the setting of medication adherence for people with type 2 diabetes, we use a pretrained LLM to develop a bank of text messages based on BCTs curated in a recent systematic review [68]. This work is intended to act as a blueprint for future research to create a more transparent, replicable, and scientifically rigorous look into the content creation process for brief message interventions and serve as a starting point for subsequent studies to analyze the safety, efficacy, and viability of AI-generated messages.

In this paper, we show the potential of generative AI as a tool for transparent and replicable content creation in brief message interventions. Inspired by a list of 46 BCTs and their corresponding theoretical constructs and behavioral strategies, we engineer attributed system and user prompts for GPT to generate 25 messages for each of the 46 BCTs, for a total of 1150 messages in the specific setting of medication adherence for type 2 diabetes. Our findings reveal that a significant majority of the generated messages were compliant with both message length and complexity considerations (1034/1150, 89.91% and 928/1150, 80.7%, respectively), making them well-suited for SMS-style interventions.

The diversity of generated messages is analyzed through the distributions of their embedding vectors with 2 popular pretrained natural language processing systems: BERT and ADA. The generated messages showcase a diversity in message content that is comparable with an existing publicly available data set of brief messages and reflects similar distributions among BCTs from the comparison data set while also maintaining variability between messages of the same BCT, thereby demonstrating the capability for generative AI to craft a plethora of unique and contextually relevant communications with only a very standardized change in input between BCTs. The data set and source code for message generation and analysis are available in Multimedia Appendix 1.

As generative AI is an extremely new field growing at a rapid rate, new algorithms and “versions” of LLMs are being released regularly. These updates can often fundamentally change the assistant output generated by the same system and user inputs, leading to a lack of consistency in experiments conducted with the same prompt over a long period of time. Especially for users of chatbot-style LLMs like ChatGPT, these updates can come suddenly and without permission, making research on such platforms difficult. While this issue is somewhat mitigated using an API (which generally does not force immediate adoption of the newest models), many current LLMs will eventually be depreciated, albeit at a more gradual rate. The only true mitigation of this limitation is the use of open-source models (such as, for example, Orca [77] and Llama 2 [78]) that can be fully downloaded and deployed on the client side; however, this comes at the cost of the technical proficiency required to set up such a system. Therefore, while prompts like ours serve as great examples of attributes to consider and the language one might use when constructing a prompt for a particular setting, the definitive construction of a singular, unchanging prompt to support intervention research is generally unfeasible in practice.

Another potential limitation of generative AI for very large-scale message generation is the finite context window size of the given LLM. In our results, we generate 25 messages per BCT, equating to an average of 998 tokens for each interaction, which keeps us well within the current 4000-token context limit bounds of gpt-3.5-turbo-0301. If a larger bank of messages is required, according to our current standardized single-prompt structure, one could feasibly increase the number of generated messages to 100 or more before nearing the limit. However, for extremely large-scale data generation, even a single-prompt interaction will likely be insufficient, and because LLM context is cumulative, this restriction will likely be most prominent for multistep interactions.

It is important to note that our use of a singular templatized prompt across multiple BCTs can sometimes fail to capture the nuanced essence of each distinct BCT. This “one-size-fits-all” approach may yield inconsistencies in the quality and accuracy of the generated content for certain BCTs as compared with others. However, prior studies in behavioral science have also suggested that some BCTs may not be capable of being delivered effectively in an SMS format regardless of the method of creation [90], perhaps indicating that difficulties in accurately representing some BCTs may have less to do with the limitations of generative AI and more to do with the inherent complexities of some BCTs in the given setting of brief message interventions.

Generative AI, while versatile, is heavily reliant on the specific language of the prompts provided; thus, when prompts rely heavily on external data fitting into a templatized structure, small, seemingly insignificant differences in the style or authorship of that data can potentially affect results [91,92]. For those aiming for AI-generated content that would be used in practice without human review, the design and fine-tuning of dedicated prompts for each unique contextual change (in our case, BCTs and their corresponding information) are likely a necessity. Therefore, while our approach serves as a proof of concept for efficient single-prompt content creation and demonstrates the vast potential of pretrained LLMs in this domain, achieving universally accurate results for all BCTs would necessitate a more granular, tailored approach to prompt design, attending to the individual nuances and requirements of each BCT within its specific context.

As with any state-of-the-art technology, ethical considerations for the implementation of generative AI are paramount, guided by the core principles of transparency, privacy, accountability, and fairness [93,94]. However, the inherent unpredictability of LLMs becomes acutely significant in health care contexts, where the consequences of misinformation or inappropriate AI-generated suggestions can be dire for patients. Given that ensuring fully reliable, safe, and accurate information generation by LLMs is deemed “fundamentally impossible” [95], human checkpoints become indispensable before, during, and after AI employment. Researchers and domain experts must meticulously review AI-generated content for accuracy, safety, and equity, especially when evaluating its usability for complex patient-facing health care interventions.

It has been demonstrated that diversely attributed, complex prompts can reduce biases [81], but it is imperative to strike a balance, as excessively long prompts may increase the likelihood of undesired model behaviors [96]. In our approach, we acknowledge the crucial role of domain experts in both prompt design and content review, effectively mitigating the risk of malicious actor involvement and enabling the use of longer, more attributed prompts. Generative AI should not be seen as a standalone solution but as a tool that augments and accelerates the work of researchers. While it dramatically enhances efficiency in content creation, the responsibility for upholding rigorous standards of accuracy, safety, and fairness in health care interventions remains firmly with human experts, and it is the symbiotic collaboration between this tool and the human research team that ensures the delivery of ethically sound and clinically effective interventions.

To the best of our knowledge, this is the first work to propose the use of generative AI as a tool for content creation in SMS health interventions. However, a similar study [28] detailing a traditional content creation process was undertaken using the same systematic review [68] as a theoretical framework. A workshop was held for content creation and subsequent focus groups and surveys were used for review, resulting in the production of 371 messages informed by the selected BCTs in the context of medication adherence for type 2 diabetes. However, despite efforts toward transparency, this work does not reveal a detailed account of the actual content creation process, and the data set of generated messages has not been made publicly available for review or comparison.

Previous studies have looked at traditional content creation for brief message interventions, with a specific focus on the selection and review of BCTs and their corresponding messages [27,97,98]. More broadly, investigations into mHealth interventions have been a hot topic in health care research for years [11,14,15,26], and several works investigating the specific incorporation of behavioral science into brief message interventions have been previously undertaken with positive results [99-101].

In addition, one-shot, zero-shot, and few-shot approaches to prompt engineering have seen an explosion of interest following the expansion of LLMs within the public and academic mindset, leading to a large body of research on the methods and frameworks of prompt design for a variety of contexts and use cases [45,102,103].

This is the first in a series of works detailing the process for responsible and efficient use of generative AI in the development of brief message health care interventions. The next step involves assembling a team of qualified behavioral scientists and other domain experts to conduct in-depth analyses of the generated messages, focusing on their adherence to safety standards, adjustments to meet the technical requirements of an SMS delivery system, a formal review of BCT coding for each message, and general checks that the generated messages meet best practice standards for content design.

Moreover, the development of a subsequent interaction with the model could be used to self-adjust the generated results based on designer feedback. Using a multistep prompting method, the model could, for example, be directed to self-assess for safety and equity considerations, as well as edit more individually for the given use case based on specific critiques provided by the research team. Such iterative developments of the model should necessarily involve rigorous patient testing and feedback—a crucial step in ensuring that the AI-generated content resonates with patients’ experiences and needs to further personalize and refine the developed health care interventions.

Standardizing the realization of individual BCTs within brief message content represents another critical research direction. Many interventions currently withhold both their messages and their content creation processes, potentially introducing unintentional biases and skewed outcomes due to inherent differences in writing styles and other design-related confounding variables. By advocating for the transparency and standardization of content design, we can enhance the research efficacy of interventions by reducing such confounders and further ensuring that the results of such interventions are truly tied to the theoretical frameworks and behaviors being tested.

Finally, the ambitious goal of generating hyperpersonalized health care communications tailored to individual patients becomes a promising possibility with LLMs. Future iterations of content could be further personalized by implementing features like translating content for different languages and localizations; tailoring content for cultural relevance and sensitivity in the use of examples, metaphors, and references; adjusting the complexity levels of the text to cater to different educational backgrounds or cognitive abilities; and providing accessible formats adjusted for individuals with disabilities. While conventional content creation methods struggle with the impossible number of potential messages required for personalized content for every recipient, attributed prompting with generative AI offers the potential to create individualized messages that can significantly enhance patient engagement and outcomes and change the way health care interventions are experienced.

In this study, we explore the practical application of generative AI for content creation in the development of brief message health care interventions. We illustrate the potential of using pretrained LLMs as a tool to aid researchers in the resource-intensive process of content creation by generating a data set of 1150 messages inspired by 46 BCTs selected for the setting of medication adherence for type 2 diabetes. Building on the foundations laid by former health care intervention development studies, this paper differentiates itself by the following:

Ultimately, the value proposition of using generative AI in this domain lies not in the perfection of the initial generated content but in its adaptability and capacity to rapidly produce a multitude of messages that can subsequently be refined and curated by human experts. This combination of AI-driven speed and human-driven supervision presents an efficient, transparent, and scalable method for developing effective and replicable brief message interventions. While follow-up studies are needed to ensure the safety and usability of the generated messages and provide potential refinements to the proposed prompts for individual settings, the use of generative AI in health care intervention development opens new doors for the scalability and potential standardization of content creation within health care intervention design and research. Given the time- and cost-intensive nature of crafting interventions traditionally and the current opacity of the content design process, our study underscores the potential of generative AI as a significant efficiency tool poised to revolutionize the creation of behavior change interventions for medication adherence and beyond.