The intricate relationship between health and the environment is well established, particularly for global youth. Adolescents experiencing environmental challenges in their communities, such as poor waste management or stagnant water in the streets, are more likely to develop diseases such as dengue fever, Zika, parasitic diarrhea, and acute dermatitis, as well as mental health issues such as anxiety and depression [1,2]. For example, research in Indonesia showed that children living in households where trash was not routinely collected and removed were significantly more likely to be dengue seropositive [2]. These hazards often reflect persistent infrastructure deficits such as poor drainage, insufficient waste services, and limited maintenance of public spaces, which can accumulate over time and reinforce long-term health inequities. Despite widespread recognition of these risks, many communities lack the resources, authority, or partnerships needed to translate knowledge into sustained environmental improvements [3,4]. This gap is especially pronounced in low- and middle-income settings, where government capacity to implement meaningful change is frequently constrained [5].

Youth-led citizen science strategies aimed at enhancing local environments have demonstrated significant benefits, including improved health outcomes and increased feelings of empowerment among adolescents [6,7]. These strategies enable adolescents to become investigators of their environment and promoters of positive change. However, successful implementation of these interventions is challenging, as it requires meaningful engagement from adolescents, strong stakeholder commitment, and well-planned, feasible, and efficient solutions [2,7].

To increase appeal and impact, efforts to engage youth in community solution-building, using approaches such as the Our Voice method [8], can leverage the digital technologies that young people around the world engage with fluently on a daily basis [9]. Such technologies are evolving rapidly, characterized by the widespread adoption of artificial intelligence (AI), which has demonstrated an increasing capacity to process complex data, transform images, and generate insights to solve problems. AI can model environmental changes with a level of detail and accuracy previously unattainable, enabling more informed decision-making that could lead to more effective interventions, such as optimizing renewable energy resources, improving waste management, and predicting natural disasters [10-12]. Additionally, AI has the potential to provide tools that enhance action-oriented citizen science and transform adolescents’ environments.

This AI application is ready for broad exploration, and several tools offer opportunities for initial testing of potential impacts on youth engagement in community health improvement. For example, GPT language models can provide constructive feedback related to solution-building. Furthermore, photo transformation tools can improve community members’ ability to visualize potential solutions and enrich their subsequent engagement with stakeholders. However, these capabilities come with considerable risks, particularly the potential to amplify existing inequalities if AI research and development are concentrated within already privileged populations or contexts [13]. Recent scholarship further cautions that generative AI systems may produce plausible but inaccurate outputs, reflect embedded biases in training data, and require structured human oversight to ensure ethical and contextually appropriate deployment in mental health–related settings [14]. If not applied judiciously, AI use might also risk replacing adolescents’ voices rather than amplifying them [15].

In this study, we adapted and evaluated the use of 2 specific AI tools to support community-driven action research (1) providing feedback through GPT, and (2) image transformation. We then applied these tools to a citizen science project that engaged adolescent residents of the vulnerable neighborhood of El Pozón in Cartagena, Colombia.

OpenAI ChatGPT and similar AI systems can offer critical feedback on problem-solving strategies, especially in community-based initiatives. For instance, adolescents proposing the draining of stagnant water to improve public health in their communities can amplify their impact by leveraging AI to refine their ideas and provide constructive feedback. ChatGPT can assist in identifying potential challenges to proposed strategies, suggesting alternative approaches, and offering evidence-based recommendations to enhance the effectiveness of their solutions. As noted earlier, by engaging with AI, young people can potentially deepen their understanding of public health issues, gain insights into the environmental and social implications of their actions, and develop more comprehensive and sustainable interventions. This collaborative process between AI and human creativity may support communities in addressing local problems more effectively and innovatively [15,16].

AI technologies like Adobe Photoshop AI can reshape our perceptions of the future by transforming visual representations of the world. This technology empowers designers, architects, and urban planners to create compelling visions of sustainable cities, clean energy landscapes, and resilient communities that inspire action and drive societal change. In the context of community-driven change, these tools enable us to reimagine the “undesirable present” and project a “desirable future” with clarity. Such AI-generated imagery may arguably galvanize public support for environmental and social initiatives, making the abstract tangible and the future visible [17].

El Pozón, in Cartagena, Colombia, has witnessed rapid urbanization and glaring social disparities over the last 2 decades. Community members in this area, established as an informal settlement in 1969, still grapple with a lack of essential services, multidimensional poverty, and unfavorable living conditions [18]. In this intricate and challenging context, Bradbury and colleagues [19] made a compelling case for the need to incorporate diverse sources of knowledge to advance positive community-level change. They emphasized that involving individuals intimately connected to the issues is imperative to mitigate the pervasive consequences of poverty and unhealthy environments. In this regard, the constructive collaboration between action research and citizen science emerges as a potent catalyst, harnessing the abundance of local human knowledge and understanding to effectively and sustainably address pressing community problems.

The Stanford University Our Voice citizen science method is a form of participatory action research that engages community residents as scientists through a multidimensional mixed methods approach [7]. In a 4-step process, participants become citizen scientists through a user-friendly, multilingual mobile app called the Discovery Tool.

Using the Discovery Tool app, citizen scientists take georeferenced photographs and record narratives, either through audio recordings or text notes, describing the elements in their environment that either diminish or enhance their health and well-being. To conclude this Discover step, the citizen scientists upload their anonymized Discovery Tool data to a secure server. Then, during the Discuss step, the citizen scientists participate in facilitated meetings to discuss, analyze, and prioritize their group’s collective findings and then brainstorm potential changes and action steps.

The next step involves 1 or more facilitated meetings in which stakeholders are invited to the citizen scientists’ presentation of their findings and recommendations (advocacy step). Ideally, the citizen scientists then work collaboratively with other community stakeholders to co-organize and implement realistic next steps to address the pressing problems they identified during the Our Voice process (change step). While the Our Voice platform has been shown to be accessible and effective, to date, the integration of AI to enhance the community-driven change process has not been fully tested [15,20].

This was an action research study aimed at investigating AI’s capacities to further enrich the Our Voice process by amplifying and activating the collective voices of vulnerable adolescents (also known as citizen scientists) as they endeavored to improve their community in El Pozón, Colombia.

The study consisted of several phases. First, during a preparation phase, we (the researchers) evaluated the initial applicability of 2 AI tools: image transformation and feedback generation using SecureGPT (a privacy-enhanced version of ChatGPT 4.0; OpenAI). During this phase, we conducted several small experiments, including transforming images with various prompts and refining solutions with SecureGPT. The results of these experiments allowed us to reflect on the effectiveness of the AI tools and determine the best ways to adapt them for real-world applications.

Second, we applied insights from the preparation phase to tailor AI tools for the Our Voice project in El Pozón. We used these adapted tools during the Our Voice discuss and advocacy steps of the El Pozón project and evaluated their effectiveness in enhancing the Our Voice process. Figure 1 provides a schematic overview of the full study workflow.

This study consisted of 2 phases. In the first phase, the researchers tested different AI tools to prepare for their use in the second phase. The second phase involved implementing the AI tools in a citizen science project conducted on the Caribbean coast of Colombia.

During the study preparation phase, the researchers tested the integration of AI-supported interactive conversations using Stanford Medicine SecureGPT (beta), a privacy-enhanced version of ChatGPT 4.0. During this phase, which did not involve external participants, the research team tested SecureGPT by introducing a realistic community problem and corresponding proposed solutions derived from a prior Colombian Our Voice project conducted with youth citizen scientists [7]. We then tested 3 different prompts designed to elicit feedback on the proposed solutions. Each prompt was tailored to simulate a particular stakeholder’s perspective. For example, we asked SecureGPT to assume the roles of a community leader, a civil engineer, and a government official reviewing the proposed solutions. We then asked SecureGPT for feedback based on these assigned roles.

This study was reviewed and approved by the Universidad de los Andes IRB (IRB approval number 2024300727) and conducted under Stanford University’s general Our Voice IRB (IRB approval number 40379). Two parents or caregivers provided written informed consent for each participating adolescent, and adolescents provided assent. Participant privacy and confidentiality were protected through the deidentification of all collected data. Photographs were limited to public spaces and were reviewed before analysis and dissemination to ensure they contained no identifiable individuals or other personally identifying information. Participants were compensated with school materials (eg, water bottles, notebooks) for their time at each project stage (discovery, discussion, and advocacy), receiving approximately US $15-$20 per stage.

The results obtained during this study phase were critical in preparing for the real-world implementation of AI during the Our Voice project in El Pozón.

We initially tested DALL-E’s capacity to transform pictures to reflect a desired realistic future. As the seed (original) picture, we used a photo taken by an adolescent citizen scientist from a previous Colombian Our Voice project.

With DALL-E, we did not find the capability to produce realistic images. Interestingly, DALL-E even suggested using an editing tool such as Adobe Photoshop. Thus, after this first iteration, we tested Adobe Photoshop AI to assess its capacity to transform photos (Figure 2).

The Adobe Photoshop AI–generated images demonstrated that picture transformations with this program were very realistic, and the process was most effective when prompts were not only specific, such as “remove water” or “dry the street,” but also more conceptual, such as “make this street healthier for adolescents to walk.” Additionally, it is important to note that a well-transformed image was often the result of an iterative process. Prompting Adobe Photoshop AI with specific requests using the lasso tool, an area-selection feature, enabled us to transform certain areas of the picture into a realistic image of the desired future. This visual representation of potential solutions can help enrich and reinforce citizen scientists’ voices, which in turn may accelerate the research-to-action process.

These findings suggest that the potential to transform pictures could significantly enhance citizen scientists’ options for discussion. We therefore determined that during the implementation phase of this project, we would use Adobe Photoshop AI and the lasso tool to select and modify specific areas of community photos collected through the Discovery Tool app.

Based on insights gathered from the previous Our Voice steps, which identified waste management and stagnant water as the most critical issues, the research team, together with the citizen scientists, invited different stakeholders and decision makers to participate in the advocacy step of the Our Voice method. This meeting was attended by 38 decision makers representing PaCaribe (waste management), the regional Environmental Protection Agency, the District Administrative Department of Health, Foundation Grupo Social, the Public Space Office, the Community Participation Office of the State of Bolivar (Cartagena is the capital of Bolivar), and the Nuestro Esfuerzo (Our Effort) school leadership.

During the meeting, the youth citizen scientists presented their AI-refined solutions along with AI-enhanced images illustrating them. This provided stakeholders with clear insights into the citizen scientists’ vision and recommendations. The stakeholders then organized into committees to address the pressing issues identified: waste management and stagnant water. By the end of the meeting, the stakeholders, working with the citizen scientists and project facilitators, were able to define and commit to actions for 1 and 3 months.

Among the most relevant commitments that emerged in response to the AI-enhanced proposals were (1) planning and conducting a collective walk to identify critical points for waste disposal that would inform where containers are needed; (2) placing waste containers inside the school (PaCaribe waste management group in charge) and managing them adequately (school in charge); and (3) using geo-localized coordinates of planter holes to install the lids conceptualized by the students using Adobe Photoshop AI (research team, District Administrative Department of Health, and the Police Inspector in charge).

Over the 2-month period following the advocacy step, the research team observed that the Adobe Photoshop AI–enhanced lids and the containers conceptualized by the students became the stakeholders’ objectives. As the Police Inspector noted, “We need to place those lids, those that were shown in the picture, before the rainy season starts.” In addition, PaCaribe agreed to install 2 smaller waste containers inside the school to prevent vandalism and theft. Currently, researchers, citizen scientists, and PaCaribe are in discussions to explore options for securely placing a smaller container outside the school, as it was determined that the Adobe Photoshop–created container was too large and heavy for collection and transportation. Notably, the AI-generated images continue to be referenced as aspirational goals, even if they may not represent the most feasible solutions at present. This highlights the power of visual representation to inspire stakeholders and guide discussions, as these images can serve as a motivating vision of future possibilities, even amid current practical constraints.

Across iterations, AI outputs varied in usefulness. Some SecureGPT responses were repetitive across stakeholder role prompts and required facilitator guidance to focus the discussion on locally relevant considerations. At times, outputs were overly long or generic, prompting refinement of the 3-3-5 prompting structure to improve specificity. Early experimentation with DALL-E for image transformation did not yield sufficiently realistic place-based modifications, leading the team to rely on Adobe Photoshop AI within a researcher-mediated workflow. These observations informed adjustments to prompting strategies and facilitation procedures throughout the implementation phase.

The integration of human-facilitated AI tools into an Our Voice Citizen Science project showed potential for supporting youth-led community engagement and advocacy efforts aimed at addressing critical environmental health issues. Recent randomized controlled trials have evaluated AI-powered chatbots as direct therapeutic agents for reducing symptoms of depression and anxiety, demonstrating short-term clinical improvements in structured individual treatment contexts. By contrast, our study examines AI not as a replacement or substitute for human therapy, but as a structured, facilitative scaffold embedded within a collective, youth-led participatory action framework [21]. This shift from AI as therapist to AI as deliberative support represents a distinct application that extends beyond symptom management toward environmental health support and civic participation. Specifically, the use of SecureGPT and Adobe Photoshop AI during the discussion and advocacy steps of the Our Voice method provided valuable support for refining and visually representing solutions proposed by citizen scientists, thereby facilitating clearer communication and more effective stakeholder engagement compared with prior projects in this region [6].

Notably, the images generated by Adobe Photoshop AI (version 25.11) left a lasting impression on the stakeholders, who continued to refer to the lids and containers, even though they were only visual concepts. Initial testing of these AI tools during the study’s preparation phase was essential for ensuring their agile and practical use during subsequent interactions with the citizen scientists. For instance, based on insights gained from the study preparation phase, we employed a structured prompt asking SecureGPT to identify 3 strengths, 3 weaknesses, and 5 reflective questions. This structured feedback subsequently enabled the citizen scientists to constructively refine their solutions, supporting a qualitatively expressed increase in the perceived feasibility of their proposals. Additionally, AI tools appeared to support citizen scientists in integrating information and perspectives that were typically beyond their immediate knowledge. Finally, facilitating consensus among citizen scientists regarding the prompts provided to Adobe Photoshop AI and SecureGPT was critical for improving teamwork and ensuring effective group decision-making [22-24]. Prior scholarship examining the intersection of machine learning and citizen science has primarily focused on AI as a tool for large-scale data classification and environmental monitoring. By contrast, this study explores AI as a structured cognitive and visual support tool within the deliberative and advocacy phases of participatory research. Rather than replacing human analysis, AI functioned as a structured support embedded within facilitated group processes [20,21].

In addressing both waste management and stagnant water issues, SecureGPT provided valuable feedback that aided in refining and strengthening the voices of the adolescents. For waste management, SecureGPT highlighted strengths in the adolescents’ proposed strategies, such as community awareness when accompanied by infrastructural change. It also identified important weaknesses, including over-reliance on influential adult actors, potential sustainability challenges, and the need for clearer implementation strategies. These notes prompted deeper reflection among the citizen scientists on practical considerations such as long-term sustainability and resource allocation. Similarly, for the stagnant water issue, SecureGPT’s structured feedback helped the citizen scientists critically assess their proposed interventions. For example, they initially suggested filling unfinished cement planters, but later refined the approach to installing protective lids to prevent water accumulation without spoiling public space elements. The AI-generated insights emphasized the importance of technical feasibility, long-term maintenance, and awareness of potential unintended consequences, thereby facilitating a more comprehensive understanding of the complexities involved and underscoring the health risks associated with dengue fever and other infectious diseases.

Adobe Photoshop AI was used to transform participant-generated photographs into visual representations of potential environmental changes based on the citizen scientists’ observations. These images were used during the Our Voice advocacy step, when the citizen scientists presented their priorities and proposed actions to stakeholders, including local authorities, health officials, and community members. In this context, the images provided a concrete reference point that stakeholders could discuss alongside the youth’s recommendations. At the same time, highly realistic AI visualizations can introduce a feasibility gap when images imply infrastructure options that are not operationally workable. If not made explicit, this gap may inflate expectations and reduce stakeholder confidence. In this study, facilitated discussion and stakeholder review were used to frame AI images as aspirational prototypes rather than technical designs and to support iterative refinement toward feasible options.

The advocacy step itself demonstrated the practical impacts of integrating AI tools into citizen science methodologies. Stakeholders, including representatives from the District Administrative Department of Health, the local police inspector, nongovernmental organizations, and waste management companies, actively engaged with the AI-refined solutions and committed to concrete short- and medium-term actions. For example, the placement of waste containers by PaCaribe and the educational workshops by the District Administrative Department of Health on dengue prevention reflect initial actions addressing the identified environmental health challenges.

Despite these successes, several limitations should be acknowledged. First, although Adobe Photoshop’s AI tools generated realistic visualizations, producing the final images required substantial human oversight and multiple rounds of iterative refinement. This oversight requirement aligns with broader concerns that large language models and generative systems may produce plausible yet contextually inaccurate outputs without domain-specific supervision. Ensuring that AI-generated suggestions were reviewed and contextualized by facilitators was therefore key to maintaining technical feasibility and preventing misinformation [25]. Second, Adobe Photoshop is proprietary, paid software; reliance on such platforms may limit accessibility and could exacerbate existing disparities in technological resources across users or institutions. Sustainability beyond the funded research period remains uncertain if access to proprietary AI tools is withdrawn, underscoring the need to test lower-cost or open-source alternatives. Third, the intervention was implemented within The Ecology Group, an already well-established and highly functional organization. As a result, the observed positive outcomes may not generalize to newly formed groups or teams with less cohesion, fewer resources, or less organizational structure. Fourth, documentation relied on structured field notes and analytic memos rather than verbatim transcripts, as facilitating group discussions, operating AI tools, and recording observations simultaneously create cognitive demands susceptible to recall bias. To mitigate this, field notes were expanded into analytic memos immediately after each session and reviewed in research subgroup debriefs, with audio recordings used to support documentation accuracy.

Fifth, although the first author served as group facilitator and domain expert responsible for identifying inaccuracies in AI-generated infrastructure recommendations, prompt entry and note-taking were carried out by dedicated research team members. Nevertheless, the concentration of facilitation and technical oversight within the research team introduces researcher mediation bias that may have shaped which outputs were flagged and how facilitation proceeded. Future implementations should test whether community facilitators without specialized technical training can effectively oversee AI outputs to assess replicability beyond researcher-led contexts. Sixth, the 2-month observation period following the advocacy step was initially insufficient to verify whether stakeholder commitments translated into durable environmental change. At the time of the original data collection, reported outcomes reflected documented engagement and stated intentions rather than verified infrastructural transformation. However, at the time of this resubmission, follow-up communication with school staff indicates that sustainable changes have been implemented and that students have continued to use AI tools independently to advocate for further community improvements. While encouraging, these updates are informal and were not captured through systematic follow-up, underscoring the need for future studies to incorporate structured, extended follow-up periods to rigorously assess advocacy impact. Finally, while the intervention was conducted in Spanish, AI tools can vary in accuracy and linguistic nuance across languages, which may have shaped how prompts were interpreted and outputs were generated, highlighting the need for future work to examine language-specific performance and its implications for equity.

Future work should evaluate alternative tools with more intuitive interfaces, lower barriers to access, or both, and assess performance across a broader range of organizational contexts (including newly formed or less established groups) to better understand generalizability and equity considerations. In addition, future research should establish clear guidelines and best practices for effective use across diverse populations and settings, identify strategies to expand equitable access to these technologies, and develop structured facilitation processes to optimize implementation and impact. More rigorous study designs could also strengthen the evaluation of AI integration relative to the Our Voice approach alone. For example, matched schools or communities within the same locale could be randomly assigned to an AI-supported Our Voice condition, while control groups follow the traditional Our Voice method without AI [14].

This exploratory action research study examined the integration of 2 generative AI tools within a youth-led Our Voice citizen science project addressing environmental health concerns in El Pozón, Cartagena, Colombia. Structured SecureGPT prompting, using a fixed strengths, weaknesses, and reflective questions (3-3-5) format, supported more systematic deliberation and feasibility-oriented refinement of adolescent-proposed solutions. Adobe Photoshop AI enabled rapid generation of place-based visual prototypes that strengthened shared understanding and provided concrete reference points during stakeholder advocacy. Across both tools, effective use depended on facilitator-mediated workflows, real-time group review, and domain-informed oversight, highlighting that these systems functioned best as structured supports within collective decision-making rather than as autonomous authorities. For participatory action and citizen science researchers, these findings suggest a practical approach for using generative AI to structure reflection, prototype solutions, and support stakeholder alignment while preserving youth agency and attending to feasibility and equity constraints.