This is the first in a three-part series sharing findings from our AI for Climate Action project, which is mapping community-centered AI tools for climate action across Latin America, Africa, and Asia. This series explores three themes emerging from our research. In this first post, we look at how organizations make decisions when considering AI for climate action. The next posts will explore data practices and the role of AI in supporting social movements.
When people talk about AI and climate change, the conversation tends to split into two narratives: AI’s environmental cost and its potential to help address the climate crisis. This framing leaves little room for smaller, locally grounded AI initiatives that can help us imagine the role AI can play in climate action beyond this binary, and what it can do for communities in the frontlines of environmental defense. Building on the tradition of participatory AI, which seeks to redistribute power in how these systems are designed and governed, our AI for Climate Action project set out to explore what a community-centered approach to AI for climate action might look like.
Over the past three months, The Engine Room mapped more than fifty AI initiatives across Latin America, Africa, and Asia and interviewed twelve organizations exploring how locally grounded AI was used for climate action throughout their lifecycle – from design and model training to deployment and eventual decommissioning. Rather than neat technological lifecycles, we came across non-linear journeys in which decisions to use AI (or not) were shaped by the possibilities brought about by the technology, as much as by community priorities, ongoing social and environmental struggles, and a spirit of creativity, resourcefulness, and experimentation.
This post focuses on the earliest part of that journey and the questions organizations ask when considering whether to build an AI tool. At the end of the post, we also share more details about our upcoming cohort learning program for organizations thinking of integrating AI into their climate action work.
Deep and active listening as the starting point, always
A recurring theme in our interviews was that AI was rarely the starting point. Organizations described lengthy periods of listening, testing, and reflecting with their communities before deciding whether AI had a role to play in addressing the climate problem in front of them, particularly in light of ongoing struggles for climate justice.
This was explicit in the collaboration between Diversa, a Quito-based AI and data organization, the Yaqui community of Vicam, Sonora, in Mexico, and Técnicas Rudas, a Mexico-based organization who partnered to build a data dashboard supporting the Yaqui community’s defense of their water rights. Diversa’s team was clear that the project did not begin with a proposed algorithm or model:
“The first significant moment was when they asked us: we want to understand access to water, what’s happening with the Río Yaqui, we wanted to understand water governance; because these were the community’s priority needs. [The project] was born out of a territorial necessity, not from a pre-existing technological solution.”
Before any technical development began, Diversa and Técnicas Rudas held workshops, surveys, and participatory activities and “a lot of active listening” with Yaqui community members, including children, women, and community leaders to identify the community’s priorities and the questions a tool would need to answer. Those conversations ultimately shaped what the dashboard measured and how it supported the advocacy efforts of the Yaqui community. For example, in addition to including environmental data, it also included an archive of Mexico’s water-related legislation for community use in assemblies and legal proceedings.
A similar logic shaped the Voice of the Ogiek Messenger project in western Kenya, an AI-powered messaging system built with the Ogiek-recognized custodians of the Mau Forest who have faced repeated evictions carried out in the name of conservation. The team situated these evictions within a longer history of conservation efforts that displace the communities protecting the land. As one researcher explained, the project began from that tension:
“There’s a long history about how conservation has harmed Indigenous communities who are actually the conservators of those spaces. In looking at the Ogiek case that had been going through the African human rights courts and the whole conversations of the government pushing them out, we went back and began the conversation. How about we found a messaging system that the community can use to share their concerns about climate justice, but do it in a way that is very sustainable.”
Answering that question meant preserving the possibility of refusal from the start. The team described this explicitly as “recognizing refusal as a methodology, that the community could refuse, and we needed to respect that. They could redirect us. We needed to respect that.” The team brought the idea to the community’s Council of Elders. During the meeting, the Elders discussed the risks involved, including concerns about jeopardizing ongoing court cases, as well as the team’s commitment to developing the project in a non-extractive way. Work began only after these discussions and the Elders granted their approval.
For organizations considering AI in their own work, these experiences suggest the value of engaging in meaningful reflection and conversations with the people at the frontlines of climate action or most affected by climate change before engaging in the design or adoption of a specific tool.
A frugal approach can open up more imaginative and effective ways to use AI
While many of the organizations we interviewed faced constraints in their contexts, they found that designing for unreliable electricity, limited connectivity or access to hardware became an opportunity to build more context-appropriate tools. In fact, several projects argued that smaller, task-specific AI systems can perform better than bigger models, while requiring fewer financial, computational and environmental resources. In the words of the Diana from the Diversa team:
“You don’t need a rocket to move from your house to the store. Maybe you need to walk, maybe you need a bicycle but you don’t need something that big. There are other AIs, other algorithms, smaller models, that don’t consume so many resources, that don’t need so much data. Those conversations are being forgotten in the rush toward generative AI.”
Through our desk research, we found a series of tools designed to be technically responsive to low resource contexts and low-bandwidth environments, such as AI and SMS integration for non-smartphone users, initiatives like Nsukka Yellow Pepper Project in Nigeria using solar powered tools and local hardware based on Raspberry Pis to bypass cloud storage especially in low broadband environments, or Skiliket in Mexico that used low-cost microcontrollers to build sensors that collect environmental data without relying on proprietary infrastructure.
We also spoke with the team behind Situational Intelligence Open Source Software (Siti OSS), which powers PetaBencana.id in Indonesia and MapaKalamidad.ph in the Philippines, who have deliberately relied on smaller, targeted systems paired with human verification, rather than pursuing increasingly large predictive models, for disaster forecasting. As the founder explained:
“In the global AI conversation, what’s dominant right now is that we are being bombarded with this idea that there’s only one way to build AI. But actually there’s not. There’s actually a lot of different ways to build AI. And starting with massive datasets is not the only way. In fact, it’s a very inaccurate way. Besides being far more energy efficient and better for the environment, smaller better-curated datasets are also much more accurate than large models […] If you go with that model, which is right now dominating the public imagination, it means that only tech companies with the largest resources, that are already the monopolies, only they can intervene in this space. And it’s a useful narrative for those folks, but it’s really not an empowering narrative and it’s not true at all.”
Based on the experience of these organizations, frugal design, in addition to responding to material constraints, can open up more imaginative, meaningful, and less harmful ways for communities to engage with AI technologies. For organizations introducing AI into their work, what possibilities become visible when energy use, connectivity, hardware access, and ways of transmitting knowledge are treated as starting points for design?; and how can these conditions lead to tools that communities can shape and sustain on their own terms?
Designing AI systems communities can own
Finally, it is also worth reflecting on how technical architectures can make community ownership structurally possible and seed new possibilities for social movements to get involved in AI technical design. Questioning who owns and controls the architecture offers a productive space to consider how communities can use AI with greater agency and autonomy and in a way that is aligned with the protection of land and intergenerational environmental knowledge.
For example, IARAA, a chatbot developed by Brazil’s Landless Workers’ Movement, the World March of Women and Baobab (International Association for Popular Cooperation) to support agroecology and land reform, was built on open-source large language models, such as DeepSeek and GLM, paired with custom retrieval-augmented generation architecture. The chatbot draws exclusively from a curated library assembled by the movements themselves, rather than from the open internet or the training data of a commercial model. As one member of the development team explained, engaging with AI as developers, rather than solely as end users, opened the possibility of shaping the tool on their own terms:
“I don’t think it’s possible to have that perspective based solely on a user’s viewpoint. But when you position yourself as a key player in the tool’s development and begin to understand the possibilities for adapting it through changes in its architecture, those possibilities start to emerge.”
Similarly, an AI of Our Own, an initiative supporting Indigenous language and cultural preservation across Cambodia, Ghana, and South Africa, arrived at a similar architectural choice from a different direction. During early conversations, community partners raised pointed concerns about what it would mean to place oral traditions and cultural knowledge within an AI system. As one Cambodian weaver and Smot chanter put it:
“If you can build an AI that can help us to teach and to preserve this tradition, then it’s good. So if I put this thing into an AI, who is going to be responsible for my data, who is going to be storing my data and where is it going to go? If I put this data into an AI system, what if it gets distorted? What if AI misinterprets my melodies? What if it sings with the wrong voice? Who is going to look after this knowledge if I put it into that system? After I’m gone?”
Like IARAA, an AI of Our Own responded also using a specific architecture (i.e. retrieval-augmented generation) that gave knowledge stewards control over the data used to inform the tool, coupled with local data centers designed to keep knowledge in community hands.
For organizations exploring AI, these examples raise questions about how technical architectures can embed meaningful political and social commitments in the design itself, and how stewards of land, environmental knowledge, and cultural memory can meaningfully shape its design.
What this means for CSOs using AI in their climate action work: Join our peer learning cohort
For social justice organizations weighing whether to introduce AI into their work, the initiatives in this mapping suggest a set of starting questions to guide their reflections:
- Have we listened closely enough to the communities most affected to determine together whether AI can be part of an appropriate response?
- What electricity, connectivity, hardware, and language constraints will our intended users actually be working within and have we designed for this context?
- Can we choose an architecture that makes community ownership technically possible?
These questions will be central to our upcoming peer-learning cohort, designed for practitioners working at the intersections of AI, climate action and social justice who are actively developing, implementing, or exploring AI tools in their work. Whether you’re experimenting with community-based environmental monitoring, designing AI systems for climate resilience, or considering if AI is the right tool for your context, the cohort will offer a space to reflect with peers, share experiences, and learn from diverse approaches across the Majority World.
Guided by The Engine Room, participants will explore emerging trends, practical challenges, and community-centred approach to AI design, data practices, and governance. Together, these discussions will inform a Responsible AI for Climate framework, co-designed with participants to support more context-sensitive, care-centred, and responsible uses of AI for climate action. Stay tuned for our open call to join the cohort. If you’d like to learn more or be notified when applications open, get in touch with us at lesedi@theengineroom.org.
We’ll also continue these conversations at Mozfest 2026. Join us for our session “Frugal AI grows like wild herbs”, where we’ll explore these initiatives, practice key dimensions of the emerging Responsible AI for Climate framework, and imagine what a forest of thriving alternative AI infrastructures could look like.
We would like to give special thanks to the twelve organizations who generously shared their stories for this study, as well as our regional researchers: Madhuri Karak, Radhika Jhalani, and Yosr Jouini, who conducted the desk research and interviews in Asia and North Africa.