Frugal AI can support SDGs by expanding who builds and benefits from AI

Low-cost edge artificial intelligence (AI) could help communities in developing countries move from being AI users to creators of locally relevant technology, according to a new study published in Sustainability which finds that Tiny Machine Learning can lower barriers to AI innovation while supporting sustainable development.

The study, Empowering Local Frugal Edge AI Innovation Based on Participatory Citizen Science in Developing Countries, examines 50 participatory TinyML initiatives across 22 countries and finds that low-power, low-cost AI tools can support local skills, grassroots entrepreneurship and Sustainable Development Goal-aligned innovation when combined with citizen science, open education and institutional support.

TinyML offers a low-cost route to inclusive AI innovation

Governments, universities and development organizations are looking for ways to close the global AI divide before the 2030 deadline for the United Nations Sustainable Development Goals (SDGs). While AI has become imperative to economic growth, climate planning, healthcare, agriculture and public services, its benefits remain unevenly distributed. Low-resource regions often face poor connectivity, limited computing infrastructure, high costs and shortages of technical training.

Tiny Machine Learning, or TinyML, offers a practical way to address that imbalance. Unlike large cloud-based AI systems that require powerful servers and continuous internet access, TinyML allows machine learning models to run on low-cost microcontrollers and embedded devices. These systems consume only small amounts of energy and can operate close to where data are produced, making them useful in environments with limited power, connectivity or technical infrastructure.

That edge-based model fits the wider principle of Frugal AI. Frugal AI focuses on systems that are affordable, resource-efficient and adapted to local conditions. In developing countries, where high-performance computing and cloud infrastructure are often difficult to access, such systems can widen who gets to build and deploy AI. They can also reduce environmental costs by limiting energy use and dependence on large data centers.

The study presents Frugal Edge AI as a pathway for community-driven innovation, where citizens, students, local organizations and entrepreneurs can define problems, collect data, build models and test solutions. This participatory approach shifts AI development away from a top-down model dominated by large companies and high-income regions.

Citizen science is crucial to that shift. The paper defines participatory citizen science as an approach where non-experts take part in data collection, problem framing and knowledge production. When paired with TinyML, citizen science enables local participants to go further than data gathering. They can train and deploy models on inexpensive hardware, creating AI systems for agriculture, environmental monitoring, public health, assistive technologies and local infrastructure.

The study identifies strong relevance to several development goals, including quality education, reduced inequalities, decent work, innovation, climate action and sustainable communities. TinyML projects reviewed in the paper include applications for soil nutrient detection, crop disease identification, irrigation prediction, animal movement monitoring, low-cost medical screening, air-quality monitoring, waste classification, fire detection, beekeeping support and assistive tools for people with disabilities.

These projects are not simply classroom exercises, they show how low-cost hardware, open educational resources and hands-on workshops can turn technical learning into practical prototypes. In several cases, participants moved from introductory training to functional systems within days or weeks. That speed matters because it shortens the path from learning to experimentation, allowing communities to test AI solutions against immediate local needs.

Citizen science builds skills, prototypes and grassroots entrepreneurship

The paper is based on a qualitative multiple-case analysis of 50 participatory TinyML initiatives from 22 countries. The cases were drawn from public repositories and program archives, including TinyML4D, TinyMLedu and the ITU AI for Good Innovation Factory. The sample covers workshops, student projects, community deployments, research demonstrations and early-stage entrepreneurial efforts.

The study finds that participatory TinyML initiatives have a strong capacity-building effect. Participants gained practical skills in sensor integration, data collection, data labeling, model training, and on-device inference. They also improved their understanding of machine learning concepts, the difference between cloud and edge AI, and the role of Frugal AI in low-resource settings.

Nearly all reviewed initiatives produced functional prototypes. Many were lab-tested rather than fully deployed, but the study treats this as consistent with the education-focused nature of the programs. More than 85 percent of the cases remained at the functional prototype stage, while about 15 percent progressed toward field deployment or scaled implementation. This indicates that participatory TinyML is effective at lowering the barrier to experimentation, but the transition to sustained deployment remains limited.

The entrepreneurial signal was clear but still early. More than half of the cases were tied to competitions, showcases or similar early-stage activities. Fewer than 20 percent showed incubation uptake or venture formation. The paper argues that these ecosystems function primarily as entrepreneurial catalysts: they help participants recognize opportunities, build proof-of-concept systems and gain visibility, but they often require additional support to become sustainable businesses.

Academic institutions were involved in more than 60 percent of the cases, often serving as the main facilitators of technical training, mentorship and access to resources. The study portrays universities and research centers as trusted hubs that can connect education, innovation and community engagement. In low-resource environments, they can provide infrastructure, open curricula, workshop spaces and links to local partners.

The research also highlights the role of global platforms and competitions. Initiatives such as TinyML4D and TinyMLedu provide educational resources and public venues where participants can present work, receive feedback and connect with mentors. The AI for Good Innovation Factory adds a broader startup pipeline by connecting SDG-focused AI ventures with investors, international organizations and technical experts.

These networks matter because technical training alone is not enough to sustain innovation. The study finds that prototype refinement and entrepreneurial activity are more likely when participants have access to follow-up mentorship, maker spaces, innovation hubs, seed funding or competitions. Without those supports, promising prototypes may remain isolated demonstrations.

The paper also identifies a missing baseline for technical progression. Many initiatives teach participants how to build simple TinyML prototypes, but fewer provide a clear pathway toward deployment, maintenance, orchestration and business viability. To move from workshop projects to durable solutions, participants need skills in embedded systems, sensor integration, model optimization, data governance, system reliability and low-power deployment.

A major challenge is scaling beyond single devices. The study notes that many projects succeed at the level of individual prototypes but do not address the operational complexity of managing networks of distributed edge devices. That includes updating models, coordinating data flows, monitoring system performance and maintaining hardware over time. Without these capabilities, TinyML projects may struggle to become reliable public services or viable ventures.

The study finds that Frugal Edge AI can support digital sovereignty in developing countries. By keeping data processing local and reducing dependence on foreign cloud infrastructure, TinyML can help communities retain more control over data, models and services. This is especially important in regions where connectivity is unreliable, data governance frameworks are evolving and dependence on external platforms can limit autonomy.

Ethical governance and institutional support remain weak links

The study’s strongest warning is related to ethics and governance. While participatory TinyML initiatives show high innovation potential, explicit ethical and governance practices appeared in less than 10 percent of the cases reviewed. Most projects relied on informal or ad hoc approaches rather than structured frameworks for consent, data protection, fairness, accountability or risk assessment. This gap is significant because many community AI projects collect sensitive or locally important data.

Applications in health, agriculture, environment and livelihoods can affect privacy, economic decisions, social inclusion and public trust. Citizen-generated datasets may also reflect unequal participation, local power imbalances or hidden bias. Without formal safeguards, participatory AI can reproduce some of the same problems associated with large-scale AI systems.

The paper evaluates the initiatives against principles associated with UNESCO’s AI ethics framework, including human-centered development, transparency, sustainability, fairness, inclusion, accountability and capacity-building. The strongest alignment was found in education, open resources, participation and sustainability. The weakest areas were formal governance, data protection, post-deployment monitoring and structured fairness checks.

The study recommends that ethics become a required part of TinyML workshops and project templates. This includes standard fields for data provenance, consent, potential harms, fairness checks, hardware reuse, environmental footprint and community accountability. Responsible AI should not be added after a prototype is built. It should be embedded from the beginning of problem definition, data collection and model design.

Environmental sustainability is a stronger point for TinyML. Because models run on low-power devices and avoid constant cloud processing, these systems can reduce energy use and infrastructure dependence. Hardware reuse and low-cost components can also lower waste and make experimentation more affordable. The paper positions Frugal Edge AI as both an inclusion strategy and a response to concerns over the environmental cost of large AI models.

To strengthen impact, the study introduces the Frugal Edge AI Lean Canvas, a planning framework for innovators working in low-resource settings. The canvas helps teams define the development challenge, target beneficiaries, edge AI value proposition, data strategy, deployment pathway, partners, costs, impact metrics, revenue model and ethical risks. Its purpose is to align technical choices with SDG outcomes, local constraints and responsible governance.

The study says such tools are needed because conventional startup models often assume stable connectivity, cloud infrastructure and revenue-driven scaling. Those assumptions do not fit many low-resource or sustainability-focused AI deployments. A Frugal Edge AI model must account for intermittent connectivity, limited energy, local data ownership, low-cost maintenance, public or nonprofit partnerships and measurable social impact.

The paper also calls for stronger policy support. National AI strategies and digital transformation plans often prioritize large-scale AI infrastructure and cloud-based services. The study argues that governments should also support participatory Frugal AI, especially through STEM education, vocational training, digital public goods, micro-grants, local innovation hubs and public-private-academic partnerships.

Education systems have a major role to play. Integrating TinyML and citizen science into STEM and vocational curricula could help students learn AI through real-world problem-solving rather than abstract instruction. Open educational resources, low-cost hardware kits and multilingual materials can make these programs more inclusive. Gender-aware facilitation and community-driven problem selection can further expand participation among groups historically underrepresented in technical fields.

The findings also point to the importance of long-term research. The paper notes that evidence on skill retention, employment outcomes, venture survival and community-level impact remains limited. More longitudinal studies are needed to track whether participants continue using TinyML skills, whether prototypes become services or businesses, and whether community AI systems deliver lasting benefits.