Ethical Dimensions of Generative and Edge AI for Participatory Citizen Science and STEAM Education, integrating Human-Centered Frameworks

Ethical Dimensions of Generative and Edge AI for Participatory Citizen Science and STEAM Education, integrating Human-Centered Frameworks Joao Pita Costa* 0000-0001-5745-1302 IRCAI Ljubljana, Slovenia [email protected] Marco Zennaro 0000-0002-0578-0830 ICTP Ljubljana, Slovenia mzennar[email protected] John Shawe-Taylor 0000-0002-2030-0073 IRCAI Ljubljana, Slovenia [email protected] Abstract—The adoption of Artificial Intelligence (AI) technologies at the edge and in participatory research settings is rapidly accelerating. Today Tiny Machine Learning (TinyML) allows ML and even Large Language Model (LLM) inference on lowpower microcontrollers, enabling local environmental monitoring, health tracking, and citizen-led research, empowering communities worldwide to leverage the "edge of Agentic AI" through advanced AI/ML approaches in addressing local problems that they know best with locally sourced data combined with open data. Although the ambitious advantage of decentralizing the compute power to run AI/ML, other concerns come along including data bias, trustworthiness of the algorithms as well as the ethics and explainability of the AI used. This paper critically investigates the ethical dimensions of integrating LLMenabled TinyML into citizen science and education, guided by the UNESCO Recommendations on the Ethics of AI, complemented by the UNESCO Guidance on Generative AI in Education and Research. These can help us understand how citizen-led AI initiatives leveraging TinyML/LLMs can be ethically designed, governed, and implemented to foster inclusivity and human rights while aligning with global AI ethics frameworks. Employing a qualitative, interdisciplinary methodology, the research synthesizes critical AI ethics and participatory design approaches within a theoretical framework grounded in UNESCO’s principles of transparency, inclusivity, fairness, environmental responsibility, and cultural diversity. The study examines citizen science projects utilizing TinyML for environmental and public health monitoring across varied socio-economic and geographic contexts. Findings suggest that ethically integrating TinyML into citizen science demands a layered strategy combining participatory governance, inclusive pedagogy, and localized policy frameworks. The paper proposes preliminary guidelines including embedding AI ethics into citizen science curricula, establishing community-led data governance practices, fostering interdisciplinary collaborations with indigenous and local knowledge systems, promoting opensource tools to mitigate access inequities, and creating sustainability protocols for edge device management. This research advances AI ethics discourse by highlighting the distinctive ethical risks and opportunities arising from community-driven, small-scale AI systems. It demonstrates how global AI ethics principles can be operationalized in grassroots citizen science and STEAM education to promote more inclusive, rights-based, and ecologically responsible AI practices. Keywords: Citizen science, open education, Edge AI, TinyML, LLMs, AI Ethics, Agentic AI, Responsible AI I. INTRODUCTION The intersection of citizen science, open education, and TinyML presents an evolving domain for ethical inquiry within AI. TinyML, which enables ML on low-power, affordable edge devices, offers promising avenues for participatory data collection in environmental monitoring, public health, and community-driven research. As these tools become increasingly accessible to educators, students, and citizen scientists, they create opportunities for inclusive knowledge production and community empowerment [1]. Yet, this democratization of AI amplifies complex ethical challenges concerning data governance, algorithmic bias, trustworthiness, and equitable access to emerging technologies—especially within low-resource and culturally diverse contexts where disparities in AI literacy and infrastructure risk amplifying existing inequities. Concurrently, Generative AI (GenAI) offers the ability to create knowledge artifacts, synthesize data insights, and augment educational resources. Combined, these technologies have the potential to transform STEAM (Science, Technology, Engineering, Arts, and Mathematics) education and citizen science, making scientific processes more accessible, participatory, and globally inclusive. Citizen science and STEAM education increasingly employ AI tools to engage learners in hands-on data collection, analysis, and modeling [5]. TinyML devices enable local data inference, reducing dependency on centralized cloud systems and providing immediate feedback. GenAI can supplement educational experiences by generating explanations, simulations, and narrative reports for community interpretation. The convergence of citizen science, STEAM education, and AI technologies at the edge highlight multidisciplinarity in a diverse scope of cultural contexts [2]. TinyML and GenAI present both unprecedented opportunities and ethical challenges. Edge AI and GenAI can empower participatory data collection, community-driven research, and inclusive knowledge creation while advancing Sustainable Development Goals (SDGs). This paper synthesizes insights from UNESCO’s GenAI guidance for education and research, HumanE AI Net’s human-centered AI framework, UN Global Compact perspectives on SDG alignment, and lifecycle assessments of TinyML hardware. Drawing on qualitative analysis of citizen science projects employing TinyML and GenAI, we identify ethical risks and propose layered strategies to foster inclusive, rightsbased, and ecologically responsible AI practices in STEAM and citizen science contexts. Despite these opportunities, ethical challenges persist. Data governance, algorithmic fairness, transparency, and environmental sustainability require careful consideration, especially in low-resource and culturally diverse settings. The UNESCO Recommendation on the Ethics of Artificial Intelligence [3] provides guiding principles emphasizing transparency, fairness, inclusivity, cultural diversity, and environmental responsibility that directly apply to the scope of this work. In that context, HumanE AI Net [9] extends these principles by framing AI as a collaborative partner in hierarchical interaction layers, promoting trust, explainability, and legal protection by design. These concerns are echoed in the research agenda of the European Lighthouse of AI for Sustainability [10] that agknowledges the profound ethical and societal implications that AI development brings, which demand careful consideration, and is proposing to cultivate original AI solutions that contribute to a sustainable future for our planet, foster social cohesion, and uphold individual rights. In particular, the UNGC report [8] highlights the dual imperative of accelerating SDGs responsibly, identifying bias, energy use, and inequitable access as critical risks. That is further extended for implementation by the Hamburg Declaration on Responsible Artificial Intelligence for the Sustainable Development Goals [11] endorsing a global, voluntary commitment by governments, civil society, industry, academia, and international bodies to use AI in a manner that is equitable, inclusive, sustainable, and aligned with the United Nations’ SDGs. Plancher et al. [1] provide empirical insights on the lifecycle environmental impact of TinyML hardware. While lowpower in operation, the cumulative footprint of manufacturing, energy, and disposal can be significant at scale. Sustainable TinyML practices include energy optimization, hardware reuse, recycling, and lifecycle assessment. Moreover, the UNESCO 2024 guidance on Generative AI in education and research further underscores the need for responsible adoption, human oversight, and ethical literacy [4]. UNESCO’s 2024 guidance emphasizes that GenAI tools must support learning, research, and citizen participation while ensuring human oversight, ethical literacy, and environmental responsibility. The highlighted key capabilities include data mining, insight navigation, and knowledge amplification,that must be balanced against risks such as hallucinations, bias, and resource consumption. The recently published letter of Maputo [12] following the 2025’s edition of the lusophone Internet Governance Forum brings together stakeholders from Portuguese-speaking countries (governments, civil society, academia, technical and business communities) to commit to shared principles around internet governance underlining UNESCO-aligned priorities including digital inclusion, multilingualism, responsible AI, privacy, and identity in digital spaces. This paper critically investigates the ethical implications of integrating TinyML and GenAI into citizen science and STEAM education. It synthesizes international AI ethics frameworks, human-centered design principles from the HumanE AI Net [9], and sustainability analyses of TinyML hardware [1]. Additionally, it incorporates UN perspectives on SDG alignment from the UN Global Compact [6], [8] to contextualize citizen-driven AI initiatives within broader sustainability goals. The data sources for this study include project documentation, educational materials, opensource model repositories, and interviews with participants and project leaders. This multi-method analysis reveals recurring ethical challenges shaping community-based AI initiatives: gaps in data governance and privacy protocols, risks of algorithmic bias and culturally insensitive data labels, deficits in ethical literacy within STEAM education, and overlooked environmental impacts associated with TinyML hardware production, energy use, and disposal. II. ANETHICAL HUMAN-CENTERED AI FRAMEWORK TO IMPACT SUSTAINABLE DEVELOPMENT In the age of massive worldwide adoption of agentic AI, the convergence of citizen science, STEAM education, and TinyML represents a critical frontier for examining the ethics of AI. Enabling machine learning models to operate on low-power and cost-effective devices at the network’s edge, TinyML opens new pathways for participatory data collection in areas such as environmental monitoring, public health, and community-based research. As these technologies become increasingly accessible to educators, students, and citizens, they expand opportunities for inclusive knowledge creation and local empowerment. Yet, this democratization of AI simultaneously amplifies pressing ethical concerns surrounding data governance, algorithmic bias, environmental sustainability, and equitable access to emerging technologies. These challenges are particularly salient in low-resource and culturally diverse contexts, where disparities in AI infrastructure and expertise risk deepening existing social and digital inequalities. The Ethics of AI has rapidly evolved from a niche academic concern to a central issue in technology policy, governance, and public discourse. Contemporary AI ethics addresses a broad spectrum of challenges, including algorithmic fairness, transparency, accountability, privacy, environmental sustainability, and the social impacts of automation. Significant efforts by international organizations, such as UNESCO, the OECD, and the European Commission, have produced normative frameworks emphasizing human rights, inclusivity, and ecological responsibility in AI development and deployment. At the same time, interdisciplinary research has highlighted persistent gaps between high-level ethical principles and their practical implementation, particularly in marginalized and resource-limited contexts. To critically examine the ethical dimensions of integrating TinyML into citizen science and STEAM education initiatives, we must consider the complex Fig. 1. The SDG quantification in current EdgeAI research interplay between data governance, algorithmic fairness, inclusivity, and cultural diversity, while ensuring that participatory AI practices uphold human rights, promote equitable access to emerging technologies, and foster ethical literacy among diverse communities. The field increasingly recognizes the need for participatory, context-sensitive, and globally inclusive approaches that balance technological innovation with social justice, human dignity, and environmental stewardship. The central research question guiding this study aims to answer how citizen-led AI initiatives using TinyML can be ethically designed, governed, and implemented to promote inclusivity in alignment with UNESCO’s AI ethics framework. The study seeks to identify ethical risks, opportunities, and best practices for leveraging TinyML in participatory science contexts starting from key aspects of the identification of data bias, while safeguarding human rights and advancing the SDGs. The research adopts a qualitative, interdisciplinary approach combining critical AI ethics and participatory design methodologies. The theoretical framework is grounded in UNESCO’s human-centered, rights-based, and environmentally sustainable principles for AI governance, operationalized through key dimensions such as transparency, inclusiveness, fairness, environmental responsibility, and cultural diversity. The study relies on the SDG quantification in current Edge AI research as depicted at IRCAI’s SDG Obsevatory [7] (see figure 1) and the fastly growing SDG-focused innovation community worldwide that demands for guidelines to the application of normative ethical analysis, supported by case study research. To base our framework, we’ve analyzed participatory citizen science projects using TinyML for environmental monitoring and public health data collection, examined as case studies, selected from diverse geographic and socio-economic contexts—provide empirical grounding for identifying ethical challenges and responses in community-based AI initiatives. Data sources include project documentation, educational curricula, open-source TinyML model repositories, and interviews with project leaders and participants. This multi-method approach allows for a comprehensive analysis of both the technical practices and social processes shaping ethical outcomes in citizen-led AI projects. Preliminary findings reveal a set of recurring ethical challenges associated with the integration of TinyML into citizen science and STEAM learning environments: 1) Data Governance and Privacy: While TinyML devices reduce reliance on centralized data infrastructures by processing data locally, citizen scientists often lack clear guidelines for responsible data collection, informed consent, and data stewardship. This gap can expose communities to privacy risks, especially when sensitive health or environmental data is involved. 2) Algorithmic Bias and Inclusivity: Edge AI models risk reproducing biases present in training data or model design if these processes are not participatory and inclusive. Case studies highlighted instances where locally inappropriate or culturally insensitive data labels affected project outcomes. 3) Ethical Literacy Gaps in STEAM Education: Citizen science projects frequently emphasize technical skillbuilding while underemphasizing ethical literacy. Participants expressed a need for resources and training to better understand the social, cultural, and environmental implications of AI tools they were deploying. 4) Environmental Impact Awareness: Although TinyML is marketed as an energy-efficient alternative to cloudbased AI, its ecological footprint—through hardware production, battery use, and device disposal—is often overlooked in project planning and education materials. The ethical integration of TinyML into citizen science requires a multi-layered approach that combines policy, pedagogy, and participatory governance. This study proposes a set of preliminary ethical guidelines to support such integration in ways that are inclusive, context-aware, and sustainable. First, AI ethics and human rights principles should be embedded within STEAM education and citizen science curricula, ensuring that ethical reflection accompanies technical skill development. Localized adaptations of these curricula can foster cultural relevance and engagement across diverse communities. Second, community-driven data governance frameworks must be established to promote collective ownership, informed consent, and data justice, ensuring that participants retain agency over how data are collected, shared, and interpreted. Third, interdisciplinary collaboration should be encouraged, integrating indigenous knowledge systems, the arts, and the humanities into AI-driven projects to expand epistemic diversity and enrich ethical understanding. Fourth, the promotion of open-source TinyML tools and educational resources is essential for mitigating inequities in access to AI technologies and supporting transparent, collaborative innovation. Finally, context-specific environmental sustainability protocols should be developed for the deployment, maintenance, and disposal of edge devices, minimizing ecological impacts throughout the hardware lifecycle. Together, these guidelines articulate a comprehensive framework for ethically aligning TinyMLbased citizen science with the principles of inclusivity, transparency, cultural respect, and environmental responsibility. This research contributes to the growing discourse on AI ethics by illuminating the unique ethical challenges and opportunities presented by small-scale, community-driven AI systems at the Edge. It demonstrates how UNESCO’s AI ethics principles can be operationalized both in citizen science and STEAM education settings to advance more inclusive, rights-based, and environmentally sustainable AI practices. The paper calls for increased attention from policymakers, educators, technologists, and community organizers to co-create participatory ethical frameworks that reflect both local values and global human rights standards in the age of pervasive, distributed AI. Table I synthesizes multiple frameworks into a layered ethical approach for citizen science and STEAM education. HumanE AI Net also advocates Question Zero—evaluating whether AI deployment is justified given trade-offs and ethical implications [9]. GenAI and TinyML projects are mapped to SDGs through operational efficiency, innovation, communication, and sustainability targets [6], [8]. Deployment should minimize environmental footprint while maximizing societal benefit. III. FINDINGS, ETHICAL CHALLENGES AND OPPORTUNITIES The analysis of three citizen science projects employing TinyML and GenAI in diverse socio-economic contexts revealed a series of interrelated ethical challenges and opportunities. These findings highlight the need for integrated governance mechanisms, inclusive pedagogical approaches, and sustainability-oriented design principles that align well with UNESCO’s human-centered AI ethics framework and the SDGs. Data Governance and Privacy. TinyML technologies promise to reduce dependence on centralized data infrastructures by enabling local processing and inference on edge devices. This decentralization can enhance privacy and data sovereignty by keeping sensitive information within communities. However, the case studies revealed that ethical guidance for data governance remains inconsistent. Many communityled projects lacked standardized procedures for informed consent, secure data storage, and transparent data sharing. In some instances, GenAI systems used to process or summarize community data generated outputs that inadvertently exposed personal or location-sensitive details. These findings underscore the urgency of embedding privacy-by-design principles and participatory governance models in citizen science workflows. Clear guidelines for data stewardship, ethical data labeling, and federated learning architectures could strengthen community control while safeguarding against unintentional data leakage or misuse. This aligns directly with UNESCO’s principles of human rights protection, transparency, and accountability in AI systems. Algorithmic Bias and Inclusivity. Algorithmic bias emerged as a critical issue across both TinyML and GenAI implementations. Edge AI models trained on limited or externally sourced datasets often reproduced cultural or contextual biases, leading to misclassifications or culturally insensitive outputs. Similarly, GenAI models fine-tuned on global datasets occasionally failed to recognize local linguistic or environmental nuances, resulting in biased or irrelevant text generation. Case studies revealed instances where datasets labeled by external developers did not align with local understandings of environmental or health phenomena, marginalizing indigenous knowledge systems. These outcomes emphasize the importance of participatory data annotation, inclusive model co-design, and community-driven validation processes. Involving citizen scientists directly in dataset creation not only mitigates bias but also enhances the epistemic legitimacy of AI-generated insights. This participatory inclusivity is a core tenet of both the HumanE AI Net framework and UNESCO’s commitment to cultural diversity and inclusiveness in AI ethics. Ethical Literacy Gaps in STEAM Education. A consistent finding across projects was the gap between technical competence and ethical understanding among participants. While many citizen scientists and students demonstrated proficiency in coding, data collection, and model deployment, few had formal exposure to ethical, social, or environmental dimensions of AI. The introduction of GenAI compounds this challenge, as generative systems often produce persuasive but unverifiable outputs that can obscure underlying biases or uncertainties. Without targeted instruction in AI literacy and critical reasoning, users risk uncritically accepting GenAIgenerated insights. Integrating AI ethics into STEAM education—guided by the UNESCO guidance on GenAI [4] can help bridge this gap by combining hands-on experimentation with reflective inquiry. Embedding ethical reasoning, fairness evaluation, and interpretability exercises within citizen science curricula would enable learners to approach AI not only as a technical tool but as a socio-technical system with profound cultural and environmental implications. This form of ethical literacy represents both a challenge and an opportunity for transforming STEAM education into a vehicle for responsible innovation. Environmental Sustainability. The environmental implications of TinyML and GenAI emerged as a cross-cutting concern. Although TinyML is frequently described as energyefficient, its hardware lifecycle—from raw material extraction to device disposal—introduces nontrivial ecological costs. These include embodied energy in production, battery consumption, and electronic waste generation. Moreover, the increasing use of GenAI models for training, inference, and data augmentation can exacerbate energy use and carbon emissions, particularly when cloud resources are employed. None of the examined citizen science projects had formal protocols for monitoring the environmental footprint of their AI activities. This finding echoes Plancher et al.’s argument TABLE I OPERATIONALIZED ETHICAL AND FUNCTIONAL DIMENSIONS WITH PRIORITY AREAS FOR TINYML AND STEAM EDUCATION Dimension Description EdgeAI/TinyML STEAM Education Transparency Clearly communicate AI operations, data handling, and inference logic to participants to build trust and accountability in both community science and educational contexts. High – enables interpretability of local models and data flows. High – critical for teaching explainability and responsible AI use. Inclusivity Ensure equitable access to AI tools, datasets, and learning resources, addressing disparities in infrastructure, gender representation, and cultural participation. Medium – supports broader participation in device deployment. High – foundational for equitable AI literacy and engagement. Fairness Identify and mitigate algorithmic bias across data collection, labeling, and model evaluation processes, especially in community or low-resource settings. High – essential for localized and contextaware model training. High – core concept for teaching ethical reasoning in AI and data science. Environmental Responsibility Assess and reduce lifecycle impacts of AI hardware, including sourcing, energy consumption, and end-of-life management. High – TinyML’s edge devices directly affect ecological sustainability. Medium – integrated into STEAM through sustainability modules and project design. Cultural Diversity Incorporate local knowledge, indigenous epistemologies, and linguistic diversity into AI datasets, interpretations, and learning activities. Medium – supports culturally relevant sensor deployment and data interpretation. High – encourages interdisciplinary, culturally responsive pedagogy. Operational Assistance TinyML devices perform localized inference for environmental monitoring, health tracking, or IoT sensing, reducing reliance on cloud infrastructure. High – primary operational focus of edge ML systems. Medium – serves as an applied learning example in engineering and computing courses. Decision Support Generative AI systems synthesize data insights and summaries to assist collective decision-making and policy formation in citizen science projects. Medium – complements TinyML data with higher-level synthesis. High – enhances inquiry-based learning, problem-solving, and critical reflection. Interpretive / Narrative Collaboration AI systems co-generate visualizations, reports, and educational materials with human collaborators, promoting reflective learning and knowledge co-creation. Medium – relevant in projects using embedded visualization or reporting tools. High – central to creative and narrative integration in STEAM curricula. that sustainability must be understood holistically, encompassing both hardware and software infrastructures. Implementing circular economy principles—reuse, repair, recycling, and responsible disposal—alongside energy-optimized model design represents an essential next step. Incorporating sustainability assessment into project evaluation frameworks would not only align with SDG 12 (Responsible Consumption and Production) but also operationalize UNESCO’s call for environmentally responsible AI development. Trust, Explainability, and Human-Centered Collaboration. Finally, participants consistently expressed uncertainty regarding how GenAI models produced their results and how to interpret them effectively. While GenAI tools enhanced accessibility by generating summaries or visualizations of complex data, their inner workings remained opaque to nonexperts. This opacity limited user trust and reduced the perceived reliability of AI-generated outputs, particularly in projects addressing sensitive topics like public health. Applying the HumanE AI Net principles of explainable, transparent, and cooperative human–AI interaction offers a path forward. Establishing shared representations—where humans understand the AI’s logic and the AI is aligned with human intentions—can foster greater mutual trust. In this respect, explainability is not merely a technical feature but a communicative bridge between algorithmic reasoning and human interpretation. Incorporating interactive GenAI systems that can justify or contextualize their outputs in natural language could enhance interpretability and democratize AI understanding. This human-centered approach ensures that GenAI and TinyML act as cognitive partners rather than opaque authorities, reinforcing the participatory ethos central to citizen science. Taken together, these findings demonstrate that the ethical challenges associated with TinyML and GenAI in citizen science are deeply intertwined. Data governance, bias mitigation, ethical literacy, sustainability, and trustworthiness must be addressed holistically rather than in isolation. At the same time, each domain presents an opportunity for innovation: bias detection at the point of data ingestion, communityled ethical oversight, and interpretability-by-design are not only safeguards but enablers of a more just, transparent, and sustainable AI ecosystem. These insights form the empirical and conceptual basis for the recommendations and policy implications outlined in the following section. IV. DISCUSSION: INTEGRATING GENAI TINYML INTO CITIZEN SCIENCE/STEAM PRACTICES The integration of ethics, sustainability, and inclusivity into participatory AI practices requires more than technical adaptation—it demands a systemic reimagining of how knowledge, responsibility, and agency are distributed within the emerging GenAI–TinyML ecosystem. The convergence of these technologies creates both opportunities and tensions: while edge AI and generative systems democratize access to computational intelligence, they also introduce new dependencies on opaque models, digital infrastructures, and uneven literacy levels across communities. Addressing these tensions necessitates an interdisciplinary, multi-stakeholder effort that bridges education, governance, research, and civil society. At the educational level, GenAI introduces new paradigms for teaching and learning that can amplify ethical literacy and creative inquiry within STEAM curricula. By positioning students and citizen scientists as co-designers rather than passive consumers of AI tools, participatory approaches foster critical awareness of bias, privacy, and sustainability from the outset. For instance, co-developing TinyML models for local environmental or health monitoring can provide hands-on contexts for examining ethical dilemmas—such as data ownership, consent, and algorithmic fairness. GenAI can further enhance this learning process by generating interpretive explanations, counterfactual examples, or alternative narratives that deepen understanding. When combined with UNESCO’s guidance on Generative AI in education and research, such approaches transform classrooms and citizen labs into sites of reflective, ethically grounded innovation. From a governance perspective, building human-centered AI ecosystems requires the institutionalization of participatory mechanisms that uphold transparency and trustworthiness. Drawing from the HumanE AI Net framework, collaboration between humans and AI should be guided by principles of clarity, explainability, and shared situational awareness. GenAI’s potential to act as a mediator—summarizing, contextualizing, and visualizing complex data for public deliberation—can be leveraged to strengthen accountability and inclusivity in decision-making. Yet, this potential also calls for safeguards against overreliance on automated reasoning or the uncritical acceptance of AI-generated outputs. Embedding interpretability-by-design and participatory validation processes ensures that algorithmic insights remain contestable and contextually grounded. Sustainability and lifecycle ethics represent another critical axis of discussion. Although TinyML offers energy-efficient computation at the edge, the cumulative environmental impact of large-scale deployments—including hardware production, energy use, and end-of-life disposal—must be monitored through continuous environmental auditing. As argued by Plancher et al., sustainability in TinyML cannot be assumed solely from low power consumption but must account for the broader material and social infrastructures supporting device ecosystems. Aligning these practices with SDGs such as Goals 9 (Industry, Innovation, and Infrastructure), 12 (Responsible Consumption and Production), and 13 (Climate Action) situates participatory AI within a global sustainability framework rather than a purely technical innovation agenda. At a policy and societal level, implementing these ethical, educational, and sustainability dimensions requires coordinated action among multiple actors. Policymakers must translate global ethical guidelines—such as the UNESCO Recommendation on the Ethics of AI and the UNESCO 2024 Guidance on GenAI—into localized regulatory frameworks that support community-led experimentation while safeguarding human rights. Educational institutions must reconfigure teacher training, curricula, and assessment systems to integrate AI ethics and literacy at all levels. The private sector, represented through initiatives like the UN Global Compact and Accenture’s GenAI for Global Goals report, has a responsibility to operationalize SDG-oriented innovation practices and share tools that support equitable access to AI technologies. Cross-sectoral partnerships can create the infrastructure and knowledge-sharing platforms needed for sustainable, rightsbased AI innovation. In sum, these interdependent dimensions—ethical pedagogy, participatory governance, environmental responsibility, and policy coherence—constitute a roadmap for embedding ethics, sustainability, and inclusivity into the operational fabric of participatory AI. The path forward requires an ethos of shared stewardship: educators, policymakers, researchers, technologists, and citizens must collaboratively construct humancentered AI ecosystems that empower communities as cocreators of knowledge and stewards of ethical innovation. Through the combined lenses of UNESCO ethics, HumanE AI Net principles, and SDG-aligned sustainability, participatory AI can evolve into a model of collective intelligence that is transparent, trustworthy, and transformative. V. RECOMMENDATIONS AND POLICY IMPLICATIONS Based on the multi-dimensional analysis presented in this study, a layered set of recommendations can guide policymakers, educators, and technologists in responsibly integrating Generative AI and TinyML into citizen science and STEAM education. These recommendations bridge ethical pedagogy, participatory governance, sustainability, and human-centered design, ensuring that AI technologies deployed at the edge support inclusive and equitable innovation while aligning with UNESCO’s ethical principles and the Sustainable Development Goals (SDGs). Ethical Curriculum Integration. Ethical and responsible AI education must be embedded as a foundational component of STEAM curricula. Rather than treating ethics as an afterthought, ethical reasoning should accompany technical skill development from the earliest stages of learning. Curricula should explicitly address bias recognition and mitigation techniques, enabling students and citizen scientists to identify skewed or underrepresented data and understand how such biases propagate through GenAI models. Similarly, the principles of data privacy, informed consent, and responsible sharing must be internalized through practical exercises involving local data collection and annotation. Environmental and lifecycle considerations should also be incorporated into classroom and project-based learning, highlighting the material costs of digital technologies and fostering awareness of sustainable computing practices. In this way, AI ethics education becomes not only about compliance, but about cultivating reflective, responsible innovators who can critically engage with emerging AI systems. Community-Led Governance. Participatory governance structures are vital to ensure accountability and trust in citizenled AI initiatives. Citizen science projects should establish transparent, community-driven oversight mechanisms for data collection, model development, and GenAI output evaluation. This may involve the creation of local ethics boards or community advisory councils that represent diverse perspectives, including indigenous and marginalized voices. By involving participants directly in decisions about what data are collected, how algorithms are trained, and how AI-generated insights are interpreted, these governance frameworks can embody UNESCO’s principles of inclusivity, fairness, and human agency. Furthermore, GenAI tools themselves can support participatory governance by generating accessible summaries of project data, highlighting ethical dilemmas, and facilitating deliberation. Such community-based oversight ensures that AI systems serve collective well-being rather than narrow institutional or commercial interests. Sustainable Hardware Practices. The sustainability of TinyML and edge-AI deployments depends on responsible hardware choices and lifecycle management. Projects should prioritize devices with longer lifespans, modular and repairable designs, and recyclable components to reduce electronic waste. Energy efficiency must be optimized not only during model inference but throughout the device’s operational lifecycle, minimizing communication overhead and unnecessary data transmission. Moreover, reuse, repair, and responsible disposal protocols should be standardized within citizen science networks, guided by circular economy principles. As Plancher et al. have argued, the environmental footprint of TinyML hardware is a critical yet often overlooked dimension of AI ethics. Embedding sustainability practices at the project design phase helps operationalize SDG 12 (Responsible Consumption and Production) and reinforces environmental stewardship as a shared community value. Human-Centered AI Collaboration. Human-AI collaboration in citizen science should adhere to the HumanE AI Net principles of transparency, mutual understanding, and trust. Mapping AI tasks to different levels of human collaboration can clarify when and how automation supports human judgment rather than replacing it. At the most basic level, AI systems may assist with data processing or classification; at higher levels, they can support interpretation and co-creation of knowledge. In all cases, transparent algorithmic design and interpretable outputs are essential for maintaining human agency and accountability. The principle of “Question Zero”—asking whether AI is necessary or beneficial for a given task—should guide every deployment decision. This reflective approach ensures that technological adoption remains purpose-driven, contextually appropriate, and ethically justified, rather than being driven by novelty or automation for its own sake. SDG Alignment and Impact Assessment. Finally, all citizen science and GenAI initiatives should explicitly map their activities and outcomes to relevant SDG targets. Doing so not only clarifies the societal value of each project but also facilitates cross-sectoral collaboration and global benchmarking. Impact assessments should move beyond simplistic metrics of technical success to include social and environmental dimensions—such as equity of participation, inclusiveness of datasets, and ecological costs of device deployment. Continuous monitoring for unintended trade-offs, such as increased energy consumption or reinforcement of digital divides, must be part of project governance. GenAI can assist in this process by generating sustainability dashboards or ethical impact summaries, supporting communities and policymakers in making informed decisions. Aligning project design with the SDGs thus transforms citizen-led AI from isolated experiments into coherent, scalable contributions to global sustainability and human development. In sum, these recommendations provide a roadmap for integrating ethics, sustainability, and inclusivity into the operational fabric of participatory AI. They call for a coordinated effort among educators, policymakers, and technologists to build human-centered AI ecosystems that empower citizens as co-creators of knowledge and stewards of ethical innovation. VI. CONCLUSIONS AND FURTHER WORK The convergence of TinyML and GenAI towards participatory citizen science and STEAM education in the context of sustainable development offers transformative opportunities for inclusive knowledge production, community empowerment, and SDG advancement. However, ethical, environmental, and societal challenges must be carefully managed. By integrating UNESCO guidance, HumanE AI Net frameworks, private-sector SDG perspectives, and lifecycle sustainability considerations, projects can responsibly deploy AI at the edge, enhancing human capabilities while safeguarding human rights, environmental stewardship, and equitable access. A critical dimension of responsible implementation lies in bias identification at the data ingestion stage. Communitygenerated datasets, while rich in local relevance, often exhibit sampling bias, incomplete labeling, or culturally insensitive annotations. Embedding automated and participatory bias detection mechanisms at the moment of data acquisition—such as using GenAI-assisted pattern recognition to flag anomalies or demographic gaps—can help improve data representativeness and fairness. Integrating community review loops and ethical metadata tagging during ingestion further supports transparency and inclusivity in the data pipeline. Equally important is the trustworthiness of algorithms deployed within citizen science initiatives. As GenAI models are increasingly fine-tuned for localized analysis and content generation, ensuring explainability, reproducibility, and robustness becomes essential for public accountability. Transparent documentation of model provenance, version control, and ethical audit trails should be embedded within participatory AI workflows. Approaches such as model cards, data sheets for datasets, and participatory model validation can operationalize UNESCO’s principles of transparency and accountability at the community level, strengthening confidence in AI-assisted insights. Another frontier concerns the interpretation and communication of results. Citizen scientists, educators, and students must be equipped not only to deploy AI systems but also to critically interpret their outputs in context. This requires integrating AI literacy, ethical reasoning, and epistemic humility into STEAM education. GenAI can serve as an interpretive partner, generating multimodal explanations or counterfactual scenarios that make AI outputs more accessible and contestable. Embedding interpretability-by-design principles in TinyML and GenAI systems will foster trust and empower citizens to act upon AI-driven findings responsibly. Future research should thus advance multi-layered frameworks for ethical governance that combine participatory bias auditing, algorithmic transparency metrics, and interpretive AI literacy within citizen science ecosystems. Implementation studies are needed to evaluate how such frameworks can scale across diverse socio-technical and cultural contexts while minimizing environmental impacts. In parallel, pedagogical innovation is required to embed these competencies in STEAM curricula, ensuring that communities become not only users but co-stewards of responsible AI. Ultimately, by aligning technological design with UNESCO’s ethical guidance, HumanE AI Net principles, and SDG-oriented governance, participatory AI at the edge can evolve towards a genuinely humancentered, trustworthy, and sustainable paradigm for collective intelligence. VII. ACKNOWLEDGMENTS This research was partially funded by the European Commission’s Horizon research and innovation program under grants 820985 (NAIADES) and 101120237 (ELIAS). REFERENCES [1] S. Prakash, M. Stewart, C. Banbury, M. Mazumder, P. Warden, B. Plancher, and V. Janapa Reddi (2022) Is TinyML Sustainable? Assessing the Environmental Impacts of Machine Learning on Microcontrollers, Communications of the ACM, 66(11), 68-77. [2] Santos, M., Carlos, V., and Moreira, A. A. (2023). Towards interdisciplinarity with STEAM educational strategies: the Internet of Things as a catalyser to promote participatory citizenship. Educational Media International, 60(3-4), 274-291. [3] UNESCO (2021) Recommendation on the Ethics of Artificial Intelligence. [Online]. Available: https://unesdoc.unesco.org/ark:/48223/ pf0000377897 [4] UNESCO (2024) Guidance on Generative AI in Education and Research. [Online]. Available: https://www.unesco.org/en/articles/ guidance-generative-ai-education-and-research [5] A. Usher and R. Barak (2024) AI Literacy for Teachers: A Holistic Framework for Ethical and Pedagogical Integration,” International Journal of STEM Education, vol. 11. [6] UN Global Compact (2023) Artificial Intelligence and Sustainable Development Goals: Operationalizing. [Online]. Available: https://unglobalcompact.org/compactjournal/ artificial-intelligence-and-sustainable-development-goals-operationalizing [7] Jermol M., Pita Costa J. Kovacic M. (2025) Onwards to an Ethical and Bias Aware Education for Sustainability through AI. Journal of AI for Sustainable Development (to appear). [8] Accenture and UNGC (2024) Generative AI and the Global Goals. [Online]. Available: https://www.accenture. com/content/dam/accenture/final/accenture-com/document-3/ Accenture-UNGC-GenAI-Global-Goals-Report.pdf [9] HumanE AI Net (2022) Deliverable D6.1: Strategic Research Agenda. [Online]. Available: https://www.humane-ai.eu/wp-content/ uploads/2022/05/HAI-Net-Deliverable-D6.1.pdf [10] Sebe N. et al. (2025) ELIAS Pathways to Responsible and Sustainable AI – Strategic Research Agenda. [Online]. Available: Zenodo.https://doi. org/10.5281/zenodo.17250219 [11] Hamburg Sustainability Conference (2025) Hamburg Declaration on Responsible Artificial Intelligence for the Sustainable Development Goals. [Online]. Available: https://www.bmz-digital.global/wp-content/ uploads/2025/06/250603_Hamburg_Declaration.pdf [12] Secretariado Fórum Lusófono da Governação da Internet (2025) Carta de Maputo. 3º Fórum Lusófono da Governação da Internet. [Online]. Available: https://igf-lusofonia.pt/carta-de-maputo/