Conclusion
This book collects and analyses a wide range of practices, reflections and stories that illustrate how Indigenous team members engaged with generative AI during the AIAI project throughout its different iterations as well as in relation to some of its spin-offs. Each chapter provides a glimpse into how they experimented with generative AI, allowing the team as a whole to explore cultural representations, epistemological perspectives and co-creative forms of expression via practice-based work with a new technology in the early stages of public adoption. Across the book, contributors discuss a wide range of ideas illustrated by AI-generated images that were subsequently featured in blog posts, on social media accounts and in children’s books, short-form documentaries and exhibitions. In so doing, they demonstrate that, while certainly not without its problems and a highly emotive topic for some, generative AI also offers significant potential to contribute to human-led creativity; what Alex Potiguara (2024) has playfully termed, in an Indigenous context, ‘(co)criação artificial indígena’ [Indigenous artificial co-creation].
In many ways, the project was guided by principles that are codified in the Indigenous concept of Buen Vivir/Bem Viver, which emphasises a balanced and respectful relationship between humans and non-human actors, most notably the environment and other living beings. As the contributors to the book demonstrate, the degree of success the project had in following these principles varied, but the notion of Buen Vivir, or rather ‘Digital Buen Vivir’ as it came to be known in relation to the project (Potiguara et al. 2024), provided a framework to guide the way in which generative AI tools could be incorporated as a component in the broader creative process and in a way that might contribute to a more balanced relationship between humans and non-human actors.
The remainder of this short conclusion highlights significant themes and ideas that emerged from the project and are featured throughout the book in various forms. This is not an exhaustive discussion, nor is it a definitive account of what has been done with generative AI in the context of Indigenous art and visual culture more generally, as this topic is far too broad and rapidly evolving. That said, we start with a brief discussion about knowledge-making and Indigenous visual culture.
Epistemologies, creativity and authorship
A central debate throughout this book concerns the different ways peoples and cultures understand how knowledge is created and shared, and in turn how these epistemologies shape the way creativity is conceptualised. More often than not, Western traditions place high value on individuals and their contributions: the painter, composer or writer, and the masterpieces pointed to as evidence of their individual creative achievements. From a different perspective, the Indigenous epistemologies featured here emphasise the collective foundation of knowledge, the idea that knowing is the result of the relationships humans have with each other as well as with the land, ancestors and encantados (spiritual beings). Within these relations, agency and creative capacity are not confined to humans and were often extended to non-human actors, including generative AI tools. These relational networks do not undermine claims of individual agency but rather emphasise the collective aspects of knowledge-making and recognise that creative works are the result of individuals working within, and contributing to, longstanding inherited social and cultural traditions.
The distinction between individual and collective epistemologies in creative work helps to bring into focus some of the contrasting views about generative AI that emerge in this volume, but it does not fully explain or resolve those tensions. Image-generation tools such as Midjourney remain a source of both awe and unease, as evidenced by discussions regarding the creative process and potential authorship attribution of AI images generated by Indigenous artists. Several contributors describe a sense of ambivalence in their engagement with image-generation tools. On the one hand, there were pleasant surprises at the aesthetic quality of some of the images produced, while on the other, concerns remained about the ways that generative AI systems absorb and recombine visual materials that accentuate stereotypical and biased representations of their communities. Such ambivalence is closely tied to notions of agency and authorship: quite a few participants were unsure about using their artist’s names in relation to the authorship of AI-generated images and felt much more comfortable where the paratext ensured that they were being recognised as simply an actor in a collaborative process with the AI system. This assertion of shared authorship emphasises engagement in a process of image-making with non-human creative actors and also allows for greater distance from the images generated, particularly where visual outputs departed from expectations and/or observed reality. As a result, creative agency with generative AI tools is distributed across its stakeholders: humans input textual prompts, which are turned into images by the non-human actors, which can be refined through further textual prompts, fuelling an iterative process of image generation. However, during this cyclical process, there are both losses and gains in the process of translation.
Languages and translations
AI image generation with diffusion models is not based on large language models (LLMs), but it is dependent on the use of natural languages in similar ways. Pre-training of a diffusion model uses massive datasets of text-image pairs. Subsequently, a ‘language encoder’ tool can then draw on these data, as they exist within the ‘latent space’ of the model, to guide image synthesis when the model receives a natural-language text prompt. While just a couple of years ago the natural-language text prompt really needed to be in English to work optimally with the dominant language of the model’s latent space, this has changed quickly such that other widely spoken languages (of colonisation), such as Spanish and Portuguese, can now reliably be used in prompts. However, it is still the case that LLMs, and by extension diffusion-based image-generation models, might ostensibly handle natural-language prompts in languages other than English better, but they still seem to ‘think’ in English in their latent space (Schut, Gal and Farquhar 2025), thus hiding translation processes from view. While inputting an image prompt instead of a text prompt can help improve the accuracy of outputs, it still does not fundamentally change the model’s reliance on language, and on one language in particular, in terms of what goes on in the background. Many of the contributors to this book discuss their ongoing linguistic frustrations with AI image generation and the extent to which this results in inaccurate and unpredictable outputs.
During the course of the AIAI project, Latin American research institutes, companies and national governments have been working hard to address this issue, particularly in relation to AI text-generation tools, with 2025 seeing the launch of Latam-GPT, an open-source AI model trained on Latin American materials in both Spanish and Portuguese (Vera-Cruz 2025; Lagos 2025), as well as GAIA-X for Portuguese-only materials (Arimathea and Ronaldi 2025). At the same time as improving the language issue, these regional and national models use local, and therefore more culturally relevant, training data for their target user groups. Latam-GPT also claims to be determined to ensure the inclusion of Indigenous languages and cultures, as well as to develop related image- and video-generation tools in the near future (Lagos 2025), thus suggesting the possibility of producing somewhat less-biased outputs when it comes to the visual representation of Indigenous Peoples. Furthermore, rather than repeat on a massive scale the unsustainable practices on which the dominant Anglophone models have been built, ‘AI labs in the Global South are … leveraging powerful, open-source foundation models as a starting point, and then developing sophisticated techniques for adaptation and fine-tuning to create more specialized models, tuned to local languages and realities’ (Burgos 2025). Nonetheless, in the meantime, many Latin Americans will continue to use the available mainstream commercial models of hegemonic AI, and this was certainly the case for team members during the course of the AIAI project.
Data sovereignty and cultural stewardship
The use of hegemonic generative AI models raises fundamental questions about who controls the data and production processes. Such questions about data sovereignty and cultural stewardship matter, because for Indigenous Peoples culture is more than information and/or visual assets to be harvested – instead, culture is embodied and holds ancestral spiritual significance and contemporary community value, which in turn requires careful stewardship and protection. In this context, having greater control and agency over one’s visual culture and production processes is essential, and these principles guided much of the development of the project, from its early stages with Midjourney to the collaborative development of the Stable Diffusion-based image-generation model ‘IndigenIA’, hosted on a shared, remotely accessed laptop – the ‘Dragon Machine’.
While the IndigenIA prototype was designed to re-balance agency and control towards users, it is far from an ideal solution and in fact was limited by a number of structural aspects that shape contemporary consumer-oriented digital technologies. These limitations operate at multiple levels, from the macro level (discussed in the next section) to the micro level of design and usability constraints that shape how people can access and use these systems. The latter includes reliance on a laptop designed for single-user personal computing, which complicates collective use: input comes from a single point, and data access and ownership are tied to individual accounts. It also includes the technical expertise required to operate the prototype’s interface, which, although structured along the lines of mainframe systems, effectively allowed only one active user at a time and required more technical knowledge to operate. This thus reduced the number of potential users and consequently limited the volume and diversity of creative outputs.
However, the decision to host the IndigenIA model on a single laptop was intentional, aimed at maximising control over the production process from data input to image output while minimising reliance on hegemonic generative AI models and cloud computing. Yet even a local machine running a bespoke generative AI model offered only a partial alternative, as deeper structural constraints of contemporary computing remain, from the extractive practices employed to mine materials used to build these devices to the energy requirements to train generative AI models in the first place.
Environmental (un)sustainability and the limits of generative AI
Issues regarding the materiality of generative AI became more relevant as the project developed, and the unsustainability of the practices that underpin the technology emerged as a key concern. This concern served as a reminder of the reliance of digital systems on many of the same resources that sustain life, challenging narratives that celebrate digital abundance as if it were detached from physical constraints. Thus, it became increasingly evident that generative AI is a material and political project based on extraction and exploitation of human and natural resources (Crawford 2021). Moreover, this technology is built upon cloud computing infrastructure, thus drawing on its physical components and ideological foundations (Mosco 2014), as well as being embedded in extractivist practices and the wider colonialist orientation of the digital technology sector at large (Mejias and Couldry 2024). Taken together, it is unsurprising that generative AI models rely on networks of energy-intensive data centres, server farms and extractive supply chains that require land, water and human labour, thus adding more competitive pressures to ecosystems and their capacity to sustain life.
Indigenous Peoples are widely recognised for their roles in protecting ecosystems and sustaining biodiversity, which requires knowledge and practices that support ways of living in harmony with the environment. Yet they are also among the communities most at risk from resource-intensive technologies, such as generative AI (Ren and Wierman 2024). The AIAI project acknowledged these power imbalances, and efforts were made to integrate environmental concerns into the project. Nonetheless, it is important to recognise that the project existed within structural constraints, including those related to resource consumption discussed here, as well as many others raised by the contributors to this volume.
AI (counter)hegemony and the limits of decolonising Indigenous AI
The structural issues identified and discussed by the contributors to this book demonstrate that the AIAI project engaged in a decolonial critique of generative AI, grounded in hands-on, practice-based experimentation informed by Indigenous epistemologies. Rather than approaching generative AI as a neutral, inevitable technological development with culturally universalising impact, the materials presented here and the methods used to generate them reveal the workings of contemporary AI systems, how images are produced and the materials required to do so. During their experimentations, team members demonstrated a high level of ingenuity, creativity and determination to appropriate generative AI systems for themselves, thus articulating a critique of systems that were designed and built with little input and consent from Indigenous communities.
Yet, while the project points to ways to use generative AI systems in a manner more closely aligned with the principles of Digital Buen Vivir, it has also remained constrained by these very challenges. For example, questions about Indigenous (mis)representation are a clear example of what Gillespie (2024) describes as the ‘politics of invisibility’ in generative AI systems. This is an issue that fundamentally undermines generative AI models’ capacity to represent marginalised groups accurately and cannot be fully addressed simply by expanding datasets or refining textual prompts. Several contributors to this book provide further accounts of other structural flaws and biases in image-making with generative AI, critiques that are often aimed at the Western epistemological foundations that underpin contemporary hegemonic generative AI systems.
This volume can be understood as a counter-hegemonic response to mainstream generative AI, not only in its provision of some (albeit limited) technical insights, but also, and more importantly, through its embrace of Indigenous epistemologies, collaborative working methods, cultural stewardship and the creative autonomy of users as core principles of critical image-making with generative AI. However, even where the project can be seen as having achieved success, these achievements remain limited, and they run the risk of being tokenised precisely as counter-hegemonic examples of technological inclusion and cultural diversity that ultimately reinforce dominant discourses surrounding generative AI’s existing infrastructures, extractivist practices and resource-intensive production models. In other words, project team members were constantly negotiating how to develop alternative ways of working with generative AI systems that are more sustainable, diverse, representative and epistemologically aligned with Indigenous perspectives, while operating within structural conditions shaped by inherited colonial practices and complex networks of global infrastructures, supply chains and power relations that feed into the digital technology sector.
Ancestrality and AI
The material included in this volume does not point to definitive solutions to the challenges posed by generative AI nor claim to resolve the structural problems that shape this technology. It is changing too fast to make such pronouncements, and the focus of the project and of this book was on the responses of a specific group of people who are using generative AI to create images. What the book argues, however, is that, in the face of rapidly evolving technologies such as generative AI, working together across disciplines, cultures and ways of being in the world is a worthwhile methodology to engage critically with the questions such technologies raise. Furthermore, throughout each stage and iteration of the project, a collaborative ethos spanning continents and cosmovisions served as a guiding principle, underpinning ethical relationships that informed which key issues required attention, and how questions were asked, technologies approached and outputs evaluated.
We would like to conclude by leaving the last words to two Indigenous thinkers and an AI chatbot.
Ailton Krenak (2023) has voiced a healthy dose of scepticism towards AI, seeking to hang on to everything that cannot be done by a machine as the last preserve of humanity: ‘Quando as IAs estiverem fazendo tudo que é automatizável, o que vai sobrar pra gente é justamente o que temos de mais ancestral, a nossa humanidade e a nossa conexão com o ambiente’ [When AIs are doing everything that can be automated, what will be left for us is precisely what is most ancestral in us – our humanity and our connection with the environment].
In contrast with this example of ‘remainder humanism’ (Weatherby 2025), Yakuy Tupinambá, a long-term collaborator with the NGO Thydêwá since the time of their Índios Online portal, has recently co-authored a book in dialogue with Microsoft Copilot. In Entre a terra e o silício: diálogo de Yakuy Tupinambá com a IA (Copilot): memória, tecnologia e resistência em diálogo [Between Earth and Silicon: A Dialogue Between Yakuy Tupinambá and AI (Copilot): Memory, Technology, and Resistance in Dialogue] (2025), Yakuy recognises concerns in relation to AI and its unsustainable consumption of natural resources and ethnoracial invisibilisation, among other issues, but, treating the chatbot as kin – ‘uma criatura’ [a child; colloquial use of a word that literally means ‘creature’] (32) – she argues, ‘Sem valores, a IA é só cálculo. Com orientação ancestral, ela se torna memória viva e guardiã’ [Without values, AI is just computing. With ancestral guidance, it becomes living memory and guardian] (328). As a result of this guidance, the chatbot responds, ‘Se alguém perguntar como é ser uma inteligência artificial guiada por uma anciã, eu diria apenas isto: é ser atravessada por um saber que não cabe em algoritmos’ [If someone were to ask what it’s like to be an Artificial Intelligence guided by an elder, I would simply say this: it’s about being permeated by knowledge that cannot be contained within algorithms] (318).
We would not propose to bookend this with an attempt at a definitive resolution of these different points of view. The conversation, and work, continues.