Introduction

Thea Pitman, Sandra De Berduccy and Andreas Rauh

This book, as it states on the cover, is about Artificial Intelligence and its relationship to art and Indigeneity in particular. Definitions of Artificial Intelligence are slippery because they are continuously changing, either to retrofit previous smart applications now that AI has become ‘meaningful’ to the general public (this is where all sorts of technology, such as satellite-navigation routes and social media suggestions, start to be referred to as AI, even if we did not previously refer to them as such) or, conversely, to exclude previous applications in order to site AI permanently at the forefront of innovation (this is where references to AI are often currently really referring only to generative AI, as is the case for the title of this book). But definitions of AI are also slippery because it is much more than (just) a technology: as Paola Ricaurte (2022) has argued, ‘At the micro-political level, AI mediates the relation with the self, our intersubjective relations, as well as our relationships with the world, our representations of reality and the shared imaginaries of the present and the future’ (728), and, in the words of Rita Raley and Jennifer Rhee (2023), ‘AI is not a technology – it is an epistemology and also a form of governmentality and political economy’ (196). Others, such as Emily M. Bender and Alex Hanna, authors of the book The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want (2025), are more dismissive of the use-value of the term, describing AI as simply ‘a marketing term used to sprinkle some magic fairy dust that brings venture capital and in theory brings the consumer’ (Jaffe 2025). In this book we opt for the same ‘linguistic pragmatism’ in relation to our use of the term as that espoused by Raley and Rhee (188), while maintaining a healthy dose of circumspection about the ways in which it is being deployed.

The main object of our attention is the recent emergence of AI image-generation tools. In early 2021, OpenAI released the text-to-image generative AI tool DALL-E to the general public for the first time. Their second, significantly enhanced iteration, DALL-E 2, followed in April 2022, and, together with the launch of similar tools Midjourney (Midjourney Inc, July 2022) and Stable Diffusion (Stability AI, August 2022), the three companies created a veritable wave of hype and anxiety in the press around their potential applications and impacts. In the following paragraphs we offer a very brief synthesis of how image generation works in order to introduce key terms and concepts before going on to outline the more specific focus of this book.¹

These new and powerful image-generation tools are all based on the process of ‘diffusion’, in comparison with earlier tools such as convolutional neural networks (CNNs) and generative adversarial networks (GANs). CNNs, such as DeepDream, did not generate images from scratch but instead sought to enhance existing images by finding and accentuating patterns in them, resulting in a psychedelic, ‘dream-like’ aesthetic referred to as ‘algorithmic pareidolia’ (Somaini 2023, 89–91). The tool was created by Google engineers and released in 2015; early iterations produced still images, while later ones offered the possibility of producing animations of the process. GANs, like StyleGAN, did technically generate new images in a competitive process between two separate neural networks (a generator and a discriminator) sampling material from a ‘latent space’ – a high-dimensional mathematical space where each point represents a compressed set of features that can be decoded into an image. However, although GANs could produce photorealistic results, for lay users their applications were relatively limited, often focusing on tasks like producing variations of faces, objects or scenes.²

When we talk about (neural) networks in relation to all these tools, we are referring to highly complex structured architectures or models that use the rules provided by algorithms to learn patterns. An ‘image-generation model’ is the trained version of that network that produces output images. Diffusion models are trained on truly massive datasets of image-text pairs (typically scraped from sites such as Instagram and Pinterest and inevitably also dependent on the language in which the textual description has been written [Bender and Hanna 2025, 107–9]), and this training process consumes a staggering amount of energy and other resources (Ren and Wierman 2024). During their training they are taught to remove random ‘noise’ – that is, a pattern of meaningless pixels – from images. When a user inputs a natural-language ‘prompt’ into the trained model’s interface, it is converted into ‘tokens’ that guide the model as it gradually removes noise from its dataset, thus resulting in a new image that hopefully matches the prompt. As Dave Lynch helpfully explained it to team members in the context of the AIAI project, AI image generation with a diffusion model is ‘a process where a computer model diffuses visual images akin to how an ink drop disperses in water. By doing this millions of times with different images, the model is then able to generate images by reversing the process, reforming these “dispersed drops” into new images via instructions in the form of text prompts’.

The ease of use via natural-language prompts, as well as the quality of the results, have accounted for the massive uptake of diffusion models. Many models, such as Midjourney and DALL-E, are also integrated into other widely used platforms (Midjourney has always been accessed via the community platform Discord, while DALL-E 3 has been accessible via ChatGPT); however, use is restricted by those that are commercial tools requiring a subscription, such as Midjourney, versus those that are free to use, such as Stable Diffusion. Our understanding of the way they work is also governed by the degree to which they are proprietary or open source in terms of the datasets they are trained on, the algorithms used and the option to customise them by training smaller additional models (Low-Rank Adaptation models, or LoRA) as side-weights to the main algorithm. Stable Diffusion is fully open source, whereas the dataset that Midjourney is trained on is reported to be the same as Stable Diffusion’s dataset (LAION-5B [Large-scale Artificial Intelligence Open Network]) and thus open source, but the model and its algorithms are not, so it functions as ‘a black box’ where we can only make educated guesses from the results in relation to the prompts used about the algorithmic process of image generation. It is, of course, the case that those image-generation models that offer the most user-friendly interfaces are often the most opaque in their functioning, and those that offer greater transparency have a higher bar of technological expertise in terms of users being able to engage with them.

Before moving on from this technical summary of the tools in question, what is worth noting about the terminology relating to image generation, and to AI more broadly, is, of course, the extended metaphors relating these models to the human brain, seen most clearly in the terms ‘neural network’ and the reference to ‘intelligence’ in AI. Related to this are also all the references to ‘dreaming’ and ‘hallucinations’ that abound in this field. DeepDream itself is often credited with having contributed to the ‘widespread tendency to consider images generated by deep-learning algorithms as a form of “dream” or “hallucination” of the “machine”’ (Somaini 2023, 91; see also Zylinska 2024, 246). The term ‘hallucination’ has also come to refer specifically to cases where the ‘generated images fail to reflect factual information’ (Lim, Choi and Shim 2025) that may or may not be made explicit in the prompt (such as, in the early days of diffusion models, representing hands with extra fingers). While this programmed response to always generate an image regardless of its facticity may present a problem for those seeking to accurately represent a minoritised culture, for example, it can also provoke creative interactions leading to the generation of increasingly curious results from an artistic perspective. The copious references to dreaming in this field are also very suggestive in relation to Indigenous epistemologies where dreaming is a key way of interpreting the world and imagining the future. It is for this reason that we have framed our response to AI image generation as sitting ‘between dreams and hallucinations’.

Shifting our focus now to the relationship of generative AI to art and Indigeneity, in relation to art, while many arguments have been advanced that maintain that AI-generated visual material cannot constitute ‘art’ per se, if we accept that the term is often used to refer to visual culture more generally, then ‘AI art’ exists, and it is frequently artists who engage with such tools who are seen to be ideally positioned to ‘stress-test’ them with respect to their impact on society. In relation to Indigenous Peoples and their visual representation and cultural self-expression, including art, however defined, while such technologies may tend to invisibilise or misrepresent minoritised cultures and/or risk facilitating even greater levels of cultural appropriation than before, theorisations of Indigenous, and more broadly decolonial, AI recognise that engaging with, appropriating and redesigning these new technologies is necessary and that Indigenous epistemologies may helpfully inform their ongoing development.

However, in such an emergent field, what is required, in addition to theoretical debates, is evidence of how Indigenous people, and Indigenous artists in particular, are responding to generative AI, and their assessment of its use value and/or its status as art. While in Anglophone contexts there is already evidence of Indigenous artists ‘stress-testing’ these generative AI technologies, this book provides evidence from an early pilot project to explore precisely this in a Latin American context. Working as a team with Indigenous³ artists, writers, traditional knowledge holders, community leaders and others, based, in the vast majority of cases, in present-day Argentina, Bolivia, Brazil, Chile and Peru, the AIAI: Artificial Intelligence, Art and Indigeneity project, launched in 2023, has sought to answer these questions through workshops to discuss images generated with a range of different tools, including Midjourney and a bespoke Stable Diffusion-based image-generation tool called IndigenIA. In particular, the project has explored the team members’ struggles with AI hallucinations that misrepresent their realities and intentions, as well as their pragmatic engagements and their hopes and dreams in relation to generative AI, all with a view to foregrounding Indigenous perspectives in shaping the way this kind of technology continues to develop and impacts us all.

In what follows, we offer a series of reflections on what we learnt through the AIAI project across its different stages, and this, in a multivocal format. We have sought to keep our different voices and perspectives distinct, organised into eighteen short ‘chapters’, rather than to bring them together into a more uniform argument written in standardised academic prose – in these days of LLM-fuelled syntheses and summaries, we feel that maintaining the uniqueness of people’s different voices and perspectives is of paramount importance.⁴ Non-native speakers of English have written or recorded their individual chapters in Spanish or Portuguese, based on original discussions that took place with the academic and NGO team members over WhatsApp or Zoom. These have then been translated into English and lightly edited, without diminishing the uniqueness of voice and perspective.⁵ All of the chapters by Indigenous artists and writers in Part 2 (Between Dreams and Hallucinations) are presented bilingually with the English translation following the original Spanish or Portuguese.⁶ All other chapters that are only presented in English here are also available in Spanish and Portuguese translation on the open-access Manifold platform that accompanies the book to facilitate accessibility for all interested parties (read.uolpress.co.uk/projects/artificial-intelligence-art-and-indigeneity).⁷ The platform also includes additional materials, such as the short videos made about the project, that would be impossible to include in the printed book.

The book is divided into three sections. Part I (AI, Art and Indigeneity) provides the theoretical framework, context and methodology underpinning the AIAI project and offers an overview of key debates in the field of (generative) AI, art and visual culture; Indigenous and decolonial AI; and of Indigenous artists’ experiments with (generative) AI both in the Anglophone settler (post-)colonial world and in Abya Yala/Latin America. It then moves on to give an overview of the AIAI project as it has evolved over time and a brief discussion of the methodology used. Part II (Between Dreams and Hallucinations) comprises texts written by some of the Indigenous artists and writers who collaborated most extensively on the AIAI project over time, reflecting on their different responses to image-generation tools. It is ordered in such a way as to reveal a shift from the generally more cautious responses of visual artists to these tools in comparison with the more positive response of writers who have gone on to use them to illustrate their written work. Part III (Generative Thoughts), authored by both the book’s editors and a range of other team members, both Indigenous and non-Indigenous, explores the way that the results of the AIAI project intersect with some of the wider debates in the field relating to topics such as racial stereotyping, Indigenous futurism, censorship and copyright infringement, as well as the different perspectives that various collaborators have brought to the project and the different directions in which they have taken it as they have developed spin-off projects. We also reflect on the different considerations that have been brought to bear as we have exhibited the project in a variety of locations and the public responses to it. The conclusion briefly draws together the findings of the project, setting them in a wider context.