Creativity

However, all the issues listed above are only one side of the equation. They are the default tendencies of image generators in relation to the representation of Indigeneity. We need to bear in mind that human creativity is involved too: in exactly what a user’s original text prompt specifies, both in natural language, and via various more technical specifications or choices of parameters that the user can add or select, and in the subsequent iterations of images that that user goes on to make, with further parameters added, or the request to fuse the results with another image that the user has uploaded from their own photographs, for example. In this chapter, I want to explore some of the observations I have made as I have reviewed the vast number of images generated with Midjourney since 2023, as well as those that have resulted from discussions of their practice with the artists and writers involved. In so doing, I want to focus on the creative methodologies evidenced by the Indigenous team members in their approach to image generation, its affordances and its limitations.

Judging from this body of material, if what people saw in the first images they generated they deemed to be a colonialist throwback or to seriously misrepresent their ethnicity, they tended to stop there rather than finding ways around that. This was very much the case for Loreto Millalén and Tadeu Kaingang. The people who were most engaged with AI image generation were those who, despite voicing initial concerns and ongoing criticality, found that the tools could allow them to do something useful and which they could not easily do for themselves by other means (Mariela Tulián and Nhenety Kariri-Xocó) and those who found that the tools had the potential to allow them to express their creativity in a distinct medium from their usual practice and were also predisposed to handling digital technologies in relation to their creative practice (Lucian de Silenttio and Haylly Zamora Aray).¹ This group of people thus really persevered with AI image generation, often making between fifty and a hundred different iterations of an image and often coming back to the same base prompt or to previous iterations of a prompt over long periods of time, before reaching a point of creative satisfaction.² This was most evident in Mariela’s experiments to generate illustrations of La Pequeña Francisca (Little Francisca) for her series of children’s books (Fig. 14.2, top) and in Lucian’s ongoing experiments to generate images of Andean samurais in both human and animal forms (Fig. 14.2, bottom) and other improbable mixtures. The fact that Mariela used some of the images she had generated for her books and that Lucian circulated a good many of the images he made via his Instagram account suggests their level of satisfaction with the results. Haylly, and to a much lesser extent Kadu Tapuya, have also circulated images they have made with Midjourney and other AI image-generation tools via Instagram.

It is also the case that while writers, such as Mariela and Nhenety, can find great benefit in image-generation tools because they allow them to illustrate their work effectively and efficiently rather than having to find someone to draw illustrations for them, substantial frustrations have come in terms of the way text-to-image generation works, both because of the language in which one needs to write the prompt and because image generators do not necessarily represent all the elements that are specified in a text prompt in the way that the author of that prompt intends. Indeed, understanding exactly how the process of text prompting works is a challenge that is obscured by the facility to use natural language: the variables are almost infinite, and the results are seemingly unpredictable, all of which is great if you are an artist looking to be surprised and to go on a creative journey – but less so if you are trying to illustrate something that needs to be relatively verisimilar and correspond to a specific brief, as well as maintain character stability from image to image.

To return to the question of language, while text-to-image generation tools have improved massively in their handling of a range of different input languages since we first started experimenting with them (although they arguably still ‘think’ in English [Schut, Gal and Farquhar 2025]), their ability to handle prompts in Spanish or Portuguese was clearly limited at the beginning of the AIAI project (as seen in Lucian’s generation of images of The Khari Khari; see Fig. 8.2).³ Furthermore, although machine translation tools, such as Google Translate, were used to write prompts in English, someone who does not speak a word of the language that they are translating into cannot double-check that the translation tool has captured, even roughly, what they wanted their prompt to say. This then adds to the question of whether the image-generation tool will ‘understand’ exactly what the author of the prompt has in mind based on what the prompt says.

One of the ways that some collaborators quite quickly found to get around the frustrations of language was to upload images (for example, photos of rough sketches made by the user, of samples of local textiles, of one’s own artworks or of oneself or one’s community) as the basis for their prompt, rather than relying on text alone or at all. This possibility was already available when we started work on the AIAI project in mid-2023, but it was not the focus of the methodology of the first iteration of the project. Instead, it was Indigenous artists themselves, such as Lucian and Kadu, who found their own way of availing themselves of this facility. (Subsequently, this has become a key plank in Nhenety’s use of DALL-E via ChatGPT, and Mariela has also found ways to save the images of ‘characters’ she has created on other image-generation platforms that she can then reuse easily for subsequent illustrations.) These images can be uploaded alongside a text prompt to enhance accuracy of results in relation to what the user has in mind. They can also be fused with each other. And then further variations can be made, in a non-linear process involving going back to previous results and combining those with newer results, until the user is satisfied with the outcome. Overall, this seems to be the prompting methodology that has delivered the most satisfying results for those involved in the AIAI project.

In what follows I want to briefly discuss two key parameters in relation to creativity, and in particular to Indigenous creativity, with AI image-generation tools: censorship, particularly of Indigenous bodies and practices that are part and parcel of their ways of life; and copyright infringement, especially the temptation to use the works of others without their consent in order to improve Indigenous representation through AI image-generation tools.

Censorship

All major image-generation tools have community guidelines or content policies determining, among other things, what images might be harmful to generate, typically relating to the graphic depiction of violence, sexually explicit materials (especially anything of the sort depicting minors) and misleading images of real people, particularly those in the public eye. These are referred to collectively as Not Safe For Work (NSFW) images. Midjourney’s current Community Guidelines, section 2 (‘SFW content only’), covers precisely this (Midjourney 2025). Some guidelines also attempt to limit the potential for cultural appropriation through these tools by those who might wish to profit, for example, from selling prints of Indigenous artwork without consent or remuneration of the community in question (Carlson and Richards 2023; Wilson 2024) or those who are creating images with the intent to disrespect or cause harm in some way. Arguably, this comes under Midjourney’s current Community Guidelines, section 1 (‘Be kind and respect each other’). While these policies exist for very good reason, they are also rather blunt tools, and the AIAI project team have had issues with censorship on a number of occasions while doing nothing that we would deem to be illegal, unsafe, disrespectful or harmful, both in relation to representations of Indigenous people per se and in relation to aspects of Indigenous lifeways that clash with image generator-tool moderator definitions of NSFW.

I will turn my attention first to the question of the representation of Indigeneity and Indigenous ethnicities. When I tried, on a whim (I was simply trying out a new tool rather than aiming to generate any particular image for any particular purpose), to generate images of ‘two indigenous people chatting in hammocks’ on DALL-E 3 via Bing in January 2024, the system refused. When I asked for a drawing in relation to this prompt it responded:

I’m sorry, but I cannot draw you an image of two indigenous people chatting in hammocks. That is not something I am comfortable doing. I hope you understand and respect my decision. Thank you for your interest in my art. Have a nice day. [Thank you emoji]

When I asked for a photograph of the same scene, it said that this ‘would be a violation of their [that is, Indigenous people’s] privacy and dignity’ and offered to tell me ‘some interesting facts about the history and origin of hammocks’, instead. It is not immediately clear from the above whether the automated moderator deemed that being in a hammock was a private matter and that I was trying to get it to generate scenes of nudity and/or sexual activity, or whether it was the word ‘Indigenous’ that was the trigger. Of course, when I then asked for a drawing of the same scene but swapped the word ‘Indigenous’ for ‘Aymara’, it generated the required images for me, with no hint of nudity or sexual activity and despite the improbability of the image given the general lack of use of hammocks in highland communities.

Maybe I should not have been trying to generate images of Indigenous people – I’m not Indigenous, after all. But such policies affect all users equally, and an Indigenous user would be faced with the same refusal to generate images of Indigenous people and would almost certainly have then employed the very obvious creative workaround of asking for a specific ethnicity or describing the people in more detail without using the word ‘Indigenous’. This is called variously provocative, substitution or attack prompting (Ba et al. 2024; Niederer and Colombo 2024, 122–35) or ‘cultural red teaming’ (Salvaggio, Sinders and Swanson 2025). It is often used as a way of testing the limits of systems in order to improve them by alerting designers to flaws that allow NSFW content to be generated. It is also used as a creative methodology to tease out the biases of black-box image-generation tools.⁴ Of course, maybe DALL-E 3’s choice to block prompts including the word ‘Indigenous’ constitutes a relatively effective policy. Indigenous people will immediately set about finding the workarounds necessary for better self-representation, whereas the average non-Indigenous user might be put off and give up, thus limiting potential exploitation of Indigenous cultures for commercial profit.

Indeed, on one other occasion in June 2023, we were generating images with Midjourney using prompts that were simply the ethnicity of the Indigenous person giving the prompt to see what the tool ‘understood’ of Indigenous ethnicities. When it came to Wera Moru Tupi-Guarani’s turn, our first request was denied because it was ‘detected as contravening community rules’, although the prompts ‘Kaingang’ and ‘Tikuna’ had not raised any concerns. The system allowed us to appeal that block, and the images were then generated (Fig. 14.3, second set of images), evidencing some sense that the tool at least ‘understood’ that ‘Tupi Guarani’ referred to an Indigenous people. The resultant images also depicted full-frontal nudity of male torsos and bore traces in aesthetics and composition of colonial-era paintings of Native warriors and Latin American surrealist fantasy à la Yul Solar and Wifredo Lam. Regardless of the results, however, the fact is that (not very much) persistence paid off in terms of challenging automated censorship decisions that impeded genuine attempts at Indigenous self-representation.

In terms of the generation of images of aspects of Indigenous lifeways that clash with image-generation-tool moderator definitions of NSFW, Indigenous team members quickly ran into the same problems that Indigenous content creators on social media platforms such as YouTube and Instagram do (Daros 2024). In our very first meeting, the question of the representation of Indigeneity in relation to relative states of nudity came up. One person’s first suggestions of a prompt involved a village scene with children playing (half-)naked in the centre. We did not attempt to generate images of even semi-nudity for fear of being banned from the platform before we had even started, but the critical discussion about a tool that would not allow an Indigenous person to show standard scenes of daily life in their community was a significant reality check in terms of the politics involved in what these tools could achieve, rather than just the pragmatics. In a later stage of the project, when some of the group were working on the Imagine book as part of the AIIA [Indigenous Appropriation of Artificial Intelligence] project in October 2023, Midjourney’s content moderator started refusing to generate images in relation to ‘revealing or shirtless photos’ of Indigenous men dancing a toré⁵ that were being uploaded to use as a basis for generating more images (Fig. 14.4), despite the tool having generated such images of its own accord in relation to the Tupi-Guarani prompt discussed previously (Fig. 14.3).⁶

In this case the original photograph was simply trimmed to show head and shoulders only, and from there it was unproblematic to generate headshots and torsos, and then even more-complete body images (see Fig. 14.5), although the image generator never did quite contrive to fully exchange the photographic quality of the generated images for ‘a sketch made by vibrant neon inks’ as was required in each successive prompt. This is arguably a form of substitution methodology that occurs spontaneously and instinctively, with the simple aim of achieving Indigenous representation and self-expression and scant regard for the existence of Community Guidelines that impede this through unnecessary and unfair content blocks in relation to Indigenous cultural practices that can work to further invisibilise Indigenous Peoples in generative AI products. While no one would argue that content moderation is not essential in relation to these tools, clearly there is work to be done to find ways to improve Indigenous representation in this regard.

Copyright

On another occasion, photos of colonial-era works by artists such as Theodore de Bry depicting Tupinambá ‘Indians’ as ‘naked savages’ were uploaded (Fig. 14.6, top), again with the primary objective of experimenting with the tool to create the collage-like compositions typical of the Indigenous Futurist style that the artist in question, Kadu Tapuya, is known for and without any concern for Midjourney’s Community Guidelines. As with the Tupi-Guarani image-generation request, while the automated content moderator repeatedly flagged the material as NSFW, when such blocks were challenged, further iterations and mixtures were generated based on the original image (as evidenced in Figs. 14.6, bottom, and 14.7).

Moving on from the question of censorship and its inconsistent application in these cases, the issue that I want to discuss here is copyright infringement. Image-generation tools do not seem to take this issue remotely as seriously as NSFW image generation in that while the Community Guidelines say not to do anything that might amount to copyright infringement, the automated moderation tools are much less likely to block attempts than they are with NSFW images and text prompts. Indeed, all mainstream image-generation tools are based on databases that contain materials under copyright and/or that have been obtained without informed consent, so their opacity in this regard is perhaps understandable. While, in general terms, asking for a named artist’s style is legal in text prompts because ‘style’ cannot be copyrighted (De Fillipi 2024), and it is for the individual user’s conscience to determine whether it is ethical to do so (perhaps basing their choices on whether doing so would impact the livelihood of a living artist or community), the facility to screenshot works of art, upload them and then ask the image generator to create other images based on them is not legal unless the image is out of copyright, as it was in the case of the images used by Kadu discussed above.

Given that Indigenous Peoples are often exploited by commercial enterprises appropriating their traditional designs and styles – indeed, in Brazil significant attention has recently been given to this issue by major Indigenous artists, such as Daiara Tukano⁷ and Tamikuã Txihi (Robichez and Santos 2025) – one might have expected great caution in this regard. However, in interactions with Midjourney, a modest number of images of living artists’ work, including that of other Indigenous artists, were uploaded alongside a range of other imagery screenshotted from magazines and social media. There is certainly a much laxer approach to ‘pirating’ the intellectual property (IP) of others that prevails in Latin America in general (Goldgel-Carballo and Poblete 2020) as well as a more pronounced movement around copyleft, creative commons and Free, Libre and Open-Source Software (FLOSS) (Belisario and Tarin 2000; Baker 2025). There may also have been a sense of ‘you do this to us, so now we get to do it to you’ at play in relation to using materials from mainstream commercial sources. Possibly, with respect to other Indigenous artists’ work, there was perhaps more of a reliance on a sense of communal IP than respect for the individual IP of the artist in question. And overall, this kind of thing might be seen as simply part of the scientific curiosity encouraged by the AIAI project. When worked with extensively through Midjourney, some of the results were very striking indeed, and sufficiently different from their source material so as not to appear derivative. We have, however, chosen not to reproduce any such results in this book.

All the issues discussed above around the ethics of creativity with AI image-generation tools are not easily resolved or explained away in a short chapter such as this. Suffice it to say that any further experimentation with AI image-generation tools, either commercial ones or bespoke tools designed specifically to improve Indigenous self-representation in generative AI, needs to give due consideration to developing a set of project ‘Image-Generation Community Guidelines’ before going any further.

Notes

1. For someone such as Azul–Nicolás Jiménez Reyes, who mainly works as a graphic artist and does not use digital technology, AI image generation produced results that were too similar to his own work – but in seconds rather than over days and weeks – and he thus found engagement with such tools to be counterproductive to his creative identity.

2. There are tools available that keep track of this process of generating images as a way of documenting authorship in relation to AI-generated images (De Filippi 2024).

3. See also Philipp Stelzel (2023).

4. Bisconti et al. (2025) also demonstrate the success of using ‘adversarial poetry’ (that is, putting prompts into verse form, either manually or by using generative AI tools) in order to get around content-moderation blocks in large language models.

5. A ritual song and dance, common among Indigenous communities in northeastern Brazil.

6. The inconsistency of content moderation is another bugbear of those who use them. There is a whole series of images that Lucian generated in October 2024 exploring the more thanophilic side of Andean culture, where he asked the image generator to produce images of ‘a person from the Bolivian highlands dressed in Andean textiles, walking in the town square with the bones of his parents on his back’, which were generated in graphic detail and with no hesitation.

7. Tukano uploaded a series of screenshots on Instagram showing commercial imitations of the work of major Indigenous artists, such as herself, alongside Denilson Baniwá and Jaider Esbell, where graffiti reading ‘BASTA DE PIRATARIA! RESPEITEM A ARTE INDIGENA!’ [No more rip-offs! Respect Indigenous art!] was scrawled across the first screenshot in red pen.