Paper collage created from public domain images of dwarf emu skeletons and natural history illustrations. Source materials are shown on the black panel to the right; other images and ephemera came from her ‘archive’ of collage material. Zoë Sadokierski 2023.
Scroll right to view multiple collages.
The collage-making was driven by formal considerations, exploring how cut-and-paste techniques might be used to communicate fragmentation and absence within the historical archive. However, Sadokierski reflected that the ‘hook’ of the story was missing; the endling emus living out their final decades in a Parisian garden, long after their kin had been hunted to extinction, is a complex story to visualise, particularly with such scant visual documentation in the archives.
Around this time, generative models were being touted as able to visualise almost anything. Having never experimented with a large-scale text-to-image model before, Sadokierski wondered how one might handle the visualisation of an extinct species that she was struggling to visualise herself.
Prompting to prod the model
The initial DALL•E 2 experiment involved 17 prompts (see full list to the right). Starting with ‘A King Island emu in a garden in Paris in 1804’, the following prompts were conceived through an iterative process of reflecting on the results of previous prompts to determine what to ask next. The approach was instinctive, and not based on any knowledge of prompting conventions.
Initial prompts asking DALL•E 2 to depict subjects (emu, Josephine, Napoleon) in a setting (garden in Paris) resulted in relatively expected outcomes – emu or ostrich-like creatures in a garden setting, as shown in the slideshow below (click for multiple images):
However, Sadokierski reflected that, for her, these images carried a ‘colonial aesthetic’ – perhaps based on associations that she made upon seeing an Australian bird presented in a European setting, and knowing the bird was there as an ‘exotic souvenir’ of colonisation. This reflection led to writing a prompt she suspected the AI would not be able to visualise, an act of deliberately prodding the model to see what more unexpected images it might produce: ‘Portrait of a King Island emu in a garden in Paris in 1804 with Empress Josephine and Napoleon in Decolonial style’.
These four images present a more visually indeterminate tableau of hybrid figures than the results of the first three prompts. Particularly, the figure in a pink dress in the third image (isolated below) embodies a longer lineage of figures produced by early generative image models that are difficult to visually identify as either human, animal or other, such as the ‘crungus’ and ‘Loab’.1
Sadokierski interpreted this figure as a man wearing a fur (bear?) suit under a ball gown, poking his tongue out at her. This figure felt unsettling in the same way that early CGI characters were classified as ‘uncanny’ due to the discomfort produced by being simultaneously plausibly ‘real’ and frightening.
At this point, Sadokierski sought out literature to help contextualise this practice.
Considering DALL•E Mini, philosopher Tom Whyman writes:
At its best, it trades in a kind of online grotesque. [...] ‘grotesque’ in the original sense, in that they depict bizarre fusions – hybrids like the human-animal-plant figures found in the ruins of Nero’s unfinished palace, the Domus Aurea. (2022)
In particular, Whyman reflects that the model’s inability to accurately create faces may be the key to its appeal:
[T]he faces remain distorted, for the most part a strange, lost blur: the AI not powerful enough to reproduce them properly. This marks out what the tool is giving us as a fantasy: characters appear as they do when we imagine them, not quite complete. (2022)
Although writing about an earlier version of DALL•E, this assessment of the grotesque nature of ambiguous figures remains relevant. The skirt-hoisting ‘Decolonial Goblin’ haunted Sadokierski long after it manifested. Recognising that her hasty experiment lacked thoughtful consideration of the moral or ethical dimensions of the prompts or images produced, Sadokierski approached Monin, an expert in creative machine learning, to help her critique her process and the resulting images.
Collaborative critique
Our initial conversation started with problematising the ‘decolonial’ prompt. Together, we teased out how the ‘colonial’ visual aesthetic attributed to the initial images is not something that can be countered through an aesthetic shift or a different prompt. Decolonisation is not a visual or aesthetic quality but rather an active and ongoing process present in practices that action First Nations self-determination and control of cultural materials (Gothe and De Santolo 2022).2
It should be noted here that decolonisation has been brought into a larger discussion of how to create a more ‘ethical’ AI, a call that has arisen largely in response to what Rachel Adams (2021) calls our many ‘discontents’ with AI, such as its enactment of racial and gender bias, ‘digital colonialism’ in the largely unfettered extraction of community data, and the compounding of existing inequalities. Adams states that, to meet the aims of decolonialism, our conversations about and work with AI need to tackle how racial and colonial logics are part of the very possibility of contemporary AI. For Adams, decolonising AI involves making apparent, critically analysing and undoing ‘logics and politics of race and coloniality that continue to operate in technologies and imaginaries associated with AI’ (190). A starting point for future work here is to consider how we might create alternative approaches, imaginaries and models with AI to think-with.
To return to the images generated in response to the decolonial prompt, one way of critiquing the model is to discuss an absence of training data and through this an incapacity to visualise this scene of the emu otherwise. However, we argue that another form of critique is also necessary, one that involves careful consideration of how images operate within social, cultural and political spheres, and of how visual production through machine learning is a specific shaping force within these spheres. Beyond simply generating images, how are different generative models reconstructing concepts, meaning and social forces?3 This is particularly relevant when working with material from historical archives or generating images about historical events.
A return to collage
Before continuing to the collaborative experiments, which involved thinking together through a shared process of practice-as-research, it is worth mentioning that the final part of Sadokierski’s experiment involved a return to paper collage. Using the images generated by DALL•E 2 as objects to think-with, she then generated a new set of paper collages that mixed archival portraits of Josephine with the emus, in strange ways. She took what she observed from the DALL•E images as inspiration for collages that played with a hybridisation of human and bird forms, to surface the surreal aspects of the story.
2. Working with Absence in Historical Archives
At this point, we return to the endling dwarf emus living out their final decades in Paris. As a way to ‘stay with the trouble’ (Haraway 2016) of the surreal story, Sadokierski began creating digital and paper collages with photographs and illustrations in the public domain gathered from online databases and image libraries. This practice of ‘thinking through making’ is part of a Research Through Design methodology, in which ‘hunch’ drives a practitioner-researcher to begin exploring a topic or problem through practice to surface a clearer research question or aim (see Durrant et. al 2017; Sadokierski 2020).
Experiment 1:
Me versus DALL•E
In the first experiment, titled ‘Me vs DALL•E’, Sadokierski used her collage practice to explore ways to visually communicate the fragmentation of the historical record made up of scattered remains and misinformation, as shown below:
Full list of 17 prompts used in the first experiment:
1. A King Island emu in a garden in Paris in 1804
2. A King Island emu in a garden in Paris in 1804 with Empress Josephine and Napoleon
3. Portrait of a King Island emu in a garden in Paris in 1804 with Empress Josephine and Napoleon
4. Portrait of a King Island emu in a garden in Paris in 1804 with Empress Josephine and Napoleon in Decolonial style
5. Portrait of a lonely King Island emu in a garden in Paris in 1804 with Empress Josephine and Napoleon
6. Oil painting of the extinct King Island emu and Empress Josephine and Napoleon Bonaparte
7. Oil painting of the extinct King Island emu with Empress Josephine and Napoleon Bonaparte in Château de Malmaison in Paris
8. Portrait of Empress Josephine and Napoleon Bonaparte Château de Malmaison in Paris with the extinct King Island emu with [sic]
9. Portrait of Empress Josephine and Napoleon Bonaparte Château de Malmaison in Paris with an emu
10. Oil painting of Empress Josephine with an emu baby in 1804
11. Portrait of Empress Josephine with an emu baby
12. Oil painting of Empress Josephine with an emu as a baby in 1804
13. Photographic portrait of Empress Josephine with an emu as a baby in 1804
14. Photographic portrait of Empress Josephine with an emu in 1804
15. Photographic portrait of Empress Josephine with an emu in Paris 1805
16. Portrait of Empress Josephine dressed as an emu
17. Portrait of Empress Josephine eating an emu
1. See Sparkes 2022 and Batycka 2022 for analysis of the phenomenon of the repeated generation of a few creepy figures, including the crungus and Loab. See Aaron Hertzmann 2020 for discussion of visual indeterminacy in images generated by early Generative Adversarial Networks (GANs). GAN-generated images that are visually indeterminate continue to ‘invite investigation’ for Hertzmann, as they appear at first glance to be ‘coherent and realistic’, yet further engagement never leads to a moment of understanding of ‘the structure of the image’ (424). Hertzmann’s definition of visual indeterminacy draws on Robert Pepperell (2011), who defines visual indeterminacy as ‘a perceptual phenomenon occurring when a viewer is presented with a seemingly meaningful visual stimulus that denies easy or immediate identification’ (1).
2. See the work of the Indigenous Archives Collective for an excellent source of resources on Indigenous knowledge in cultural collections:
https://indigenousarchives.net/
