Thea Pitman

Given that generative AI is such a recent phenomenon, the methodology for the AIAI project has stemmed predominantly from our own past experiences of working together and of taking a decolonial approach to arts- based participatory action research (PAR).¹ Nonetheless, our approach also aligns well with more recent methodological proposals for working with (generative) AI. In Rita Raley and Jennifer Rhee’s (2023) ‘Critical AI: A Field in Formation’, the authors note that examining questions of racial and gender bias is of paramount importance in critical approaches to AI, that ethnographic methods are ideal for exploring this and that, in particular, artists are ‘particularly well positioned to stress test, evaluate, and exploit [AI tools], to probe and reveal their limitations so as to communicate these to the public and even advocate for better – which is to say, fair, transparent, and accountable – data sets, models, and applications’ (191). This is, essentially, what the AIAI project seeks to do and why.

Anne Dippel and Andreas Sudmann (2023) focus in particular on the value of what they call ‘AI ethnography’, drawing on ethnography’s focus on ‘situated knowledge production’ and ‘processuality’, and its spirit of ‘learning by doing’ and co-creative practice, including with more-than-human actors (827), for allowing us to study the way that different social groups relate to these new technologies. They also recognise the more recent impact of ‘feminist, queer, and decolonial approaches’ (838) on ethnography, and it is here that our practice is situated. For the first iteration of the AIAI project, Indigenous people were actively recruited from their wider networks by the core research team (subsequent to institutional ethical review), and interview- and focus group data were gathered in relation to the process of generating images and analysis of individual images themselves. However, in subsequent iterations of the project, the team has expanded to include everyone involved as named collaborators, and from the grant-application stage onwards, people have worked with greatly increased levels of autonomy and self-direction, and have fed into the direction of the project going forward. The nature of this book as an ‘anthology’ of different perspectives on the project is proof of this change in emphasis.

Nonetheless, we recognise that our methodology has its limitations. As Fabian Offert and Ranjodh Singh Bhaliwal (2024) argue, this kind of approach based on very small-scale experiments is part of a tendency in digital humanities approaches to AI that produces results that are little more than ‘anecdote’ based on an often insufficient grasp of the technologies (3) and where ‘we are performing the techno-critical equivalent of judging a book by its cover’ (8). They also add that it is also not really the job of digital humanities scholars to do ‘humanities red teaming’ (2) (that is, systems-testing to iron out bugs, security lapses and other deficiencies) – this is the job of computer scientists and engineers employed by the companies developing these tools and we may not want to serve the system or ‘feed the machine’ in this way. Offert and Singh Bhaliwal conclude by arguing that those undertaking Critical AI studies should try to understand the technology better and/or collaborate more with those who do understand it, and they should take an approach that would produce a more comprehensive ‘close reading of [these] complex systems’, instead of recounting flimsy anecdotes (8).

While we take this critique on board, for the purposes of the AIAI project, the focus is precisely on the stories that people tell – their personal responses when faced with a new, far-from-fully understood technology. In this we align ourselves with Ruha Benjamin’s (2019) deployment in relation to the study of racial bias and emerging technologies of what John L. Jackson (2013) has called ‘thin description’ in ethnographic research. The practice of thin rather than thick (all-knowing, totalising) description, Benjamin explains, ‘push[es] back against the all-knowing, extractive, monopolizing practices’ that we find in the field of technological development – practices that ‘seek to penetrate all areas of life, extracting data, producing hierarchies and predicting futures’ (46). We thus embrace our decision to focus on a small-scale, arguably unscalable, situated approach and on a non-totalising approach to that data.

In terms of more detailed methods for generating and analysing AI-produced visual materials, as well as reflection on the choice to use proprietary or open-source tools and consideration of the advantages and challenges of designing bespoke tools, our approach sits well with Sabine Niederer and Gabriele Colombo’s discussion of this topic in their recent Visual Methods for Digital Research (2024). We have experimented with ambiguous, evocative, provocative and abstract forms of what they call ‘prompt design’ (that is, designed for research purposes rather than ‘prompt engineering’ to optimise results when accomplishing real-world tasks) (122–35) in order to reveal some of the biases and deficiencies (ambiguous prompting and counter-prompting), and in-built censorship (provocative prompting) of the tools used.² We have also sought to generate images that reveal a tool’s linguistic limitations (evocative prompting with ‘nonsense words’ or simply words in Indigenous languages that the tool does not recognise) or that can serve as ‘visual prompts for (human) speculation or reflection’ (135) (abstract prompting), for example, by using the tools to help dream imagined futures.³ Eryk Salvaggio’s work on how to understand the visual output of AI image-generation tools (2023a), and how to design experiments to artistically exploit these tools – this, pace Offert and Singh Bhaliwal, he and colleagues refer to as ‘red teaming’ (Salvaggio, Sinders and Swanson 2025) – has also been influential in our subsequent processing of data generated through the AIAI project.