The Automatic Image and its Prompt

I am not an artist, not in any conventional sense, but in writing about art I’ve become interested in how prompt-based image generators challenge our understanding of what it means to make art. Who claims the title of artist when the images produced require both human and algorithm? How does the creation of unique computational images challenge ideas around originality and conceptual expression, hallmarks of modern and contemporary art? Is the creative act in the making of an image, its reception, or its production? It is a deeply anthropocentric idea to assume that art is an exclusively human endeavour, and as scholarship expands to redefine intelligence—by better understanding the more-than-human world around us—it seems necessary to rethink art as an activity that necessitates technological collaboration.

AI image generators, of which DALL-E is the most famous, began proliferating in early 2021, and the resulting images took over social media feeds. Like its famous namesake, Spanish surrealist painter Salvador Dali, DALL-E mashes up disparate visual elements to create combinations of mostly surreal, odd, fantastical, or funny juxtapositions from any possible linguistic prompt. Because the prompts can include things that are “in the style of,” non-artists can produce images that take on attributes of famous artists, time periods, or styles. The question of whether this is truly acceptable as art, however, is an ongoing debate.

Feature image: AI-generated image using the prompt “an image generated by artificial intelligence.” Produced on Oct. 10, 2022, using Craiyon, formerly DALL-E mini, an AI model that draws images from text-based prompts. Courtesy of Jayne Wilkinson.
Image description: An abstract image with a yellow and blue ‘S’-like shape on top of a green background. Extending from the background are yellow, red, and blue striations; pixelations that appear as if a woven texture.

This image set: AI-generated images using the prompts “an image of nothing in colour.” Produced on Oct. 10, 2022, using Craiyon, formerly DALL-E mini, an AI model that draws images from text-based prompts. Courtesy of Jayne Wilkinson.
Image description: A set of abstract images made up of amorphous blobs and streaks of bright yellow, pink, green, and blue.

When I tried Craiyon, a freely available version of DALL-E, most of what I produced was boring, conventional, or too personally limited by idiosyncratic curiosities to be of much interest to others. These images were no more than anecdotal one-liners. I tried some experiments related to aesthetics—for example, what is the most beautiful or the most ugly of something, what is the most confusing—to learn something about the biases built into the system; the results were unfortunately predictable, demonstrating that the output is only as good as the input. As I scrolled around looking at accounts, the humour and absurdity of the images as visual jokes, and their timeliness to news cycles (on Weird Dall-E Mini Generations you can see things like Mar-A-Lago FBI raid Lego set, Yoda attends Queen Elizabeth II’s funeral, or lofi nuclear war to relax and study to) felt little different from meme culture, where politics can be signified through humour. At most, they offer a passing comment but with little critique of, in this case, Trump’s stranglehold on American democracy, the overhyped celebration of monarchy and colonial violence, or Putin’s maniacal threats of nuclear war. DALL-E can, it seems, create anything, and the desire for celebrity culture, absurd nostalgia, and fantasy to sit together on the same screens where we live and work has obvious appeal.

The broad possibilities for access and dissemination are also what make consumer/social media AI technologies interesting for artists. Anyone can make a painting or take a photograph, but not everyone can make a good one; the same applies here. Using consumer AI as a medium requires conceptual skill, not least in designing the prompt or writing the input sentence but also in modifying and refining a project such that the output is a useful part of one’s formal aesthetic. In this sense, DALL-E and other AI tools are aligned with histories of Conceptual art, where artists were concerned with foregrounding image-text relationships through techniques of automation, instruction, or the use of rules to produce art. Even earlier, in the nineteenth century, experiments with photography (mostly by scientists at first) were similarly concerned with the relationship between the caption and what was contained within the frame. Image and text mutually informed one another, as details of what was in the photograph could be explained through text, making the mimetic representation of reality complete. Photography automated perspective and enjoyed widespread fascination at the same time that western painting was becoming abstract, moving away from the necessities of representation. During the early decades of the twentieth century, Dadaism (and then Surrealism) became globally dominant aesthetic movements, and both were also influenced by industrial automation and new technologies to access ideas about human consciousness and subjectivity. Automatic writing, not unlike AI, was a way to generate absurdist ideas, images, and concepts that would not otherwise exist together, tapping into new ideas about human psychology, consciousness, and subjectivity.

What are the frameworks we have now for making automated images? What criteria should we use to determine when a work has conceptual meaning as an artwork, and what distinguishes art from meme culture, beyond its location? And at what point might we agree, without debate, that technologically mediated ideas are art? It took decades for photography to be considered a fine art alongside painting and sculpture, and almost a century before the market for contemporary photo-based practices really developed. For contemporary artists using NFTs and other digital-first mediums, the runway from experiment to market has been much faster. What is interesting to me is not so much a debate around “is this art” or “who is the artist,” but understanding how we are going to analyze AI works critically—what frameworks account for work that is conceptually strong, beyond experimentation. Should we (as critics) want AI-generated artworks to be accompanied by their prompts, like work details or didactic labels? Or does the disclosure of the prompt destroy some of the mystery of how—and from what—the work is produced? Are the writing prompts and AI-training techniques more important than the image or work itself?

Images: AI-generated images using the prompt “an image generated by artificial intelligence.” Produced on Oct. 10, 2022, using Craiyon, formerly DALL-E mini, an AI model that draws images from text-based prompts. Courtesy of Jayne Wilkinson.
Image description: A set of abstract images. In each image, yellow, pink, green, blue, and brown streaks extend from a central point. Striations and pixelations extend from the centre, generating an almost woven-like texture.

The power of AI has grown exponentially over the past decade, and the question of its ethical deployment has been taken up across many fields, including aesthetics. Concerns around AI reproducing the most violent assumptions and stereotypes of difference—racism, homophobia, implicit bias—are also about questioning who is in power, who funds and builds systems of intelligence, and who normalizes how the algorithms adapt.

The parameters for machine-based learning and automated images are still built by humans, and therefore already included in the same systems of inequity that structure the “real” world. For technology critic and art writer Nora N. Khan, the intersection of artificial intelligence with aesthetics does not hinge on the question of authorship—it’s about the ethics of representation that are reproduced when the (human) responsibility for creation is dispersed. Khan articulates how scholars and activists may demand better legal frameworks and smarter engineering “to counter and correct the unethical violence of these specific algorithmic systems. But the everyday person, who is forming the material for these systems’ training, has a different set of tools for understanding.”¹ The material Khan refers to is data, the image-sets and text prompts that require human input and continual refinement. As AI becomes “smarter” and is utilized in more and more fields of study and practical applications, its growth will shape how we understand ourselves as a species—the types of existential questions that have long been of concern for artists. Khan writes: “How we see or unsee is the primary ethical question in a culture and computational regime that privileges vision. And how we see, name, and know the world is increasingly influenced and shaped by how machines see, name, and know; machines read images and then produce a matrix of knowledge that deeply shapes how humans read images on the same platforms.”² Image-text relationships form the crux of conceptual art practices, and experiments with the relationships between language, meaning, identity, ideology, and vision remain at the fore of contemporary art discourses. As AI practices become more refined and the automation of imagery more integrated, it will be the more creative or meaningful prompts that produce the most unusual or unexpected images.

Using consumer AI as a medium requires conceptual skill, not least in designing the prompt or writing the input sentence but also in modifying and refining a project such that the output is a useful part of one’s formal aesthetic. In this sense, DALL-E and other AI tools are aligned with histories of Conceptual art, where artists were concerned with foregrounding image-text relationships through techniques of automation, instruction, or the use of rules to produce art.

Tools of AI are difficult to distinguish, however, from an industry that is ecologically destructive and embedded with biases. Art critic Ben Davis argues that the capacity for artificial intelligence in art has already been deeply shaped by the tech sector’s relationship to capitalism, and he critiques how embedded tech is in all aspects of contemporary life: “[…] the impact of AI on artists may be the least of all worries about this technology. But the anxieties it raises can’t be seen apart from those other, larger worries either, of technology’s fusion with unaccountable power and amoral technocracy.”³ He further argues that, “Aesthetic experience is a mixture of formal invention and social meaning, and the better AI gets at automating visual interest and narrative novelty, the more directly it will force into relief the question of what is meaningful in our aesthetic worlds. This includes the question of why our cultural energy is invested in technology that locks people ever more inside their own customized taste bubbles at a historical moment when we need to be actively working toward a collective vision.”⁴

The investment of cultural energy into as yet ethically and morally untested forms is indeed risky. The control of computing technologies by a small number of extremely powerful companies, along with the centralizing of tech wealth in the assets of multi-billionaires, is one of the most significant challenges to addressing AI as a useful tool for aesthetics. But there are also ways to counter—and to ethically challenge—the feedback loops of AI that could offer powerful avenues for new forms of art and aesthetic critique. The Indigenous Protocol and Artificial Intelligence Working Group is working to address these questions from a specifically Indigenous perspective, by querying how Indigenous epistemologies can contribute to global conversations on AI, how discussions about the role of tech can move beyond the limitations of Silicon Valley, and most significantly, how we might imagine a future with AI that contributes to the flourishing of all humans and non-humans alike. There is an optimism in this scholarship, because it is collective but not corporate, and because it recognizes that there are futures for technology that need not be driven by capitalism. Now is a critical moment for that work to happen.

As each new cycle of technology automates an aspect of creativity, there is an attendant concern, or even panic, that some essential human ability is being erased. Part of that concern, no doubt, is a lingering holdover of Enlightenment/colonial-era discourse that defined the human subject as male, white, wealthy, and heteronormative, creating the myth of individual artistic genius in its wake. Fears about technology erasing human vision were prevalent in the early days of photography as well, but art has always had a relationship to technology, even if the technology was comparatively simple; making and manipulating paint on a surface or hammering a rock into a form are still practices that require the organization of knowledge to produce a visual discourse. They require not just the use of tools, but changing technologies of vision. The difference is that technologies of seeing and imaging are rapidly advancing, and the possibilities for critically engaging with AI—as image-making tool, as material, as software, as discourse—are radically open. How then are such possibilities and critiques being resolved by artists? A second part to this essay will consider specific artists’ projects that not only use AI as a tool, but are invested in rethinking and reshaping the role of automation and technology in art.

Jayne Wilkinson is a writer, editor, and curator based in Toronto.

1. Nora N. Khan, Seeing, Naming, Knowing (Brooklyn, NY: The Brooklyn Rail, 2019), 8.
2. Ibid., 9.
3. Ben Davis, “AI Aesthetics and Capitalism,” in Art in the After-Culture: Capitalist Crisis & Cultural Strategy,(Chicago, IL: Haymarket Books, 2022), 112.
4. Ibid.