Authenticity and Audience Reactions to Stylistic Mimicry Using Stable Diffusion
Ozgem Elif Acar, Hannah Jamet-Lange, Margaret Johnston, Peter Morgan, Sophie Ogilvie-Hanson
Generative artificial intelligence tools have developed rapidly. Pour project start just as text-to-image generation software was becoming increasingly popular and readily available through programs such as Stability AI’s Stable Diffusion, Open AI’s DALL-E, and the independent research lab project Midjourney.
Each of these programs allows users to input a short text prompt that describes a scene, concept, and/or style, which will then be output as a photorealistic image, or artistic work in a variety of styles. These programs offer an exciting new way to engage in artistic creation, and provide a new realm of imagination and creativity.
While generative AI programs afford new potentials in the areas of artistic creation, they have not been entirely welcomed by the visual arts community, spurring anxieties about ownership and copyright issues for artists whose work has been fed into training sets without their consent or approval.
These anxieties point to the insecurity of artistic labour, the legitimate concerns of artists, designers, editors, and other members of creative industries being automated out of their roles, or losing freelance and commission work to generative AI, and larger concerns about the data ownership.
Beyond the economic anxieties elicited through generative AI, larger questions appear: can AI programs create art that is moving, poignant, or otherwise evocative for audiences? Does the knowledge that a work of art was created, in part or in full, using machine learning change the way that audiences react to a work? What amount of labour may be considered human, and might we consider the creation of text prompts to be considered an artistic act?
Concerns about whether AI art is ‘real art’ and the question of whether or not AI machines can create an authentic artistic expression were shared by our group members during the early conception of our project, and flagged as a common knee-jerk reaction that were interesting to explore and address. Generative AI’s diffusion has entailed panics just as much as serious debates about art, replication, and automation. As in the conception of photography, the drum machine, and Adobe’s Photoshop, reactions to generative AI reiterate concerns that automation might taint artistic creation.
How, then, might we conceptualize notions of authenticity and originality in the area of AI-generated art? These are significant and context- dependent questions which we do not aim to answer, but rather provoke and probe through the creation of our Twitter bot project.
Our Twitter bot presents the audience with a series of Twitter threads; each one consists of one tweet embedded with the composite images of 3 works from an established, well-known artist in addition to one created by Stable Diffusion, created in the style of that artist. In two cases, all four images are the creation of the Stable Diffusion image generator. The name of the artist is included in each image, above the four artworks, and each work is identified with a number, 1 through 4.
Underneath this first tweet, a poll is threaded. The poll asks the user to identify which of the four artworks they believe to be the AI-generated one by picking the number associated with that artwork in the above image. By asking the audience to interact with the Twitter poll, the hope is to encourage interactivity with the project, and potentially learn from any resulting conversation in the replies. Polls on Twitter do not reveal their results until after the user has cast their own vote, meaning to satisfy the curiosity of knowing what everyone else thinks, you must cast your own vote first, and by not revealing the correct answer anywhere in the poll, one might be tempted to scroll through the replies to see if anyone has the right answer.
By prompting the audience to interact with the bot by choosing which artwork they think is the AI-generated one, we hope to generate conversation concerning questions of artistic style and hopefully develop insight into our research questions. What is it, exactly, that makes art ‘authentic’ to the artist and to their audience? How do the parameters of labour, process, decision-making, skill, and novelty that have been used to demarcate creative activity become reimagined or reorganized through AI-generated art? Is the ability to determine the exact process through which an artwork was made an important factor in judging its artistic merit and novelty?
The design of our Twitter bot invites the Machine Agencies followers into a dialogue regarding the role of mimicry, replication, and style in the development of text-to-image generative AI programs like Stable Diffusion. We hope to introduce the larger questioning surrounding the concept of authenticity in artistic creation, as viewers become aware of their own affective responses to the AI-generated replicas we have created. By comparing and contrasting multiple artworks by a single, prominent artist and determining the ‘fake’ amongst the originals, viewers may also become newly aware of aesthetic differences between human and machine-created works. While taking note of the elements of artistic works that are best suited for AI-replication, viewers may also recognize the failures or ‘glitches’ that emerge in the process of text-to-image generation. By invoking the artistic and theoretical legacy of glitch, we draw connections between the types of unintended and unexpected elements of this new medium and the specificities of previous analogue and digital technologies.