Contribute to this project at xhairymutantx.whitney.org

xhairymutantx began by asking the question: do we get to choose how we are represented on the AI substrate?

One of the original promises of the internet was self-determination of identity online. As the famous 1993 New Yorker cartoon put forward; “On the internet, nobody knows you’re a dog”. The promise of a digital existence brought with it the promise of reinvention, of mutable, often multiple, identities and existences.

There is a common concern that with the proliferation of powerful and opaque AI models, we are entering into a new era of untruth. As such, great efforts are taken by companies training AI models to attempt to align AI outputs with what is a kind of consensus truth. As ChatGPT trains, accounts of historical events are favored from more reputable sources, and text to image models attempt to conjure as close to an objective result as possible.

This imperative to treat all subject matter as objective has utility; when most people prompt “chair”, it is reasonable to expect an image to generate that resembles what most people believe a chair to be. For more subjective areas, such as “beauty” or “art”, claims to truth become much more complicated, and consequential. The concept of “beauty” understood by CLIP (the text model that underpins text-to-image systems) reveals a certain kind of truth, albeit a kind of aggregate sampling of whatever images were captioned as “beautiful” in its gargantuan training data. An aggregate reflection of everything posted online, an egregore, an aggregore.

“Art” is weighted heavily towards what art most appears in the data, which explains perhaps the prevalence of more contemporary fantasy illustration or concept art in generic text-to-image outputs. This attempt to determine objective ground truth makes sense in a journalistic or commercial context, however in realms like art, culture and identity risks tethering us to populist, gameable or impenetrably reductive representations of areas that are crucially fluid. When it comes to an individual, and their representation within AI models, what is considered objective truth about them? We have been fortunate to be able to explore this in depth, with Holly appearing to cross the threshold of notoriety sufficient to have an embedding, or concept of her, present in most AI models. If you prompt “Holly Herndon” in a text to image system, you will be returned an image that resembles Holly. Quite what of Holly is known is something we explored in our AI portraiture project CLASSIFIED (2021), using a technique (later referred to as “Textual Inversion”) created by our collaborator Jordan Meyer to deeply excavate Holly’s embedding within public AI models.

It appears that, as for many living artists for which there are many photographs of their likeness available online, Holly’s concept within models is defined by her hair. Not her artwork, as would be the case for historical artists for which we have few documented images other than the work they left us. Barely her unique facial features, as it appears that only the most notorious of us have enough pictures of their face online to be able to be convincingly rendered as a portrait. Rather, the concept of Holly compressed into models has zeroed in on her hair as her quintessential essence. There is a funny truth to this. To stand out in a busy feed we all find ourselves increasingly compressing our arguments and appearances, or having our arguments and appearances compressed, for maximal loudness. If one were to ask a human for a shorthand account of Holly, they might reliably refer to her distinctive hair.

As such when attempting to reform Holly’s concept within AI models, we came to the conclusion it makes most sense to build upon the conceptual foundations already established, out of the fear that any training data produced that strays too far from what models understand Holly to be being filtered out as inaccurate data. The prospect of “data poisoning” has recently captured imagination as some effort to assert agency over wanton training practices, however in our own research (with projects such as Spawning’s Kudurru network), we have found that for most well established concepts, poisoning may be a less effective tactic than amplifying the cliche. Reductio ad absurdum. Playing the concept compression game. Cliche amplification in this sense would represent identifying the purported essence of a concept, and amplifying it to ridiculous proportions, sufficient that the model will be primed to accept this new mutant amplification as truth.

To this end, we began with the goal of amplifying Holly’s cliche, and constructed a costume (tailored by Franziska Muller and Lenna Stam) in which Holly was overrun by her hair. She takes on mutant, promethean, proportions, and her hair, like kudzu, begins to invade and envelop her. We used images of Holly wearing this costume to fine-tune an image model, and that model was recursively refined to produce a consistent character that is able to be spawned by anyone using the interface provided. This model can produce infinite images of this new character. The images produced by this model will mostly all, in some way, be infected by the hairy mutant.

Offering this model to the public serves two functions. First, it allows anyone to explore this model, the artwork, as a medium. Which is quite fun. Second, it allows for an abundance of new material to be produced, hosted on the reputable Whitney.org site, to be ingested for new AI training. Each image will be captioned and tagged “Holly Herndon”. We hope that what is produced by this model may play a hand in defining the concept of Holly in future models. There is little reason why it would not.

This dance, of asserting some self determination over identity, and also ceding control of the production of new identity, is consistent with principles we established with our Holly+ project. Better that whatever Holly becomes in future be something active, rather than passive, subjectively deliberated, rather than objectively assigned. We can propose, others will define.

Given the new territory of offering a text-to-image model as an artwork, we have taken some measures to constrain the output of the model. We anticipate many will attempt to produce incendiary imagery that has nothing to do with the objectives of our project. In which case, kittens will be produced. If Holly ends up metamorphosing into a kitten, it will reflect malicious attempts to weaponize this project for other means. We have also taken measures to omit the usage of living artists styles in the prompts, with a similar reasoning.

Thanks to the Whitney Biennial curatorial team, Meg Onli, Chrissi Illes, Christiane Paul and Colin Brooks for your contributions, guidance and good humor. Thanks to the Whitney Museum for being generous with your good standing in this experiment.