Positions in dialogue. Three positions that FACE & FICTION does not simply cite but has to answer – the closest comparison, the sharpest critique of its aesthetics, and the honest limit of its provenance.
Holly Herndon & Mat Dryhurst – Identity Play
With Holly+ (2021), Herndon and Dryhurst built a voice model trained only on Herndon’s own voice. Instead of forbidding its use, they call the principle “identity play”: others may perform through her digital likeness, openly and with consent. For the Whitney Biennial 2024 they went one step further with xhairymutantx: a text-to-image model trained on images of Herndon, costumed and exaggerated – above all her hair – so that her identity is deliberately transformed inside the model.
This is the closest precedent for FACE & FICTION: an artist trains a generative model on her own person and treats the model itself as the work’s material. The difference lies in direction. Herndon and Dryhurst open the model to others and distribute authorship. I keep the model closed and document every single output. Their question is: who may speak as me? Mine is: what remains of a self-portrait when the camera is replaced by a model that has learned to imitate me?
Key works: Holly+ (2021), xhairymutantx (Whitney Biennial, 2024).
Roland Meyer – The face as operative image, AI as nostalgia machine
The image and media scholar Roland Meyer (University of Zurich / ZHdK) writes about the face as a technically operative image: in Operative Porträts and in Gesichtserkennung (2021) the portrait appears less as likeness than as a means of identification. In recent lectures he describes generative AI as a nostalgia machine and an amplifier of clichés – structurally backward-looking, and especially suited to producing “genuine emotions” as instantly readable images. His term for its aesthetic is platform realism: an aesthetic of the generic.
This critique hits FACE & FICTION directly. My prompts ask for medium-format film, Ilford HP5, grain, flash, early-2000s snapshots – exactly the nostalgic surface Meyer describes. My answer is not to deny it but to expose it. The analog look is a parameter, written in the prompt and visible in the provenance block beside every image. The nostalgia is quoted, not believed. Whether that is enough is a question the work has to keep open.
Key texts: Operative Porträts (2019), Gesichtserkennung (2021), ed. Bilder unter Verdacht (2024).
Eryk Salvaggio & Abeba Birhane – Reading the dataset, and its limit
The artist and researcher Eryk Salvaggio proposes reading every AI image as an infographic of its dataset: a pattern of data inscribed into a picture, telling a story about the material and the human decisions behind it. His method starts with three questions: where did the data come from, what is in it and what is missing, and how was it collected?
The cognitive scientist Abeba Birhane has answered those questions for the datasets beneath open image models. Her audits of LAION-400M and LAION-2B – the data family behind Stable Diffusion – found misogyny, pornography, racist slurs and malignant stereotypes, and showed that hateful content grew as the datasets were scaled up.
Read together, they name both the justification and the limit of my provenance practice. Salvaggio explains why every image here carries its parameters: the image becomes legible. Birhane shows what those parameters cannot cover. The LoRA is trained on my own face, with consent. The base models beneath it are another matter. Stable Diffusion 1.5, used for the 2024 works, was trained on data from the LAION family that Birhane audited. Krea 2, used today, was trained from scratch on a pretraining set of billions of images; its technical report describes the filtering in detail – including the exclusion of AI-generated images and the deliberate retention of well-known people – but names neither the sources of those images nor any form of consent. I did not choose this data and cannot inspect it. FACE & FICTION documents the layer I control and names the layer I do not.
Key texts: Salvaggio, How to Read an AI Image (2023); Birhane et al., Multimodal Datasets: Misogyny, Pornography, and Malignant Stereotypes (2021), Into the LAION’s Den (2023).
Why these three
Herndon and Dryhurst show what a self-trained model can be. Meyer asks what its look conceals. Salvaggio and Birhane ask what its data contain. FACE & FICTION stands between them: an identity made into a model, a nostalgic surface made visible as a parameter, and a provenance record that is precise about its own limits.