SIMON TAYLOR

It struck me text-prompting an AI image-generator that the image it generated was as much the result of the text as the image pool the AI used. Not only this, the resultant image is as open to contest, for example open to the charge of cultural appropriation, as if it were a part of a wider cultural discourse, but, this time, now, a cultural discourse including images; and yet …

Benjamin posited the dialectical image. The AI image is not this but a discursive image.

The image quotes sources. It adds nothing imaginary. Image-quotes are simply recombined as per the script, the text that is the input to elicit the image as its output.

What is the relation between text and image? To say it’s a matter of input and output doesn’t get at the reason for the image, which exists as an element of discourse. It is therefore quite unlike simulacra and even if it quoted simulacra it would be.

As if they were textual fragments, the AI lifts from the images in the pool those which meet certain criteria of representation, and recombines them. Recombines because through the sifting process, by which the image-generator selects its sources, those sources are already the product of recombination. They are already quotes for being quoted; or, memories (or memes) for being remembered.

The discursive image differs from the art image not for these reasons however; and it is not because it’s a copy or a falsification that it differs, as opposed to an authentic production and a true representation (as simulacrum). It differs from the art image in general and from the digital art image in particular for being literal. The discursive image enters discourse, as the AI-generated image often does, in contestation of the ontological status of images to which we are habituated and outside of any relation to representation. It has no representational origin or goal. It enters discourse as a problem of discourse and for that reason already discursive.

It is easy to impute and impugn thereby the AI-generated image to an intention, that of the text-generating human; but it’s really in vain that the attempt to pin it back onto the textual input is made. The images are already out there in digital form to be analyzed: and it is at this (for us) sub-representational level that the fit and relation to representational criteria is established. That is, the representation providing the criteria used by the AI image-generator is textual and the image and images used to make the image outputted are digital: they are imperceptible to the eye. They are electronically perceived.

Even the idea of the image-generator reading the input-text is wrong. We don’t know how AI perceives either text or image since it neither reads nor sees. From the same place then as it parses the meaning of the text prompt and recombines the images at its disposal correctly or adequately comes the potential for failing at either task or both. In this sense, except that we know or think we know how perception in us works, it is just like us. However, what for us constitutes the advantage of an elaborate nervous system to interrupt and arrest or to transmit the response to sensory excitation by nervous stimuli AI does not have. It is more like a microbe, the sensory apparatus of which does double duty as also providing it with motility; so the AI image- or text-generator senses with the same apparatus as it generates, without interruption, without being able to select which of the sensory inputs leads to image- or text-outputs, without the delay. In this delay, in this hesitation is the key to consciousness.

Then, in retrospect, in immediate retrospect, all images that are connectible, able to be harvested by AI image-generators and combed by algorithms, form, as discursive potential and resource, discursive networks. No longer bodies of images, they are distributions, discourses. And the images in them similarly belong to discourse rather than any corpus.

First published on squarewhiteworld on 14 June 2024. Re-published in herri with kind permission of Dr. Simon Taylor.