The technologies that are employed in “AI art” derive from machine learning [41] and deep artificial neural networks, in particular [20, 35]. We will identify artistic interaction with such technology as neural media praxis. As we shall show, neural media disrupts traditional notions of artistic intent, offering instead a reflexive mediation of human presumptions about what it means to see and create when artistic intention is disrupted by the opaque algorithmic translation of AI mediation. Our interest here is in the qualitative distinction between embodied autographic (“traditional” or “tactile”; see also [8, p. 25]) and computational algorithmic (“AI”) creative practices.

Philosopher Joseph Margolis provides a strong argument for artworks being “physically embodied (by “embodied” Margolis is referring to an artefact as “possessing physically perceptible properties in their earthly instantiations”) and culturally emergent entities” [36, p. 68], denotations in the world of perceptible properties, objects that nonetheless lack fixed (invariant) nature. Margolis argues that “what we take the stable denotation to be is proposed, revised and entrenched in the interpretive process itself” [p. 95]. An artwork is a conceptual blend of the discursive and non-discursive thing, a metaphoric gestalt conflating the material artefact with its temporal reading, a suggestion of immanent meaning constituted in a transient representation devoid of any theoretical genotype or universal. Margolis [p. 67ff] thus holds that the work of art is a cultural construction with intentional referent and predicative dimensions. Taking this framework as informative upon the hermeneutics of art in the era of artificial intelligence, we seek to show how neural art is modeled after but ends up challenging assumptions about our relationship with our tools of expression. Algorithmic code presupposes repeatability, an anticipation of the mathematical invariance, but what emerges from neural media praxis is opaquely tied to the subjective intentionality embedded in the artist–technology relation, a stochastic denotation in an aesthetics of ambiguity. Neural media does not offer us singular referents but rather the intentional indeterminacy of potential images [14].

Our strategy in examining this reification of the artwork is to explore the metaphors drawn from practices of autographic image making and speculate on predicate associations with entities in the neural media ecology.

William Seeley, questioning the epistemological constructs of cognitive science of aesthetics, draws from Baumgarten [11] that aesthetics must motivate the clarification of the structure of one’s phenomenal experience through “the intuitive apprehension of the latent structure of an artwork” [57, p. 211], but that, problematically, functional descriptions of perceptual stimuli—as is typical of computational models of neuroaesthetic perception—do not explain how artwork generates critically subjective aesthetic interest. Seeley proposes that an accounting of the relation between perception and semantics (apprehensive relevance) is necessary to develop naturalized models of computational aesthetics. In line with Seeley’s underlying message, that understanding in art and science is furthered by a sharing of perspectives, Johan Wagemans [66] has suggested that the interdisciplinary nature of empirical aesthetics benefits through a collaboration between artists and psychophysics researchers that “tap[s] into [the] more private, subjective aspects of aesthetic experiences” in order to “tackle research questions of joint interest in ways that are found acceptable by both, and yielding results that are both scientifically and artistically valuable and useful” [p. 673]. This is the strategy entered into in this article, to know more about the psychology of aesthetic perception, to learn more about the relation between intuitive experience and the empirical dimensions of perception, is to better understand how technical mediation of perception perturbs and interacts with creative intention. We can therefore use the technical parameters and associated aesthetic dynamics of mediated creative interactivity as a probe into the psychophysics of aesthetic space. In neural media, intuitive interaction is encoded—by design—into the artefact.

There has been little discussion of how technologies of machine learning mediate artistic practice (for notable exceptions, see [9, 60]). Yet, these technologies assume already a kind of alterity [28, p. 97ff] if we cannot explain the aesthetic response we are likely to obtain when engaging with them. The artist learns from the machine, which has, in turn, learned from visual databases intended to represent a corpus of human visual experience. This existential relation embedded within the manipulation of images by artificial intelligence promotes creative practices that are subjectively motivated but ambiguously mediated. We want to ask, what artistic response this dynamic encultures, and what begins to emerge as the experiential features of post-phenomenological aesthetics?

Now we will turn from our theoretical stance so far and explore the phenomenology of working with neural media in our individual practices. We attempt to arrive at a deeper understanding of creative expression with neural networks through artistic illustration of the media’s algorithmic constructs. This section therefore takes an autoethnographic turn as we delve into the personal experience of AI art, exploring the creative affordances of artificial intelligence. The questions we ask ourselves are how does one work with this media? What marks can we make? What stories can we tell?

I am a traditionally trained painter interested in processes of abstraction. In these works, I am using a method known as style transfer to explore the intersection of visual form and aesthetic space (Figure 1) in an image-making process mediated by artificial intelligence. This practice currently consists of two-dimensional images and video animations created with proprietary extensions [38] to the Gatys et al. [16] style-transfer algorithm. I call these images “neural paintings.”

I started by experimenting with grids of subtle parameter incrementation across a range of image scales (Figure 2), exploring the notion of infinitesimal change in the still image. Small visual changes invoke depth perception, inducing pareidolia through resonance with embodied perceptual gestalts originating in genetic/cognitive ties (for theories in accord with my notion of resonance, where affective meaning is linked to determinacy, see [54] and [44]). So what style transfer was designed to do—pastiche representation of artistic style—is circumvented by an artistic intention that sits outside the singular image. I explore the aesthetic and psychophysical aspects of this notion through an art-as-research methodology.

The artist’s brush is a responsively variant technology that becomes embodied, familiar, and ultimately transparent in the act of painting. I wanted to explore how the “neural brush” behaves in the formation of evocative texture. But it is not like physical painting or collage, the image is not manipulable in a semantic or tactile sense. There is a very expressive element, not entirely under control but directable: more like the implicit techniques of art automatism, such as frottage (Choi is referring to a method of ‘autographic sampling’ obtained by taking transfer rubbings off surfaces. The technique was developed in 1926 by surrealist Max Ernst [61, p. 12], [13, pp. 7–14], and later extended into grattage, a method of scraping away a surface to discover the serendipitous layers below—a technique that it could be easily argued was widely adopted by American Abstract Expressionist painters in the 1950s.) than the explicit imagery usually associated with collage. Imagery hovers on the edge of discernment (Figure 3).

In my work with style transfer I think of “style” as a paint brush of variable scale, loaded with a somewhat randomized palette distributed across a range of blob sizes as if multiple operations of dabbing and swirling the brush quickly across a range of image colors had happened, which is then applied stamp-like by convolving the (computational) brush used to capture the style of the original content with the brush representing the “target” style to produce a new set of color patches for the content, which is itself weighted positively or negatively in its influence over the combined output (Figure 4). The composition is drawn out of latent potentiality more than explicit intention. In other words, I do not so much place the brush as define a spatial texture palette in a process of art automatism. Thus, “style” and “content” represent the intersection of different conceptual domains; “style” has more to do with my feeling about the work, whereas “content” reflects orientation toward lived space. (This dissociation between dimensions of cognitive processing is also reflected in perceptual microgenesis—the time course of perception. Research in behavioral [4] as well as neurophysiological studies [5] shows that the embodiment of spatial references happens very quickly, whereas style develops secondarily and is an entrained awareness. In other words, aesthetic perception is emergent.)

I do not think of this media as offering “control” so much as an interactive “play.” AI should provide an augmentation of human creative intention and for that we do not so much need control as transparency on a set of rules by which the game is played. A central concern in art making for me is the derivation of your own set of rules —perhaps this is what “style” is. The realization of intentional acts comes through skillful play. The tool needs to be “felt” more than controlled.

My neural painting process involves constructing elaborate digital abstract paintings that I think of as “multi-perspectival” visual experiences delineated into “organic” and “deterministic” workspaces. I populate these workspaces with various existential referents, the former entangled with the aesthetics of brain, flesh, and the metaphor of natural environments reflecting lived body worlds, the latter with mechanically contrived surfaces, industrial, electronic, and other impersonal grid-like structures. These two polarities are convolved against each other to arrive at liminal blended spaces where the distinction between human and machine “intention” becomes vague (Figure 5).

As an artist and cognitive scientist, I am interested in human creativity and its relationship with emotion. In making AI artwork, I find interesting irony in what I call “the ghost in the machine.” That is, although I know I am dealing with a “cold” machine, I think there is a possibility of bringing the spirit—the emotion—out of it, and I like to attempt to do that by making what I call “creativity systems” that move through a visual search space in some creatively inspired way. I am interested in latent space search strategies that reflect constructs drawn from cognitive psychology, the notion of creativity involving both analytical and associative thinking. At some point analytical process gets caught, and in getting caught one has to let go, defocus, go on an associative journey. I think that a lot of artists have this notion of fluidity between stepwise thinking and being very open and loose.

In the case of Deep Learning systems, this latent space search, this journey, is in iterative code form where I get an idea, and then using code scripts, try moving 10% in every direction from it, or in directions I intuit are the best way for me. I am looking to find new associations and new neighborhoods of interest. In Deep Learning, you can really play with these processes in deep ways by changing systems of relations.

I like to think about human creative process and try and emulate that in computational art systems, so we think about, let us say, a portrait painter: The first step would be to interview and understand a sitter, and to emulate that I use many interconnected AI systems, each acting as a subcomponent of a greater gestalt. Some systems might represent the emotional recognition of your sitter, then another system might segment subject from background, trying to derive form and meaning, and via conversation and research you would have knowledge of the sitter you would want to convey. The artist takes all this and emphasizes what is important and abstracts out the things that are less important to get at the inner beauty of the portrait. So, the notion of AI variable abstraction techniques in different regions helps bring out that inner truth to the digital canvas (Figure 6).

I like the fact that a painting is not one thing anymore, that I could decide to apply different perceptual levels and abstraction levels—painting becomes multi-dimensional and stochastically variable, entering into an intersecting space of conceptual and algorithmic association. For instance, a particular 360 viewpoint of me in a landscape (Figure 7) leads to a journey of playing with shape in new ways.

These trees can morph into this conceptual representation of the room that Van Gogh lived in (Figure 8) by emphasizing what I like most. Much of this can become more immediate, letting go to move more freely and intuitively.

When transformative “Generative Adversarial Networks” (GANs) [63],¹1

The term “GAN” is used in two importantly different ways: 1. A training architecture including an adversary to evaluate the results. 2. A model, usually a generator, trained using such an architecture, e.g. BigGAN. Furthermore, the use of GAN is not limited to generative (vector to image) applications. For instance, Pix2pix, CycleGAN, and BiCycleGAN are GAN architectures for picture-to-picture transformations.

I then turned to experimenting with generative GANs [21], which are usually trained with very large data sets in order to create a generator capable of producing a large variety of images. Such a variety runs counter to my intentional needs, as finding relevant images would still be problematic. To have more control, I started to experiment with small data sets (Figure 11). This allows me to direct the process through a careful selection of the images used in training. There is no clear separation between content and style; the generated images contain elements of style and content present in the training set, combined in often unexpected ways. The ambiguity of this separation is a useful feature in my process, as I am not after photorealistic results (Figure 12).

A major difficulty in training GANs is “mode collapse”: the network fails to produce variety. I no longer regard this as a problem; it is rather a characteristic of my process. At any given phase during training, there is only little variation, but as the training goes on, I am still getting different images. This is but one example of me, as an artist, pushing the technology in different directions than what the scientists do: toward the more specific, individual, private, while science is looking for the generally valid.

My technical background has certainly helped me in developing my tools and process, but my reasons for entering into visual art were related to expression, not technique. There is an ongoing tug-of-war between the range of my aesthetic preferences and the aesthetic range that a neural art process can provide. Techniques are but the tools of a process. Expression drives the exploration of the space between representation and abstraction.

We have presented reflections on approaches to neural media in the work of three artist researchers. We claim that threaded through all these approaches there is, as per Ihde, a sense in which the artist’s embodiment of an image-making practice is both “magnified” and “reduced” [28, p. 76] by AI technology. Here we discuss interrelated factors threaded through our exemplar neural media praxes, examining how the art object in neural media is variously polarized by magnification and reduction, lending a particular “fingerprint” to a typology of neural media praxis. In this analysis, we interpret these artists’ reflections on neural media praxis in a phenomenology of the mediation of traditional process metaphors.

Relations between the tacit and the expressed are explicitly embedded in neural media. Future developments of AI creativity support technology should be cognizant of the system of deep relations engaged in the formation of neural media artefacts. Such relations might be modeled as n-dimensional Evolving Transformation Systems [22], a mathematical model used to model natural processes, and by Nadin [47, 48] to model anticipatory systems. Another possible approach to modeling this autographic–algorithmic exchange is Predictive Coding theory (see Clark [10]), which has been shown by Muth & Carbon [45] and Muth, Hesslinger, & Carbon [46] to model predictive progress toward insight in the resolution of “semantic instability” in ambiguous visual experience. The praxis of neural media is intimately tied to tacit motivation and anticipatory meta-prediction on context-related future conditions and could provide valuable empirical data to perceptual cognitive science research if such process relations could be reflexively traced through the multivariate factors that inform them. Our approach here is to phenomenologically unravel these factors in order to inform the development of such models.

A post-phenomenology of the neural art artefact presents several comparative dimensions which we may employ to arrive at a deeper understanding of how art changes when AI technology enters human praxis. Initially, we have the basic relational structure of Fig. 13:

Artist–technology–artefact (A–t–a)

In traditional practices, this triad is conceived of as a transparency relation between the artist and the tool; the brush or expressive instrument, to the degree that it is embodied, becomes “transparent,” receding from the interactive relation with the artefact:

(A–t)–a

On the other hand, in artificially mediated creative praxis “awareness” shifts from an embodiment of the tool (transparency) along a disembodying vector where the transparency is no longer tied to human intention but instead moves toward a conflation of tool and artefact. This conflation—a computational poetic gestalt—is not easily separated, and the artist is pushed further from the state of immediate interactivity toward an emergent awareness-through-technology a state of affairs that Ihde calls alterity relations. Interactive opacity is emphasized:

A–(t–a)

We argue that neural media separates from embodied praxis along certain phenomenological dimensions that suggest possibilities for creativity support development. Modeling the apparent barriers between the human and AI in creative praxis also raises both technical and ethical issues associated with real-time human-centered control of which area in an image is being developed; to enhance continuity of response in the pursuit of illusion.

Opacity, as positioned here, is then reflective of the ambiguous dichotomy of Gombrich’s “representational illusion.” As technology advances, the perfection of illusion becomes more and more representational of the environment (technical/cultural) from which it emerges. AI image enhancement and manipulation is advancing at a pace some might find concerning, and for good reason. If we are not cognizant of the degree to which technology has taken the place of creative intention, at what point have we become cybernetic curiosities in Gottlieb’s Disneyland? No longer tourists, but the commodity bought and sold?

Praxis —the reflexive pursuit of embodied expression through artefactual manifestation — displays several featural dimensions, which we identify as those aspects of an art process that constitute its intentional relations. A feature is not an invariant property of a process but is always an interaction for some intended result. In working with neural media, we observe two featural dimensions separated according to their general situatedness in the relations of creative praxis and phenomenologically distinct from the feature’s status in autographic media. These are,