GenAI? It’s the delightful horror
Generative artificial intelligence has its own form of agency, despite its lack of intentionality. The rhetorical energy of the machine catalyses aesthetic experiences under the sign of the sublime. And it offers itself to us as an unprecedented glimpse into the world.
We are inclined to reject the idea that the generative ability acquired by artificial intelligence in recent years reflects some form of creativity. Whenever we ask large multimodal models to ‘create’ a text, an image or a sound capable of stimulating our aesthetic sensibility, we are relieved from our initial agitation by the realisation that, at the end of the day, it is an imitative game. Nothing authentic, and therefore nothing to worry about. Our primacy is safe, we tell ourselves, because the machinic processes underlying artificial intelligence are radically different from those that characterise human creativity. Although it is not at all clear what this human creativity consists of. And as much as we have always known that artistic expression also has to do with imitation, combinatorial play and mathematics.
But while we wait to better understand what creativity is, it may be interesting to formulate a hypothesis: machine artefacts exert their performative power on our sensory structures and thus possess an aesthetic force. What follows is precisely a reflection on this hypothesis. I will use two conceptual elaborations or categories:
- agentivity, which is better referred to by the expression agency;
- rhetorical energy, in the sense proposed by George Kennedy (1992).
In what sense can we say that artificial intelligence is endowed with agency and rhetorical energy?
The agency of artificial intelligence
In sociocognitive theory, agency is the human faculty to act in a social context, bringing about change (Bandura 2006). Agency manifests itself as causation, in the sense that between human action and change there is a cause-effect relationship: one determines the other. But, in socio-cognitive terms, agency must also fulfil another requirement: intentionality. Action in the social context gives substance to an objective or a desire, and its effectiveness consists in its capacity to produce not just any effect, but the intended one. This should not lead us to confuse agency with free will, since all human action is subject to constraints of various kinds, not only social and cultural, but probably also biological, as discoveries in the neuroscientific field lead us to believe (DE CARO 2011).
On the other hand, if we move away from the field of cognitive sociology, we may also be tempted to attribute agency to non-human living beings (plants and animals) and to entities that are animate but not alive (whether natural or artificial). But it is very difficult to recognise an intentionality to them. Certainly, if there is an agency of flowers, wind and machine, it is an agency devoid of will: flowers do not ‘want’ to attract bees, the wind does not ‘want’ to stir the sea, the machine does not ‘want’ to do whatever it does.
Computational performativity: influential machines
Let us focus on the alleged agency of software. It’s a point on which there is no consensus. In the past, there was no lack of scepticism about it (Miller 2007), while in more recent times some authors have come out in favour of the hypothesis (e.g. Accoto 2017). Floridi, specifically, refers to the agency of generative artificial intelligence, with respect to which he speaks of ‘agency without intelligence’ (Floridi 2023).
In any case, digital technologies do not have the character of neutral and passive tools. Rather, we can understand them as agents endowed with an evident performative capacity. It may therefore be interesting to question the nature, effects and limits of such computational performativity. A remarkable work in this direction is proposed by Miles C. Coleman with his Influential Machines (2023). The starting point consists in considering the agency of software and artificial intelligence in terms that are not procedural, but rather gestural. This involves shifting the analysis from the epistemic level (referring to a knowledge about things, i.e. a capability that software is unable to express) to the ontological level (concerning the relationship that software establishes with things).
Rhetoric as an extralinguistic dimension of communication
Coleman resorts to the concept of rhetorical energy, borrowing it from Kennedy to explain the character of computational performance. In his work, Kennedy deconstructs the historical process of the ‘literatisation’ of rhetoric, which began as a persuasive skill and transformed over the centuries into a system of language governance. The American scholar suggests that rhetoric can best be understood as a form of energy inherent in the extralinguistic and unintentional dimension of communication. It is the physical energy consumed by the agent enacting a performance (in this sense, we could speak of an actant) and perceived by the recipient of the message. This energy, Kennedy observes, is expressed through bodies, including those of non-human agents. It manifests itself in the crying with which the newborn expresses its needs and persuades the parent to act to satisfy them, but also in the bellow of deer in love, as well as in the colours with which flowers attract pollinating insects.
Coleman goes a step further than Kennedy by suggesting that rhetorical energy also manifests itself in non-living animate agents and thus, in more specific terms, in computational machines. If, both ontologically and in terms of effectiveness, we admit a connection between the energy expressed by the speaker, towards whom the eyes of the audience turn, and the performance of the sun, which induces plants to orient themselves in its direction (heliotropism), why should we not recognise a similar capacity for orientation on the part of software?
Ultimately, the agency of software could correspond to its rhetorical energy. Coleman recalls several examples of this computational rhetoric, i.e. the persuasive force that software exerts on us. The area that interests us here is that relating to the ability of computational machines to catalyse aesthetic experiences. Coleman believes that such experiences fall into the category of the ‘sublime’, which differs from the ‘beautiful’ because it is characterised by a complex combination of joy and horror: it is the sense of bewilderment that accompanies the marvellous, spoken of by the anonymous extender of the treatise Περὶ Ὕψους (Pseudo-Longino 1965), or, in other words, the ‘delightful horror’ conceptualised by Edmund Burke in 1757 with his famous Philosophical Enquiry (Burke 2014). The sublime qualifies the pleasure experienced in the presence of an object that transcends our control and understanding.
The censusAmericans bot and Refik Anadol’s immersive experience
Two examples can be cited here that point in this direction.
The first example is given by Coleman himself, in the second chapter of his book. It is censusAmericans, a bot created by Jia Zhang in 2015. The software, written in Python, processes short biographies of Americans based on data provided by the US Census Bureau, turning the numbers and data codes into mini-narratives. Every hour the bot posts a new biography on Twitter (today X) and will continue its task automatically until it reaches the end of the 15,450,265 lines in the dataset (Zhang 2015). “It will take about 1,760 years,” suggests the author of the programme herself, not without a certain amount of irony.
Zhang’s game consists of inserting strangers into our lives at regular intervals, but also biographies in which each of us can mirror ourselves, a bit like the epitaphs in Edgar Lee Masters’ Spoon River Anthology. We are in the realm of experience enabled by software through random or permutative processes. We could speak of ‘generative computational aesthetics’ and ‘computational sublime’ (McCormack & Dorin 2001). As Coleman puts it, the experience of censusAmericans is nourished by a mixture of gratification and anxiety in the face of the open and elusive character of the infinite at work, a character implicit in a computational performance in which control has been ceded to the generative activity of the machine (Coleman cit, p. 51).
The second example is still inherent to the field of permutation, but moves us from the textual dimension to the multimodal dimension (integration of text, images and audio). I refer to Refik Anadol’s immersive installation Renaissance Dreams, realised in 2020. The work was developed from a large collection of data on painting, sculpture, literary texts and architectural works dating between 1300 and 1600. These datasets were processed through GAN (generative adversarial networks) artificial intelligence techniques, designed to produce a multidimensional site-specific dynamic form. The installation was in fact conceived by Anadol to interact with the architectural space and the technical infrastructure of the immersive room of MEET, the International Centre for Digital Art and Culture in Milan.
The machine ‘reimagines’ the works recorded in the database, while the dynamic system incessantly changes their shapes, colours and sound design, performing the universe of data latent in it in an emergent manner. Once again, it is right to invoke the category of the sublime. Renaissance Dreams submits the entire Renaissance corpus to us in a completely new perspective. A perspective that questions us, but at the same time beguiles us with its grace. The machine hallucinations, a well-known problem of GPT and the other great linguistic models developed in the field of generative artificial intelligence, become here an aesthetic opportunity, because they place us in the presence of the improbable. What disturbs us is the way the machine reads Renaissance iconography, recording patterns, correspondences and suggestions that would escape the human eye. Because its computational power allows it to ‘see’ what we do not see and returns it to us with great rhetorical force.
The technical image and the loss of centrality of the human gaze
As is well known, the idea that the camera manifests its own autonomy in turning its gaze on things in the world is at the heart of Vilém Flusser’s thinking. The Czech philosopher speaks of a ‘technical image’, referring precisely to the product of an obscure process that develops within the camera (Flusser 1985). By interposing itself between us and the world, the camera manifests its own ‘subjectivity’. This interposition disturbs us, because it questions the centrality of our gaze. The dismay in front of a horizon that does not belong to us is the same felt by Martin Heidegger when he confronted with the first image of the Earth seen from the lunar orbit and referred to in the famous interview he gave in 1966 to ‘Der Spiegel’ (Heidegger 1976).
The German philosopher was alluding to the photograph taken on August 23rd, 1966 from Lunar Orbiter 1, the unmanned spacecraft used by NASA to prepare for the Surveyor and Apollo missions. That image seemed to him the epitome of man’s uprooting: the place where we live is no longer Earth. But the uprooting can also be understood in another sense. The point is not just that we have severed our roots, which until yesterday were firmly planted on Earth. It is to take note of the shift in point of view: from the human to the machine. The photo taken by Lunar Orbiter 1 certified, along with the agency of the machine, its ability to ‘see’ things from a different perspective than ours. The machine became a device of enunciation.
From the rejection of writing expressed in Plato’s Phaedrus to the present day, the unease that the experience of technology provokes in us seems to be a constant. But the power of modern technology, in which Heidegger sees the realisation of the essence of technique, appears to us today as a definitive threat. Yet Heidegger himself contested in explicit terms the idea that we are overwhelmed by technology, suggesting, if anything, that we have not yet found our way (Ibid., p. 206). And so he concluded by calling for the conquest, on the part of the human being, of a sufficient relationship (“ein zureichendes Verhältnis”) with the essence of technique (Ibid., p. 214). It is precisely in the search for this relationship, I believe, that we should renounce taking a position of utterly sterile outrage at the supposed dethroning of the human by machines. And we should, if anything, critically deconstruct a naive anthropocentrism, considering the computational gaze on the world as a valuable instrument of knowledge.
