Anthropomorphism in the Intersection of Artificial Intelligence and Art

Apr 14

ABSTRACT

While Artificial Intelligence is gradually being used more commonly among artists, the field is full of anthropomorphic interpretations of the functionalities involved. This spans from Turing’s seminal work that introduced a human reference of intelligence to more recent advancements in the field that contributes to the visual applications. In particular, generative models – such as Google DeepDream and GANs which were among the models that first caught the attention of new media artists – are subject to a significant amount of anthropomorphic presentations. We hear about how a model “hallucinates” about something, how it “remembers” a certain class of images, or how it “understands” the features in an image. Although some of the fundamental parts of the field have been deeply inspired by human biology and cognition, any further emphasis on humanized interpretations of practical instances in the field can be an impediment to forming an unbiased relationship with Artificial Intelligence in the realm of artistic creation.

Artificial Intelligence is being increasingly used in a wide range of fields these days while it has also been a major source of controversy. Even the naming itself is highly debated as there is hardly a definition of intelligence that people from different fields agree upon (Goertzel and Wang 17). Although this paper discusses the influence of anthropomorphic language in some parts of the field – mainly the intersection with art – the focus is not to scrutinize the name itself or to investigate how well the name describes the field in its entirety. Rather, I will accept the ambiguity of the phrase and refer to it only as a denomination of the technological developments academically categorized under Artificial Intelligence. From this vast field of studies, the discussion will be mainly around the prominent branch of Artificial Intelligence that has also excited the attention of artists in recent years – Neural Networks. I would like to further narrow my focus in this paper to a specific family of architectures in Neural Networks known as generative models. The generative models are special in terms of their ability to not only solve problems such as classification but also to synthesize visual (as well as auditory and textual) outputs based on the data that the network has been trained on. The generative property in this branch of Neural Networks drew artists’ attention and led to the implementation of these models in many art projects. What will be explored hereafter, is the trace and the influence of the anthropomorphic language that surrounds this intersection.

Anthropomorphic views in Artificial Intelligence date back to early works in the field and expand to the most recent developments. In 1950, Alan Turing introduced a human-centred test for machine intelligence in his seminal paper Computing Machinery and Intelligence (Turing). He proposed his famous Imitation Game to replace the original question of “can machines think?” Other tests of machine intelligence like the Winograd Schema Challenge (WSC) follow the same theme of putting the machine against a human to play a game of questions and answers in order to determine whether the machine can be considered intelligent (Levesque et al.). Although these tests mostly address the topic of general intelligence or strong Artificial Intelligence, what is noticeable is the fact that human reference has been a vital part of the field since its early days. The field has also been deeply inspired by the human brain and its functionalities. This is evident in Neural Networks that account for some of the most important advancements in the field over the past decade. Neural Networks are constructed of interconnected units that are called Perceptrons (also known as artificial neurons). The concept of Perceptron was first introduced in 1943 as a simple mathematical model for a biological neuron (McCulloch and Pitts). It is no surprise then if the concept of Neural Networks is considered a simulation of the brain itself, the result of which would be the expectation of the same brain functionalities to emerge in the network – or an estimation of those functionalities at least. Phrases such as how the network “learns” is prevalent when talking about Neural Networks. Of course, learning in the context of Artificial Intelligence has a specific and accurate definition and the word “learning” indicated a similarity to our general perception of learning as a human functionality. Similarly, in generative models, what is generated in the network is essentially limited and visually similar to the dataset on which the network has been trained on. This creates the notion of how the network “remembers” what it has “seen”, where remembering and seeing have their specific scientific definitions and are not identical to the same human functionalities. However, the ability to synthesize images in generative models created a new avenue for artists to explore and resulted in some unprecedented works.

Mike Tyka is among the early artists who used generative models in his works. He was also one of the collaborators in creating the DeepDream algorithm in 2015 (Mordvintsev et al.). Blaise Agüera y Arcas, another contributor in the creation of the DeepDream, describes these works as “rather crazy, kind of cubist, surreal [and] psychedelic” (Arcas). Refik Anadol is another artist who uses AI in his works. His project Archive Dreaming is an immersive media installation that employs machine learning algorithms to search and sort relations among 1,700,000 archival documents and projects the generated results back. In the description of the project, we read: “Archive Dreaming . . . is user-driven; however, when idle, the installation “dreams” of unexpected correlations among documents” (Anadol, Archive Dreaming). Anadol also uses the term “hallucinations” in his works, both as a title or as a technique in other works such as Archive Dreaming (Anadol, Machine Hallucination). These “hallucinations” are quaint and mesmerizing morphings of different scenes into each other in a way that each individual scene can not be distinguished from what it has emerged from and what it is being morphed to.

These unprecedented results seem to be in a way magical because creating such visual outputs using conventional methods can be impossible or extremely challenging. However, it appears that the use of Artificial Intelligence makes it almost effortless to achieve them – at least from the public audience’s perspective. I would argue that these results seem magical only when you don’t have a technical understanding of the operations. Of course, not everyone has the ability nor the will to dive into the technical details of how the results are generated. Even an enthusiastic person who is not from a computer engineering background would be left with a significant amount of prerequisites in order to achieve a minimum accurate understanding of the operations that are responsible for the synthesis of these magical results. There seems to be a considerable gap between what is happening in the body of Artificial Intelligence and the public notion of these developments – and the public here includes anyone who doesn’t have the technical knowledge to understand the technical language of the field that is mostly mathematics. However, this gap of understanding is partly being filled with the use of metaphors and analogies. Terms such as “learning”, “understanding”, “seeing”, “dreaming”, “hallucinating”, “remembering”, etc are metaphorically used to describe the operations inside Neural Networks and, in my belief, these metaphors work quite effectively in providing a general understanding of the operations and scratching the surface. The issue arises when these analogies and metaphors are the only tools to provide an explanation of fundamental functionalities of the neural networks for those who are not equipped with scientific tools for a more accurate understanding. When trying to understand a concept, it is typical to use examples to help gain a minimum understanding, but it can also result in the generalization of the example. Mike Tyka addresses this issue in one of his talks about DeepDream:

It’s very important that we don’t anthropomorphize these systems. It’s very tempting to do, I know. We always use words like . . . dreams or interpret[ation]s or whatever. We have to remember these are really simple algorithms in many ways. They are just algorithms that can extract patterns from data, very large amounts of data. (Tyka)

It is also interesting to see how different the terminology in the technical discourse of the field is compared to the public discourse including the discourse around the intersection of art and Artificial Intelligence. An enthusiastic reader can simply check this by skimming through some of the seminal papers that contributed to revolutionary advancements in generative models such as GAN (Goodfellow et al.), StyleGAN (Karras et al.), Style Transfer (Gatys et al.) and DCGAN (Radford et al.) to see how these texts are empty of the use of such analogies and metaphors.

Frequent use of metaphors – that are mainly associated with human qualities and functionalities – can result in an anthropomorphic mental picture of the field. In chapter nine of her book How We Became Posthuman, Catherine Hayles describes how these mental pictures are formed. When analyzing a promotional video that was created to describe the core concepts of Artificial Life, Hayles points to the fact that verbal formulation is as capable of evoking a mental picture as actual pictures (Hayles 228). I believe these fields can be considered very similar, namely Artificial Life and Artificial Intelligence, in terms of the analogical namings and a gap between technical operations and public interpretations, but Hayles’ point is generally valid nevertheless and her argument is not limited to the Artificial Life domain. All the aforementioned terms that are metaphorically used to describe the functionalities and outputs of Artificial Intelligence to the public audience, if not explicitly clarified that are metaphors, can result in a mental image that would also be subject to personal interpretation and would essentially be inaccurate. The knowledge gap between the scientific and public representation of the concepts, that causes the inaccuracy, would hence be filled with public imagination. The formation of anthropomorphic mental images of Artificial Intelligence is not only a result of the analogical verbal formulation. There are actual pictures as well that reinforce anthropomorphism – and by actual pictures, I mean any form of visual representation of the field or the results of the field. Speaking specifically about the use of Artificial Intelligence in arts, one good example is the so-called AI robot Ai-Da, a humanoid robot that is represented to be the author of abstract paintings (Block). What we see of Ai-Da in the news, are photographs showing a robot with a human face and mechanical arms, standing by a painting canvas holding a brush and colour palette with a nice smile looking into the camera. What imagination adds to this representation is however a set of qualities from the mental picture we already have in our minds of an artist – a human artist. This mental image is reinforced with the title of the news article: “AI robot Ai-Da presents her original artworks in University of Oxford exhibition.” Now the mental image of an autonomous “artist” who does abstract paintings and attends exhibitions is completely formed. However, for someone following the trajectory of the field, it is not hard to imagine how curated these images can be.

From the early and fundamental works that introduced a human reference for intelligence to recent developments in Neural Networks that resulted in generative models that created a new avenue for artists, the footprints of anthropomorphism are abundantly found in Artificial Intelligence. Sometimes it is rooted in the ambiguity of basic definitions of the concepts in the field, sometimes it is in the form of artistic representations of the functionalities, and sometimes in the form of analogies to the closest understandable examples to fill the knowledge gap between the technical and public discourse. The critical role of human intention and supervision as essential parts of the generative models is often overlooked in the shadow of an independent and autonomous mental image of Artificial Intelligence. In all the cases, a possible consequence of this phenomenon can be a bias towards forming a rational relationship with Artificial Intelligence in the realm of artistic practices. A relationship that is more based on an adequately accurate understanding of the field that acknowledges the actualities of abilities and limitations and less based on imagination.