AI Art Stands at the Border between Mimesis and Nemesis
The concepts of ‘mimesis’ and ‘nemesis’ were introduced in 《A Study of History》 written by Toynbee, a renowned historian. In any society, there are a small number of geniuses who perform original works. For their creative works to succeed, a number of people should respond to and imitate them. This phenomenon that many ordinary people imitate a small number of geniuses is called ‘mimesis’ which means imitation or reproduction. In the process of mimesis, ordinary people may fail to follow such creative geniuses or withdraw from mimesis. This phenomenon is called ‘nemesis’ which means impossibility to conquer or retribution.
Today, the waves of the 4th Industrial Revolution are roughly sweeping the world, and one of the core technologies — AI — is now affecting every aspect of our lives. We can find AI in any area of our life such as digital assistant capable of voice recognition for elderly people living alone, autonomous driving automobile, and chatbot as well as smart financial service app, cooking assistant robot, intelligent CCTV, etc. The area of art that has been regarded as only for humans is not an exception. Recently, AI has expanded its boundary to cover even music, dance, novel, and fine arts. as fine art works created by AI are increasing drastically, various issues regarding the artistic value of AI works are controversial. AI art is in the middle between mimesis and nemesis.
AI that draws pictures
Last month, ‘Theatre D’opera Spatial’ won the first prize in the digital art section of <Colorado State Fair Art Competition>. This work was created by the game planner Jays M. Allen in utilization of the AI program ‘Midjourney.’ In October 2018, <Edmond de Belamy> which is the first portrait painted by the first computer algorithm in history was sold at Christie Auction House surprisingly at USD 432,000 (about 0.5 billion won), the price far higher than expected (USD 10,000). <Edmond de Belamy> was painted by the AI algorithm that the startup in Paris ‘Obvious’ developed.
Software enterprises of data-based businesses contributed a lot to such rapid advancement of AI technology. These enterprises have competitively presented AI painters to make known the excellence of their AI engines. Major examples include ‘Drawing Bot’ and ‘Next Rembrandt’ of Microsoft, ‘Deep Dream of Google, ‘Deep Forger’ of Twitter, etc. In Korea, the graphic AI enterprise Pulse9 released the AI painter ‘Imagine AI.’ Works of Imagine AI are exhibited at the AI gallery ‘Gallery AI.A.’
Meanwhile, there are many AI based text-to-image tools that draw pictures in harmony with texts entered by users. In May 2022, ‘OpenAI’ that opposes of AI monopolization introduced the ‘DALL-E 2’ program that provides realistic images and pictures appropriate for the recognized text. Midjourney and Artbreeder that participated in <Colorado State Fair Art Competition> are two exemplary programs of such popular text-to-image tools. Recently, some AI systems not only produce simple images but also create a short video for an entered text command.
Experiments to create artistic works in combination of a hardware robot and AI also have long been conducted. In 1974, Harold Cohen, a professor at Yale, first introduced the robot painter ‘Aaron.’ Aaron continued to evolve for 40 years until professor Cohen died, and in the year of 2007, it was permanently exhibited in San Diego. In early years, robot painters drew pictures with harmonious abstract forms, and then in the 1980s, professional level robot painters emerged and started to draw the specific figure of a target object. In 1995, robots were upgraded to the level of coloring. Aaron makes judgment based on the previously saved information, drawing quite impressive pictures with selected colors and shapes.
In 2019, the robot painter named Aida emerged. It was developed jointly by the UK robot enterprise Engineered Arts, Oxford University algorithm developers, AI engineers at University of Leeds, etc. Aida is a humanoid robot that draws pictures using pencils, brushes, and paints. Aida was named after the countess Ada Lovelace who was a female mathematician and computer programmer in the UK. It has a figure quite similar with that of humans and has ergonomic hands. Its eyes are made of cameras. While existing robot painters created images based on learning data, Aida makes a sketch of objects that it observes through cameras, using a pencil.
Types of machine learning to create AI art
Unlike the Deep-Learning technology of AlphaGo that won the board game against Se-dol Lee, AI art uses the AI learning technology called Generative Adversarial Networks (GAN). The neural network of GANs continues to learn even without a human intervention or a data set to learn from, unlike other Deep-Learning mechanisms. This advantage leads to both expectations and worries. GANs are a model of mutual competition consisting of the generator and discriminator that learn from and affect each other. The generator continues to create ‘realistic but fake images’ and puts in the discriminator, and then the discriminator compares them with previous genuine images, returning the resulting values to the generator. The generator continues to create new fake images that are similar to actual images in reflection of such returned values. It repeats this process until the image looks exactly the same as the human-drawn picture. Through this process, a GAN can create a work that looks similar to an existing work but does not actually exist. <Edmond de Belamy> is one example of works created through GAN learning based on 15,000 portraits painted in the 14th to 20th centuries. However, GANs have a limitation that they can create a new similar image, deceiving the discriminator, but cannot create a new style. This is because it is the discriminator’s role that pictures of an original style are not regarded as pictures drawn by a human. To images of a completely different style, the GAN discriminator sends a feedback that they were drawn by a machine.
To overcome this limitation, Colin Martindale, a nueropsychology professor at University of Maine, US, suggested a condition necessary to create works of an original style: When AI classifies works, they should not be classified as one of the existing painting styles such as impressionism and cubism, and at the same time, they should be of a form that can be called an artistic work.
Based on this theory, the AI team of Facebook proposed the model called ‘Creative Adversarial Networks (CAN).’ CANs are a modified version of the GAM algorithm that better fits artistic works better. With CANs, it is possible to draw pictures in a style that currently does not exist. CANs collect information to classify existing pictures and styles based on Wiki Art data sets as the basic learning data. Wiki Art classifies 81,449 images of 1,119 artists between the 15th to 20th centuries to 25 art styles (abstract expressionism, baroque, cubism, impressionism, Fauvism, pop art, minimalism, etc.). (Jeon-hui Kim et. al.)
Just as GANs do, CANs consist of the two neural sections: discriminator and generator. These two neural networks are given two tasks respectively. First of all, the discriminator’s neural network classifies images that are provided to images of human art such as those trained with existing learning data and to images created by the generator’s neural network. The discriminator’s neural network of CANs performs one more task than GANs: it performs art style classification to determine which of the 25 art styles of the training data an entered image belongs to.
The generator’s neural network as well performs two tasks: One is to create new images. Since the generator’s neural network cannot access training data, it is unable to create new images by transforming or mixing images of training data directly. Rather, it merely receives feedbacks from the discriminator’s neural network regarding whether the image it creates is classified as an artistic work or it fails to deceive the discriminator.
The other task of the generator’s neural network is to prevent its own created images from being classified by the discriminator’s neural network as one of the existing 25 art styles. Specifically, it generates styles that are totally different from existing ones for style ambiguity to deceive the discriminator. In reflection of the discriminator’s feedbacks regarding how similar it is to any existing style, the generator focuses on increasing the sophistication of ambiguity. The competition of these two neural networks contributes to generating images that are similar to artistic works of humans and different from existing human art styles at the same time. The CAN neural network secures the probabilistic distribution of AI pictures to be similar to that of existing pictures so that people would not become aware of the difference. This mechanism of CANs to imitate existing artists’ techniques and to generate works of a new style is evaluated positively in the case of abstract paintings.
AI Art: technology or art?
There are differing views among experts on the artistry of works created by AI. Some express jubilation saying that just as humans see, listen, experience, learn, and get inspired from their childhood, now is the time for AI to prove its originality. Others argue that AI art is merely a result of training with learning data selected by humans and that thus it cannot be accepted as original.
In comprehensive view of many experts regarding the issue ‘Is AI Art art?’, there are several questions to be answered: The first one is about the validity of this question itself. While excellent AI algorithms such as GAN and CAN continue to emerge, AI is an algorithm in nature. The core of an algorithm is mimesis (imitation). In this perspective, the question ‘Is AI Art art?’ is a fake question that misreads the daily linguistic grammar of humans. In other words, it may be a combination of words but it cannot be a correct question. ‘Even if a lion spoke, we would not understand what it meant,’ said Wittgenstein. Likewise, even if AI imitates the human grammar, it cannot be a form of communication just like that of humans.
The second question is, ‘Can mimesis be the essence of art?’ Deleuze stated that artistic creation is to embody pure senses on a painting and that to do it, the following three conditions need to be met: Executing a distortion that goes beyond the scope of existing knowledge. Being able to utilize incidental elements and the lines of inorganic lives. Finally, physical senses and intuitive, intellectual thinking abilities to appreciate and judge them. At present, however, AI cannot cope with coincidences or ‘regulate’ them properly just as humans do. Algorithms that cannot regular coincidences merely generate new images irregularly and stochastically and then receive feedbacks from the discriminator. Even if an algorithm that meets these conditions regarding coincidences someday, it is the destiny of algorithms to imitate humans. Unless art is redefined to cover the area of newness that is generated by a machine’s imitation, the AI’s ability of mimesis cannot reach the nature of art.
The next question is, ‘Is AI able to think as humans do?’ The answer to this question is, ‘Maybe or maybe not.’ No one can posit the possibility that there may be or may not be any singularity of AI that can think like humans or even superior to humans. However, any supposition about such singularity must be based on complete theories and verified experiments about human memory, consciousness, and mind. The complexity of human memory, consciousness, mind and cerebration is still an uncharted territory. For instance, it is still not clear whether the storage of human consciousness and memory is the synapses, brain cells, or instant connections of electric signals. Without clarifying the complete operation principle of the human brain, AI is a mere incomplete experiment imitating humans. If AI finally reaches the singularity someday in the future, could it really be an area of humanity? If machines exceed the boundary of humanity, does it not mean that machines finally reach the level beyond human understanding as if ants cannot understand finger movements of humans? As a matter of fact, ‘thinking like a human’ as well is of a ‘fake grammar’ just as Wittgenstein said earlier. The singularity of AI will be realized only when AI reaches the level not of humans but of ‘thinking like AI’ so it exceeds the human language.
The final question is, What is the ‘peculiarity of humans’ that machines cannot imitate (mimesis) but get frustrated (nemesis)? This question is about the relation between AI and humans after all. The peculiarity of humans that machines does not have is in emergence, consciousness (memory), and cooperation of humans. Many scholars believe that they can clarify the human originality and consciousness (memory) completely, but such elements may remain unknown for eternity. It is like, in the film <The Matrix>, ‘the One’ that is a set of bugs in the virtual space repeats the everlasting reincarnation. The only aspect that has been discovered regarding the human peculiarity is that humans evolve more creatively through cooperation. Could machines evolve through mutual cooperation? Could AI entities interact with each other and upgrade themselves? Unfortunately, this area still remains in the level of experiments. The question, ‘Can AI perform art?’, is directly connected to the question, ‘Could there be an AI entity that can cooperate with an anti-algorithm?’
Collaboration that describes human curiosity, originality, soul, and life
It is time for us to go beyond the level of admiring works of AI and establish a new relation between them and artistic activity of humans. Even if any AI engine achieves a significant advancement, it cannot replace curiosity, which is a unique area of human artists. Curiosity is a destiny of humans who are imperfect but need to make new challenges constantly. AI Art will evolve in a form of supporting and cooperating with the destiny of human artists. Just as photography has been established as a unique art genre, AI Art based on learning algorithms as well will be established as an art genre that expresses or imitates the human soul and life. People responded similarly when the camera was first invented. When the camera was invented in the 19th century, the belief that only painters can reproduce the world just as it looks was broken. At that time, many predicted that painters’ techniques would disappear since they could not catch up with those of photography. However, it was soon proved that the area of painters was different from that of photography. Painters concentrated on mental pictures that only humans could recognize, leading new trends of fine arts by resolving, combining, and destroying images. That was the very motivation of contemporary art. In this regard, AI Art is a new medium that refreshes and stimulates artists’ original ideas and helps their labor of creation. For example, if an AI painter learns from certain keywords and draws an abstract draft, a human painter may complete the work based on it.
We hope that artistic works created, not through competition, but through collaboration between AI artists and humans, continue to increase and become more and more diverse.
References
1. Ammed Elagmmal et al. CAN:Creative Adversarial Networks Generating “Art” by Learning About Style and Deviating from Style Norms, 8th International Conference on Computational Creativity(ICCC), Atlanta, GA, June, 2017
2. Jun-hee Kim & Jin-yup Kim, Artistic Creation in the Era of AI: Focusing on Deleuze’s Theory of Art, Art and Media, 2020, vol.19, no.2, pp. 81–112
3. Seol-min Park, ‘Could AI Conquer the Area of Art?’, Sisaweek, September 7, 2020
https://www.sisaweek.com/news/articleView.html?idxno=137339
4. Jun-hyeok Jang, New AI Technology — GAN: Self-learning AI, Samsung SDS, 2018 https://www.samsungsds.com/kr/insights/generative-adversarial-network-ai.html