For the last two years, AI has been so ubiquitous that it has popped up virtually everywhere, in plenty of different contexts. An influx or torrent of AI-based tools has revolutionised all spheres of our lives, including art and creativity. As sometimes happens with revolutions, the change occurs abruptly yet step by step, generating disputes and controversies.
While artists such as Mario Klingemann, Anna Ridler, Memo Akten, Sophia Crespo, and Sougwen Chung, who are more familiar with new technologies, have already gone through the looking glass and made art with AI without any inhibitions, others are wondering how and where they should get started, lagging behind this brand-new revolution. This division is particularly blatant between those making post-art or critical art and those focusing on classic mediums or the so-called commercial artists, for whom visual art is a source of income – illustrators, graphic artists, and special effect artists.
The first group, which views the neural networks not only as a tool and medium but also as a source of inspiration, creates philosophical, often critical, works referring to the relation between people and machines, experimental pieces demonstrating the potential, challenges, and boundaries of AI. Truth obtained along those experiments and collected data stands above the decorum – aesthetically pleasing work of art upholding the traditional rules of composition and colour combination. The second group, which so far has been relying on the masterful technique and craftsmanship developed with the use of digital and analogue tools, comes to terms with the reality in which creators more proficient in the use of AI, capable of training models, designing their own neutral networks and manipulating them can create intriguing visual experiences.
Ongoing changes evoke uncertainty and frustration, both foreign to tech optimists. Who is the artist? What is art? What is creativity? Even more – in a world dominated by AI, is there a place for creators who refuse to participate in the race? How will AI change us, the artists? Will the hand precision be ultimately replaced with the precision of prompting and coding? However, critical voices – as always – are needed so that new technology is not just a tool for accelerating production but also a pretext for reflection on its impact on culture, society, and ourselves. The art market also has doubts about the condition of a work of art and the artist – AI-generated artworks (with some exceptions) are considered as less creative, thus less valuable, according to Merel Meijer from the Rotterdam School of Management, Erasmus University (RSM). This cautious strategy has its foundations – as the rules of the game are set in place, the discussion about copyrights is ongoing, and the practices based on fair use, the DSM directive and the AI Act are still taking shape. Also, people rarely want to invest large sums of money in pieces that might be subject to total devaluation (the case of millions of works of various merit uploaded to the blockchain network as NFTs).
Where are we? Where do we go? This article will not answer these questions, but give you a certain snapshot, an attempt to outline the most important issues and a bunch of inspiration for those who want to experiment with AI for the sake of making visual art.
Levels of creativity
But let’s start at the beginning. At this point, we, the people, are still an indispensable element of the creative process of a work of art subject to copyright. The basis of generative AI is machine learning – a process in which algorithms are trained on delivered data to predict outcomes and make more accurate decisions. A significant component of AI is natural language (NLP), which allows machines to understand and generate answers in a human language instead of code, for instance. Generative models, such as Chat GPT, do not “perceive” the world in a human way. They also lack consciousness and creativity, as described by Gary Davis in his now classic book Creativity is Forever (1981) – a set of co-existing complementary traits such as perceptiveness, openness, sense of humour, curiosity, willingness to take risks, energy, emotional intelligence, need for solitude, ethics, precision, awareness of one’s own creativity, fascination with complexity and ambiguity, independence, originality, artistic ability, and ability to fantasise. As of today, AI has only some of these qualities. As a result, it views the world in a distinct, one might say more abstract, manner. Generative artificial intelligence (GEN AI) makes content based on the pre-existing data patterns used to train it. Still, it neither copies it in a 1:1 ratio nor does it create new content out of “nothing”. In a way, this process could be compared to an artist strongly inspired by learned techniques, movements and styles to make something of their own – in the case of AI, prompt plays the role of “unique vision”. The generated result is then not an exact copy of the original, though it might actually look similar.
Models such as GANs (Generative Adversarial Networks) or DeepDream by Google generate content in various ways. GANs (Generative Adversarial Networks) are based on collaboration and contest between two neural networks – a generator and discriminator – that train the model collectively. The generator creates images while the discriminator evaluates their authenticity to differentiate them from true data. The process, reminiscent of the mimicry phenomenon in evolutionary biology, leads to the generation of images that reflect reality more and more closely. Whereas DeepDream, based on convolutional neural networks, reinforces certain patterns and details present in the images, leading to the creation of some surrealist, almost oneiric compositions. Just as the lens focuses the light, highlighting selected fragments of reality, DeepDream emphasises some elements in an image, exhibiting them exaggeratedly and unexpectedly. The result is a vision of the world that reminds one of hallucination – details are multiplied and distorted, creating fascinating yet unreal visual landscapes.
In the case of diffusion models, such as stable diffusion, which are often used on popular platforms such as LeonardoAI and MidJounrey, the generative process is based on splitting images from the database at the pixel level into “noise” and recreating them from scratch to fit the prompt. This recreation allows for the compilation of abstract structures into brand-new, coherent (or not) visual content.
AI generates images that can be both completely out of the box and surprisingly close to what we deem as aesthetic. This way, it creates new spaces in art – places at the intersection of human intuition and the non-linear logic of machines, offering opportunities to explore and redefine creativity.
Unbearable diffusion of authorship
Each of the aforementioned models works only if it is trained on the appropriate dataset, which plays a crucial role as far as the quality of AI-generated images is concerned. Currently, companies offering cloud solutions for generating images and videos train their models on gigantic collections of data scraped from the internet, often without explicit consent from the creators. For example, Stable Diffusion, DALL-E, and MidJourney, which are based largely on the LAION datasets. LAION consists of the image-caption pairs sourced with the method of web scraping and allows the models to generate images similar to those contained within the training sets as well as easily emulate the style and technique of specific artists. This evokes controversies because artists, who have spent years developing their unique styles, might see their works as being reduced to a function similar to an Instagram filter.
Therefore, questions have been raised about the creative input of prompt authors and the lack of legal protection for AI-generated images. Current legislation assumes that only humans, not algorithms, have economic and moral rights. Thus, AI-generated content that bears no trace of human intervention is not protected by copyright and must be labelled as “AI-generated”. What’s interesting, a prompt – the instruction or command given to AI tools – might be protected by copyright only if it demonstrates creativity through, e.g. a unique form or literary style like a poem or an essay. In practice, this means that legal protection in the age of generative artificial intelligence is more concerned with humans as initiators of a process rather than the results of an algorithm itself.
In 2024, to balance the development of AI, the EU introduced the AI Act, the first regulation on artificial intelligence that aims to harmonise the AI rules and increase trust and protection under copyrights. AI Act requires transparency of the data used to train the models and labelling any inauthentic content, such as deepfakes. This legislation is related to the DSM Directive, which intends to give creators a right to remuneration and the option to register their work as not intended for AI (e.g. by including a relevant clause on their website to prevent scraping). However, these regulations do not apply in the United States or Great Britain.
The situation is different when it comes to the models trained by the artists themselves, who use their works and materials from the public domain. However, training one’s own models, over which one has full control, is not available to all artists. The level of entry remains high because training requires time to learn and experiment as well as knowledge of code. The more we understand how the algorithms and neural networks work, the more precisely we can control the final visual result and its origin. Yet, as Agata Lankamer, the AI artist, points out, comfortable model training is also not possible without advanced equipment, which many artists might not be able to afford.
Art and aesthetics in the AI era
Similarly to computers, the internet or software, AI propels forward the democratisation of the art creation process. On the one hand, this raises concerns among artists, critics, and audiences who realise the complexity of evaluation criteria and the potential devaluation of the position of an artist, which has been developed over the years. On the other hand, it opens the door for those who had no chance to acquire any formal art education and yet have the ideas and desire to create. As observed by Ivona Tau, an AI artist and author of the book Machine Gaze (2023), pieces that are swiftly produced, easily accessible, and lack depth will lose their value in favour of original art and research projects. According to Tau, what we are seeing in the age of AI is the shift towards the increased conceptualisation of art, its “spiritualisation”, and critical reflection.
The still-emerging aesthetic of AI gives rise to the rebirth of surrealism – at times in the utterly cliched form as the byproduct of the democratisation of those tools. A significant characteristic of this aesthetic is the morphing of shapes, forms, textures, and styles and the presence of discontinuous surfaces, creating a visual dissonance. The process of generating images by AI also displays blurred, nondescript shapes, fragments disturbed by repetitive patterns or pixelisation. These disturbances, dubbed as glitches or AI hallucinations, are the result of the algorithm seeing patterns or objects that don’t exist. Though seemingly imperfect, these errors often serve as points of departure for the artists exploring the peculiar nature of AI.
Works by one of the pioneers of AI art, Mario Klingemann, such as Mistaken Identity and Memories of Passersby I (2018), probe into the boundaries between reality and fiction by using systematic distortions to create surrealist hybrid worlds. Klingemann refers to this aesthetic as neurealism – a blend of words “realism” and “neural networks”. The artist perceives neurealism not only in terms of an aesthetic but also as a critical outlook on the possibilities and limitations of AI.
A similar play with machine errors can be observed in Learning to See (2017-) by the Turkish artist Memo Akten, who explored how neural networks “learn” to see the world by highlighting how their perception differs from humans. Akten demonstrates how algorithms interpret reality, transforming it into abstract, even surrealist compositions that reveal the internal mechanisms of working machines. In the project Mythic Latent Glitches (2019), Ivona Tau takes a step further by combining human creativity with the potential of generative AI. Glitches, digital disturbances, and imperfections become the metaphor for human perception and difficulties in grasping complicated technologies such as AI. Tau embraces it as an artistic tool to conjure visions bordering on dream and reality inspired by Lithuanian landscapes and mythology.
Whereas Anna Ridler presents a reflection on history and capitalism in her project Mosaic Virus (2018), where AI creates surrealist depictions of tulips, the colours and shapes of which are subject to dynamic changes depending on the bitcoin exchange rate. In her work, Ridler refers to the 17th-century “Tulip Fever” – the very first speculative bubble in history – to depict correlations between old financial obsessions and contemporary cryptocurrency trading. Her project is not only aesthetically mesmerising but also critically engaging as it offers an analysis of the connection between technology, economy, and culture by raising questions about beauty, manipulation, and history. Together, all these works investigate the boundaries of contemporary generative art and encourage viewers to reflect on the relationship between human creativity and machine precision.
Wasteland of possibilities
Although the development of AI and its application in art opens up a wealth of possibilities, it is also accompanied by some fundamental questions about ethics and ecology. Good equipment, custom model training or fine-tuning (feeding one’s own data to existing models) pose dilemmas about their environmental impact.
Generating a 100-word text by ChatGPT uses approximately 500 ml of water, as stated in the widely discussed article Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models (2023). Training data centres, which are now popping up everywhere, are often located in semi-desert terrains, where access to natural resources is already limited. Those large facilities consume huge amounts of energy and water for cooling the servers. The cost of this operation is often shouldered by the local community. According to the authors of the article published in The Atlantic, these practices raise the question of sustainable development in the age of artificial intelligence. Ivona Tau emphasises that training small models by artists has a minimal impact on the environment, and the responsibility for climate degradation belongs to the corporations that train large models on an unprecedented scale. Nonetheless, artists can remind people of this responsibility and exert pressure on tech companies to act more sustainably.
Is AI art more harmful to the environment than traditional techniques then? Probably yes, even as far as the scale and accessibility of AI solutions are concerned. While, as a society, we subscribe to the notion of underconsumption, more and more frequently, the consideration of sustainability in art – also the “traditional” painting – is still far from widespread. Both physical artworks and large-scale interactive installations, hugely popular in recent years, might seem superfluous, which is noted also by artists and curators (for instance, creators of the ongoing exhibition New Ecosystems of Art address critically their own practices and art they create and wonder if it could be zero waste).
Ethical issues also pertain to the global supply chains in technology. Training AI models requires preparation, labelling, and moderating enormous amounts of data, which is often done by people from African countries who work the minimum wage in harsh conditions, as detailed in the article by Time magazine. Another point of contention is feeding models data that fail to account for diversity and inclusivity since they are based on stereotypes. Malik Afegbua, an artist and designer from Kenya, protests this issue in his work The Elder Series (2019). Activists around the world are also taking notice, as evident from the attempts at creating inclusive databases that do not centre on the European or American experience (e.g. Open for Good Alliance).
Despite these challenges, AI may offer unique opportunities for creators all over the world who felt alienated up until this point. For instance, women, who often play the role of caretakers, can use AI tools to reduce the time spent on developing their ideas and use them as a kind of accelerator of creativity. Similarly to conceptual or digital art of the 1960s, generative art has not yet been dominated by a single group of artists, giving women a better chance of breaking into the mainstream. This provides a significant counterpoint to traditional disciplines, such as painting and sculpture, where women’s work is still sometimes pigeonholed as feminist art and reduced to subjects related to femininity, motherhood, or emancipation.
New horizons
Art and creative practice are subjective. We, the people, decide the rules of a game and assign art value. Art is also a tool for examining reality, an extremely sensitive to social change technological barometer, but also an analytical and speculative tool. Artists’ experiments with AI tools give us some insight into the way the art market might look in one or two decades. On the other hand, we seem to have forgotten about the wave of NFTs that swept over us right before the AI tsunami. Will this [r]evolution share the same fate? Or will AI restore the value of classical art, those “handmade” pieces, physical, imperfect, tactile artefacts – paintings, drawings, pastels, or sculpture, as I speculated in my article Predicting the Future in Visual Arts (2019)?
Perhaps this is the direction we will go as artists rely less and less on honing their craft in the traditional sense and more on prompting and quick prototyping of their visions. However, as evident from the examples mentioned in this article, art made with the use of generative AI does have a cognitive value. It facilitates the exploration of our innermost thoughts, dreams, and fears by translating them into surrealist, at times tacky, images. What is more, like the impenetrable ocean on the planet Solaris, it reveals challenges and opportunities in the relationship between a man and a machine, forcing us to consider the future of creativity.
Article was written as part of the grant programme under the title [A] inspiracje. Nowe horyzonty wizualności w erze sztucznej inteligencji ([A]inspirations. New horizons of visuality in the era of artificial intelligence) implemented by Sylwia Żółkiewska within the framework of KPO dla kultury.
The goal of the article and project is to examine the subject of AI and art and encourage artists to use the tools based on generative AI – spreading knowledge about useful tools, methods, and techniques of working with AI as well as ethical and ecological challenges.
Quotes by the artists Agata Lankamer and Ivona Tau were sourced from the discussion led by Sylwia which is available on YouTube: [A]inspiracje. Nowe horyzonty wizualności w erze sztucznej inteligencji: https://bit.ly/AI-art-discussion-2024.
