Art, Artist and Creativity
Is photography art? The answer to this well-debated question is relevant to the concept of AI as an artistic tool. At first, photography wasn’t perceived as an art form, because it was a mechanical process. Some saw it as a threat and argued against its legitimacy. Some artists embraced the new technology and began to explore its potential. As the new technology improved and spread, artists learned to control and use it. The controversy over its status was over. Photography made art-making more accessible to everyone. Today, we all can experiment with photography. Furthermore, the new technology gave the old art form new life. Leading it toward greater abstraction. AI may follow a similar trajectory as photography did. For now, without creating much controversy, AI can be a tool used by an artist to create art.
Is it possible to call AI an artist, not just a tool? Some people argue that no matter how skilful and surprising a computer’s output is, it won’t be as an artist until it manifests its personality, conscious intention or desire to express. For others, the ability to learn and generate novel and valuable ideas is enough.
The dependency between an artist and the AI system can take many different forms. Autonomy in the computational arts is a spectrum, not a binary feature. The human artist can choose the type of data used to train the model and the training parameters which influence the work. In the end, there can be a human judgment that will evaluate the final result. But it is also true that we can make the algorithm work with zero or minimal human intervention. Good examples are AARON (aaronshome .com) or the Painting Fool (www.thepaintingfool.com). They are algorithms that produce art in complete autonomy, including the evaluation phase.
The most recent works on creative AI use Generative Adversarial Networks and Creative Adversarial Networks. In the last case, the authors report that there is no human intervention during the creative process. However, the model was still trained on existing human creative works.
There has been a considerable amount of media hype around AI techniques. In the news, algorithms are often anthropomorphized and sometimes they described as artists. In fact, we don’t know what consciousness is, or what it would mean to embody it in an algorithm. If we accept that contemporary AI is not conscious and social, can we still call it creative?
The argument against the ability of AI to show creativity boils down to a statement that an algorithm can only do what it’s programmed to do. In that case, it can’t do anything new. Any product of the algorithm is a demonstration of the programmer’s creativity, not the machine’s. But, with the rise of machine learning, AI can now learn certain skills. In one of the most successful methods today, we provide the computer with lots of data. Humans learn almost the same way. We repeatedly try things and based on whether we are correct or incorrect, we adjust our approach until we reach a certain level of competence. So even though our methods of learning are quite similar, when computers do unique things, we often find ways to avoid calling it ‘creative’.
From some point of view, humans and AI don’t differ that much. Can we say that humans are not creative because any product of a human mind is solely a demonstration of nature’s creativity? We both need input data for our creative processes. Humans collect information with senses while AI does it with an input layer. Humans creativity is limited by the natural environment and our means of exploration. AI is limited by the data it is fed with and its architecture.
In 2015 Mordvintsev and his colleagues invented DeepDreams, an algorithm capable to produce hallucinatory imagery that often looks like pieces taken from a psychedelic-acid trip. Its aesthetic value was appreciated by Foster the People who used it in their music video for the song “Doing It for the Money”.
Another development which received considerable attention in 2016 was the invention of Neural Style Transfer (Gatys et al. 2016). This technique captures the content of one image and combines it with the style of another image.
Another group of researchers (Elgammal et al. 2017) proposed a system for generating an art called Creative Adversarial Networks. This approach builds on an already existing technique called Generative Adversarial Networks (Goodfellow et al. 2014) but adds a component that encourages the algorithm to act creatively. The algorithm is trained between two opposing forces. One that urges it to follow the aesthetics of the art is shown, while the other force penalizes it for emulating an already established style. The two opposing forces ensure that the art generated will be novel but at the same time will not depart too much from acceptable aesthetics standards. The original CAN model was trained with 80K images representing 5 centuries of Western art history, simulating the process of how an artist digests art history, with no special selection of genres or styles.
Researchers compared the response of human subjects to art created by human artists and CAN. The results showed that people could not distinguish between them. To the researchers’ surprise, the public ranked the automated pieces higher than the images made by humans.
The creativity of AI is also proven in image inpainting task (Yeh et al. 2017). Its objective is to automatically reconstruct missing or damaged parts of an image. Generative Adversarial Network is able to predict semantic information in the missing part and to automatically replace it with meaningful content. For instance, if an eye is missing from the image the neural network is able to correctly generate it and correctly place it.
At the end of last year, NVIDIA researchers (Karras et al. 2018) proposed a hyper-realistic image generator built on top of GAN and style transfer (thispersondoesnotexist.com). The system learns how to transparently merge styles of input images into something completely new. It became famous for generating faces but it can do the same with other types of objects like cats, cars or bedrooms.
All of this would be unnecessary if the work done by AI had no value for us. If the price is an indicator of the quality of art then how much is AI art worth?. In February 2016, a work made by the computational artist Memo Akten using DeepDreams was sold at a benefit auction for 8,000 USD. In October 2018, as reported in the auction house website (Christies 2018), Christies becomes the first auction house to offer a work of art created by an algorithm. The work was estimated between 7,000 and 10,000 USD but it sold for 432,500 USD, almost 45 times its highest estimate, questioning the experts if we are witnessing the birth of a new art market.
Text generation is another area in which AI demonstrates its creativity. Researchers from OpenAI (Radford et al. 2019) trained a language model capable to generate coherent paragraphs of text achieving state-of-the-art performance. Their model was simply trained to predict the next word in 40GB of Internet text but, according to the authors, the results were so good that they decided not to release it publicly due to concerns of malicious applications. The examples provided by the authors are impressive, although some claim that this could be the case of cherry-picking.
The beginning of AI-generated story: In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English. The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science.
AI was used by the filmmaker Oscar Sharp to generate a script for a short science fiction movie called Sunspring. The plot turned out to be telling a story about three people living in a future world and eventually connecting with each other through a love triangle. The movie was presented at film challenge contest of Sci-Fi-London 2016 and was nominated among the final top ten films.
In 2018 The New York Times reported that the novelist Robin Sloan is writing his book with the help of AI assistant that finishes his sentences with the push of a tab key. Sloan types the beginning of a sentence, “The bison have been travelling for two years back and forth,” and the program finishes, “between the main range of the city”. It is also reported that combined efforts of scientists from English and Spanish universities led to the writing of musical plot. The team of researchers and producers started with a project called What-If-Machine (WHIM) to generate multiple central premises and key characters. Then they used a storytelling computer system called PropperWryter to build core narrative arc of the show. As a result of human-AI cooperation, a musical called Beyond the Fence was created. In 2016 it was running for a week at London’s Arts Theater.
Among various fields of creativity AI has also some achievements in music composition. In 2016 an algorithm called AIVA (Artificial Intelligence Virtual Artist) became the world’s first virtual composer. The company behind AIVA offers it as a commercial product (aiva.ia) capable to generate music in a specified style. Even though the algorithm is not a real composer, its music is recognized by French music society. Till now AIVA published two albums: “Genesis” released in 2016 and “Among the Stars” released in 2018.
Manifestations of AI creativity are also visible outside the areas typical for artists. In 2015 a group of researchers (Cully at al. 2015) worked with robots to train them to walk in conditions in which their legs had been damaged. The objective was to maximize the robot’s speed while minimizing its contact with the ground. To the researchers’ surprise, the AI calculated that the robot could walk with 0% contact of the feet with the ground. When they ran the visualisation, they discovered the robot that flipped over on its back and crawled on its elbows with its feet in the air, thus reducing the contact of the feet to 0. If this task was given to a human, it seems unlikely that they would consider a robot walking without using its feet.
The recent victory of Google’s AlphaGo also gives some evidence of AI creativity. World’s best go players commented on the moves made by the AI saying:
When I saw AlphaGo’s moves, I wondered whether the Go moves I have known were the right ones.
It’s not a human move. I’ve never seen a human play this move. So beautiful.
Like a good piece of art, the AI’s behaviour caused a pleasant sensation of admiring something wonderful. The AI’s victory is even more deserved because it was trained solely via self-play.
The contemporary AI proves to be able to create objects that we humans call art, unique creations that evoke emotional reactions and have both aesthetic and real value. It also provides a unique solution or behaves in an original way in areas not related to the world or art. Research shows that people may rank art created by AI higher than the art created by human artists and pay significant amounts of money for AI art. It is not clear whether AI can be called an artist but for sure it’s capable to learn, be creative and surprise. It can assist humans in their creative work or act creatively on their own. We should not forget that AI is a different kind of entity than the human being. It doesn’t experience the world like we do. It has different possibilities and limitations. Contemporary computers are good at computations, doing narrow and well-described tasks even better than humans. They are not good in human relations, emotional behaviour, being social.
- Antonio Daniele, Yi-Zhe Song, AI + Art = Human, Association for the Advancement of Artificial Intelligence, 2019.
- Yeh, R. A., Chen, C., Yian Lim, T., Schwing, A. G., Hasegawa-Johnson, M., & Do, M. N. (2017). Semantic image inpainting with deep generative models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5485-5493).
- Multiple authors, AI in the media and creative industries, New European Media, 2019.
- Mazzone, M., & Elgammal, A. (2019, March). Art, creativity, and the potential of Artificial Intelligence. In Arts (Vol. 8, No. 1, p. 26). Multidisciplinary Digital Publishing Institute.
- Hertzmann, A. (2018, June). Can computers create art?. In Arts (Vol. 7, No. 2, p. 18). Multidisciplinary Digital Publishing Institute.
- Elgammal, A., Liu, B., Elhoseiny, M., & Mazzone, M. (2017). CAN: Creative adversarial networks, generating” art” by learning about styles and deviating from style norms.
- Cully, A., Clune, J., Tarapore, D., & Mouret, J. B. (2015). Robots that can adapt like animals. Nature, 521(7553), 503.