Can AI Be Creative?
Can artificial intelligence, or AI, be creative? According to Margaret A. Boden, creativity can be defined as the “the ability to generate novel, and valuable, ideas.” Valuable, in this case, means interesting, useful, beautiful, and so on. As for novel, it can be defined as either P-creative (one that’s new to the person who made it in the first place) or H-creative (one that is P-creative and has never been seen in history before).
An example of an H-creative work is the AARON program, created by Harold Cohen, artist and art professor at Universeity of California. AARON is a program that can make images with “world-class” coloring, as described by its creator. Artwork made by AARON has also been hung in museums around the world. However, more often than not, computer models aim at producing P-creative models, such as creating music in the style of past musicians (Cope 2001, 2006).
So how is creativity possible? According to Boden (2004), novel ideas can be made through combination, exploration, or transformation. Combinational creativity is to produce unfamiliar works by combining familiar works. This is the main type of “creativity” that people think of as a definition.
Exploratory creativity, on the other hand, is when people move through space and explore it to find what’s there (such as unvisited locations). This “space” can also be defined as cultural space or “conceptual space.” People can also set potentials and limits of the space.
Transformational creativity is when that “space” is transformed by changing one of its dimension. It is the type that brings the most controversy, as it may take some time for people to adjust to the new style. Vincent Van Gogh is one example.
Contrary to most people’s opinion, combinational creativity is the most difficult for AI to model. Of course, it is easy to spew out an endless amount of novel combinations of known concepts; the difficulty lies in getting something valuable out of it. One important thing that AI doesn’t store is culture — or peoples’ memory and experiences in their lifetime.
Within a tight context, however, AI can produce very useful contexts. For example, this Emotion Recognition Game. To play, you simply have to recreate the emojis under the time limit.
For this game, 6000 images were trained using Teachable Machine, a web-based tool made by Google that creates machine for learning models. However, since this is only a beta-version, it is far from perfect. For example, initially, more than 10 emojis were planned to be used. However, due to difficulty in identifying changes in facial expressions, angles, and other inaccuracies, only easily detectable emoji were chosen to be part of the game. To include more emojis, a larger dataset would be needed.
GAN (Generative Adversarial Networks) is a class of machine learning frameworks, in which data can be created. It is made up of two deep networks, the generator, and the discriminator. By using real images and fake images, the discriminator learns how to make certain features make the images look real. Both networks work back and forth in order to improve themselves, until the generator is able to make images in which the discriminator can’t tell the difference.
For example, this video shows a dataset of around 1000–1100 images of dresses trained using GAN. In order to make the images as ‘real’ as possible, only those with white backgrounds, no people, and from the same angles were chosen, and generated using RunwayML.
Dresses from H&M, topshop, and Pinterest were chosen and those that did not fit the criteria were manually removed from the dataset. As a result, this is a compilation of all the generated images.
The aim of this website is to possibly inspire fashion designers to create new dresses based on the generated images. Therefore, it is a sort of interactive art. Interactive programs aren’t meant to be creative, but to produce aesthetically pleasing and interesting results. Rather than the quality of the images, the interaction between humans and computer (Boden in press) is primary. The following images are inspired by a few of the generated images.
AI has even made it into the runway at Paris Fashion Week 2020. In collaboration with artist Robbie Barrat, fashion house Acne Studios used AI to inspire its Men’s Fall/Winter 2020 collection. The collection was weird but unique, with fashion designers using the mixup of AI to their advantage. For example, the AI sometimes confused borders of clothes, which led to the creation of coats with curved opening in the front.
“It is amazing to see that artificial intelligence can be freeing as a creative tool,” stated Jonny Johansson, the creative director of Acne Studios.
The general notion of people is that “computers can’t be creative because it can only do what we tell them to do.” However, through the trained models listed above, it shows that computers can do what people enable it to do. That is, computers can make p-creative and h-creative works. However, the biggest thing missing in its dataset is ‘culture’ and ‘memory.’ Because value is also considered as a part of the definition of ‘creativity,’ identifying creativity is not just a scientific question, but a social one. Whether computers can be creative involve many philosophical questions, such as what is art and creativity, the nature of meaning, and whether computer’s work could ever be considered as art. Until these questions are answered, we can never know if computers can truly be creative.
