We can all agree on the observation that Artificial Intelligence is at the center of most debates today, which makes it a very topical subject. Art, on the other side, has been around for ages, and is often considered a historical notion. If you expect an AI expert advice, or an art historian view on the matter, look away. All you will get is an AI artist view on the subject, which is a naive view, but an informed one.
“This discovery will eliminate the inferior layers of art”
“The artist runs the risk of becoming a machine, hitched to another machine”
“Painting is dead”
To establish some context, all these words have been heard around the year 1850. And they all point at a tool we are quite familiar with: the camera.
Back in those days, although it was a subject that just a few people understood, it was a very popular one.
It was a subject that made most people dream, but as we observed just before, also a subject that could be scary.
See where I’m going ?
Defining Artificial Intelligence
When we talk about AI, it is important to properly state the stage. Let’s make it simple. That’s what AI is: the manufacture of intelligent systems.
Not happy with that answer ? Well, the point is, we are pretty clear on what AI is, whereas intelligence itself is tricky to define properly. For example, a task that would consist in recognizing a cat from a dog. It would require to identify which feature is common to each subject, build general rules out of examples, and manage to compare. Now, take a task that would require to solve a deeply theoretical math problem that applies to make reality better in the long run. Making the choice to concentrate on this problem instead of feeding your reward circuit would require another level of intelligence. In the end, AI aims at building systems that answer all these problems.
A good way to define AI could be to state what we are currently able to do. Today, using machine learning, we manage to replicate simple human tasks with algorithms (a set of instructions based on statistical methods and data). We use these algorithms either to optimize (by performing complex computations), or to scale (with chatbots for exemple), even though those two notions are often linked one to another. The algorithms can learn by identifying common features in the data, and are faster than humans at analyzing it. We can call this Augmented Intelligence. Quite far from the AI that we like to fantasize, and the idea that resonate when startups, governments and companies flood us with their constant innovation promotion.
Working at Obvious involves a close relationship with AI and its possibilities. Throughout the way, it raised some questions such as “Can an Artificial Intelligence be independent ? Can an AI be creative ? Can an AI be an artist, and by procuration, to what extent will human tasks be performed by AIs ?”
Art: a perfect land of experiments
We decided to explore these interrogations through art. Why ? Art is a perfect medium that allows to experiment with the possibilities of an AI and better understand how it all works. Here are four art features we identified as being helpful in our research.
Art is tangible: it offers some concrete results.
It is accessible : most people have an affinity with some kind of art.
Art is interpretable: it offers another way to experiment, and leads to debates that are at least as interesting as the answers you can get the purely scientific field
Art is free, and it cannot be restrained by our own creativity when experiencing with it.
Therefore, art seemed like the perfect way to experiment with creativity as expressed by an AI as well as its limits.
When talking about art, we consider all types of applications that we start to see appearing around us. Music made by a collaboration between human and machines, poems, scripts, lyrics, trailers, and images made by algorithms. All these projects have in common the replacement of part of the creativity process. Each one of them is different in the level of human intervention it involves. We can say that once the whole process will have been automated, we will have created a machine that is capable of being creative, in the same way a human is.
Let us focus on that creativity process. In order to better understand it, we need to imagine ourselves as an algorithm, to which it is asked to create a picture from scratch. This exemple is extendable to the creation of a poem, a music, or any other type of creative art.
You are in the dark. No sense of sight, hearing or touch. The most likely output is the following.
Now imagine that you gain access to sight. This is done using computer vision, which allows a machine to analyze a picture pixel by pixel, and pull out some forms and colors. This is at the basis of the learning process, as it gives the algorithm access to data which is what we would assimilate as food.
Now, a picture is shown to us. From this input, the most likely output is that very same image, as it represents the only reference the algorithm has.
Now if you show more than one picture to this algorithm. It can start to learn on common features, and will likely rend an image that resemble the one depicted below.
In the same manner, Generative Adversarial Networks (GANs) analyze tens of thousands of images, learn from their features, and are trained with the aim to create new images that are undistinguishable from the original data source. These models are able to outcast any image that is not relevant (i.e. which doesn’t have enough features in common with the others).
They also reproduce the notion of novelty. Even with the same inputs, the algorithm will each time render a different result. This reflects a human creativity feature: We will never create twice the same thing, as it is impossible to ask someone to create something, twice, at the same time. The ponderation of each influence factor will have changed between those two moments.
Now let’s make things interesting. Imagine it is asked to create a picture that is beautiful. Beauty is a subjective value, and there is no right or wrong answer to this. But there is a statistically optimal one. An option would be to put labels (meta-data) on the pictures that serve as input (or food). If you can tell me which images have been enjoyed the most, I can accentuate my training on these pictures, and create an image that is closer to those.
We are now entering a highly theoretical field. If you ask me to create something useful. Iwould have to understand the many nuances of the effect a picture can produce. Bringing back memories, passing through a message, father laughter or sadness. We can imagine facing that difficulty with labels, but we would need many very specific ones, and a huge amount of information on those pictures.
Finally, if you ask me to create something that reflects my personality. I would have to develop and identify as being mine specific personality traits, and correlate those features to graphic content, in order to reflect it and gain the expected effect on the viewer. We are very far from being able to do this.
GANs are not the only models to mess with creativity, but they are attracting attention from the scientific community as they are able to model high-dimensional data distributions.
As you can see, creativity can be divided in steps, which are increasingly complex. As for now, we are not aware of all the steps on the path, and we are even less able to replicate them with algorithms.
Nevertheless, we are able to create unique pictures, that are new and resemblant to real ones. This discovery is likely to open a wide range of business opportunities.
What does the future hold ?
In order to give a shot at predicting the consequences of these models, we can look back on what the appearance of photography impacted the art market and society.
Photography automated the reproduction process. It had a great impact on professions such as copyist artist, which almost disappeared from the surface of earth. On the other side, it saw a new generation of creatives: artist photographers. You only have to open your Instagram to witness the burst in creativity that the democratization of this tool led to.
Photography also adopted the norms of classical art, and by doing so made its way into the art market.
We can imagine that if AI succeeds at doing the same thing, AI artists will become very common in the near future. And they are already appearing throughout the world. At the moment, managing those tools requires rare and expensive skills, but we can imagine that tomorrow these skills will be widely spread across the population, and that softwares will be developed to facilitate access to it.
Now, if you ask me:
“Will AI be the artist of tomorrow”,
I would be tempted to answer:
“Is the camera the artist of today ?”.
It is not. AI is a new tool, allowing the maximization of the creative potential of humans. Nevertheless, for the first time, humans also have the possibility of maximizing the creative potential of their tool.
With AI, parts of the creative process are getting automated. And those parts are the least complex ones.
New creative tasks are appearing, and creatives will soon have a new tool, a algorithmic pencil, brush, drum pad, allowing them to have a creative machine in their hands and gain in efficiency.
Science searches for the truth, not for what we want to hear. It is the applications we make of these technologies that will shape the future. Those applications need to be made with good will. Now, if we are looking for an answer to the naive question:
“Will AIs replace humans ?”
An educated answer would be that with the right amount of good will, AIs will replace humans of today, in order to allow humans of tomorrow to thrive by making them a little bit more like artists, in the way they choose to be.
Obvious, isn’t it ?