After a spring and summer of fielding questions about my role in the Art & AI Laboratory at Rutgers University, and working with computer scientists on issues of computer vision and machine learning, I though I would post some thoughts on issues that have repeatedly been raised.
I continue to be happy to talk with artists who are interested in this topic, and those from the AI and machine learning community!
I'll post in parts, addressing some of the key issues as I see them:
1. AI created art is not the same as Human created art: there are some similarities in process (learning about styles, for example) but also differences
2. AI created art is not about replacing Human artists or threatening their turf (many human artists make art based emotion, or psychological drives, or political/social intention: these are not relevant to AI)
Humans and AI do not share all of the same sources of inspiration or intentions for art making, nor should they
The CAN study provoked a series of concerns about AI as a threat or rival to art made by human beings. Yes, the study is interested in the process of art creation, and the more abstract problem of what creativity is and does. However, AI is premised on better understanding machine process and machine creativity. This does not mean simply copying or hijacking human means of making art and trying to pass off AI-generated products as the same. In fact, one of the interesting facets of this study was NOT ONLY the comparison of the CAN process’s works in comparison/contrast to other human examples in order to ascertain acceptance of the works as aesthetically recognizable as art, but also to determine that human viewers liked and even sometimes preferred the AI-generated works of art. This is not to prove that the AI is better than human creators, but instead that the AI can produce a work that is able to qualify, or count, as art and that it has qualities that make it desirable or pleasurable to look at.
Many of the contemporary artists who are asked about the study’s works resist seeing them as art because they have a definition of art that is based on modern, individual-centric definitions of creativity. As an art historian, I would point out that that is a relatively recent and culturally- specific conception. For many centuries, across many cultures and belief systems, art has been made and will be made for a variety of reasons under a wide range of conditions. Frequently created by groups of people rather than an individual creator or artist (think medieval cathedrals or guild workshops), made to the specifications of patrons and donors large and small, made to order, funded by a wide variety of groups, civic organizations, or religious institutions, and made to function in an extraordinary range of situations. The notion of a painting being the coherent expression of the individual’s psyche, emotional condition, or expressive point of view begins in the Romantic era, thus is a defining norm predominant in the 19th and 20th centuries in Western Europe and its colonies. Although this remains a common motivation for many artists working today, it does not mean it is the only correct definition of art, or that it will prevail in the future. Clearly machine learning and AI cannot replicate the lived experience of a human being, nor is this being proposed in the study. How the machine makes art is intrinsically different, it has to be. But we are asking everyone to consider that a different how, or different process of making, does not disqualify the creations of the process as art.
AI is not able to create art in the same way that individual human artists do, nor is it trying to do so. Instead the study is ultimately focused on understanding the process of creativity such that a means can be found to model that process or system that generates a creative result. One way to do this, and what this study has chosen, is to model the process by which art is taught, and then how a creator can be encouraged to synthesize that information and next create something new. To do this the machine was trained on many thousands of human-created paintings (paintings are easily readable by the machine in digital version) by a process not unlike a human artists’ experience of looking at other artists’ works, learning by example. The CAN system was then designed to encourage choices that deviate from copying/repeating what had been seen to encourage new combinations and new choices.
It is the process that is under study here, not the production of human-like art. If the process is modeled successfully, art will result. An important barometer of whether art has been successfully created through the chosen process is whether human beings can recognize it as art, and choose it as pleasing. That is what the CAN study is about.
I continue to be happy to talk with artists who are interested in this topic, and those from the AI and machine learning community!
I'll post in parts, addressing some of the key issues as I see them:
1. AI created art is not the same as Human created art: there are some similarities in process (learning about styles, for example) but also differences
2. AI created art is not about replacing Human artists or threatening their turf (many human artists make art based emotion, or psychological drives, or political/social intention: these are not relevant to AI)
Humans and AI do not share all of the same sources of inspiration or intentions for art making, nor should they
The CAN study provoked a series of concerns about AI as a threat or rival to art made by human beings. Yes, the study is interested in the process of art creation, and the more abstract problem of what creativity is and does. However, AI is premised on better understanding machine process and machine creativity. This does not mean simply copying or hijacking human means of making art and trying to pass off AI-generated products as the same. In fact, one of the interesting facets of this study was NOT ONLY the comparison of the CAN process’s works in comparison/contrast to other human examples in order to ascertain acceptance of the works as aesthetically recognizable as art, but also to determine that human viewers liked and even sometimes preferred the AI-generated works of art. This is not to prove that the AI is better than human creators, but instead that the AI can produce a work that is able to qualify, or count, as art and that it has qualities that make it desirable or pleasurable to look at.
Many of the contemporary artists who are asked about the study’s works resist seeing them as art because they have a definition of art that is based on modern, individual-centric definitions of creativity. As an art historian, I would point out that that is a relatively recent and culturally- specific conception. For many centuries, across many cultures and belief systems, art has been made and will be made for a variety of reasons under a wide range of conditions. Frequently created by groups of people rather than an individual creator or artist (think medieval cathedrals or guild workshops), made to the specifications of patrons and donors large and small, made to order, funded by a wide variety of groups, civic organizations, or religious institutions, and made to function in an extraordinary range of situations. The notion of a painting being the coherent expression of the individual’s psyche, emotional condition, or expressive point of view begins in the Romantic era, thus is a defining norm predominant in the 19th and 20th centuries in Western Europe and its colonies. Although this remains a common motivation for many artists working today, it does not mean it is the only correct definition of art, or that it will prevail in the future. Clearly machine learning and AI cannot replicate the lived experience of a human being, nor is this being proposed in the study. How the machine makes art is intrinsically different, it has to be. But we are asking everyone to consider that a different how, or different process of making, does not disqualify the creations of the process as art.
AI is not able to create art in the same way that individual human artists do, nor is it trying to do so. Instead the study is ultimately focused on understanding the process of creativity such that a means can be found to model that process or system that generates a creative result. One way to do this, and what this study has chosen, is to model the process by which art is taught, and then how a creator can be encouraged to synthesize that information and next create something new. To do this the machine was trained on many thousands of human-created paintings (paintings are easily readable by the machine in digital version) by a process not unlike a human artists’ experience of looking at other artists’ works, learning by example. The CAN system was then designed to encourage choices that deviate from copying/repeating what had been seen to encourage new combinations and new choices.
It is the process that is under study here, not the production of human-like art. If the process is modeled successfully, art will result. An important barometer of whether art has been successfully created through the chosen process is whether human beings can recognize it as art, and choose it as pleasing. That is what the CAN study is about.