Page 275 - The Black Box Book

The last type, transformative creativity, is characterized by tran-
scending the boundaries of the previous conceptual space and developing
a new structure of thinking. For a long time, this type of creativity could
not be mastered by computers. But this changed, according to Boden,
when developers such as Karl Sims and Wiliam Latham experimented Even though, we can list some examples of computer scientists who use
with so-called generative algorithms in the early 1990s, which were based extremely reductionist models of creativity (see Langley et al., 1987),
on the principle of evolutionary programming and in this way generated still the latter accusation seems a bit sweeping. Back in the 1980s, for
computer graphics of astonishing complexity (Sims, 1991, see also: Todd example, computer scientist Marvin Minsky criticized that people tend to
& Latham, 1992; Bentley, 1999). exaggerate the difference between ordinary and outstanding achievements
Meanwhile, ANNs have outdone other methods that have been in terms of their creative thinking (1982). At least in this case, there can
considered as a path toward the goal of creative AI in recent years and be no question of a cult of genius.
decades. For the time being, at least. Nonetheless, the creative potentials In general, a fundamental problem with the debate on AI is that
of ANNs are also doubted by various philosophers, artists, and humani- we typically value and admire those achievements of machines that we
ties scholars: For example, in 2016 Matteo Pasquinelli argued that ANNs also value and admire related to humans. If a computer beats the human
are fundamentally incapable of producing anything emphatically new. world champion in the board game Go, it is a sensation that is reported
According to him, ANNs would be able to recognize patterns in already globally in the media. However, there was much less public interest in
existing data purely inductively, but they would not be capable of what 2017 when an AI model was able, for the first time, to recognize actions
Peirce calls the operation of abduction. Hence, ANNs would be unable to and gestures using ANN-based methods and based on training with video
achieve, what Margaret Boden called transformative AI. data, i.e. when it was able to master, in a basic form, what is commonly
Also, the media philosopher Dieter Mersch recently reproached referred to as common-sense knowledge in relation to humans (Sudmann,
those who, like Margaret Boden, are willing to ascribe creativity to com- 2016, 2017). It was not acknowledged that the step of enabling an AI to
puters for mostly using simplistic and historically outdated definitions of visually perceive and recognize its environment is much more fundamental
art and culture as a basis and at the same time consistently disregarding and far-reaching – not least for the development of advanced systems
a specifically “epistemological dimension of the aesthetic”: that will eventually be capable of human-like creative behavior – than
the well-known high-profile developments like AlphaGo & Co.
“Persistently they ignore what art makes art in the first place:
Reflexivity as the opening up of a different knowledge. Instead, GENERATIVE ADVERSARIAL NETWORKS
under the sign of a preference for rationalism and hard sciences,
a direct connection is drawn between ‘natural’ creative activities One may also be skeptical whether Generative Adversarial Networks
such as the development of life and the ‘social’ or ‘ historical’ vir- (GANs) or their more developed successors as Creative Adversarial Net-
ulence of the arts, regardless of essential incompatibilities. If, on works (CANs) will raise the performance of an AI in terms of creativity to
the other hand, we insist on the persistence of an aesthetic reflexivity a new level in the medium or long term. CANs essentially copy the function-
as a criterion for art, it is because this reflexivity is not drawn from ality of GANs but add a decisive component to the algorithmic operations:
perceptual extensions, subjective expressive excesses, or the rep- The discriminator network still has the task of classifying images according
resentation of borderline phenomena, just as little as the properties to whether they are real or fake, but also learns to assign these artifacts
of art can be inferred exclusively through its ‘works’; rather, these to 25 art styles (like impressionism or surrealism, etc.). Accordingly,
always ‘acquire’ a transformation and displacement of art itself. Art the generative network still has the function to generate images that
is always art about art; it therefore implies in every act and artifact the discriminator considers to be real, but in addition to that it is supposed
a transformation of the aesthetic itself, whereas most models of an to make the assignment to these art styles impossible. The performance
artificial creativity are oriented to continuities and remain attached of these creative algorithms thus consists in a fixed scheme of deviation
to an anachronistic nineteenth-century cult of genius, which in turn from a norm as an iterative optimization process. Hence, CANs deviate
owes itself to a vulgarization of the Kantian definition of genius” decisively from what characterizes deviation from established art styles
(Mersch, 2019, p. 73). with regard to human creativity, namely that as an artistic operation it
is itself always up for disposition in relation to a historical norm, i.e. it is
uncertain (Elgammal, 2017).
As a result, the computer art produced by CANs may even be more
interesting than many works by human artists, and this would be solely
274 275