Page 346 - The Black Box Book

A NIGHTMARE OF IMAGINATION

Artificial intelligence is being experimented with in all areas of culture, by
artists, developers, designers, scientists, and everyday users. Frequently,
the objective is to produce content that is equivalent to or even more
compelling than that of a talented human author. Individual models of spe-
cialized learning algorithms also combine and are interlinked. In a fraction
of the time and with the appearance of autonomy, content is generated in
staggering amounts. However, these algorithms still require a great deal
of human aid in the form of decisions, and should therefore be viewed
more as another type of synthetic paintbrush or a more sophisticated
camera. What potential does learning algorithm technology provide for
the growth of artistic imagination?
Developers and enthusiasts claim that the Midjourney synthetic

generator transforms imagination and dreams into breathtaking works
of art without limitations. All the user has to do is enter a text command
in the form of a string of words and the program will generate an infinite
number of image variations based on the input. After some time of ex-
perimentation, it is not difficult to understand how to use this synthetic
brush to create digital images. Thus, A suitable selection of keywords can
be used to take advantage of the system’s limitations and circumvent
the rules established by programmers and taggers. (Figure 3).
Nevertheless, I have come across images of sexualized female
bodies in abundance on social media fan communities dedicated to this
Figure 3: An example of
a generated image using the Midjourney tool (Figure 4), with users expressing great satisfaction with their qual-
engine, shared on Facebook, screenshot, ity. Given that the learnt network can best mimic the scenes that were
Barbora Trnková, 2022
prevalent in the dataset, it is evident that it contains a significant amount
of sexualized female bodies. However, it also turns out to be a syncretic
tool that can be used by designers for sketching and designing. Indeed,
the speed of algorithms allows them to experiment quickly, which can
contribute to the development of these fields, simplifying their work,
but also putting many designers out of work.
In 2020, Vladan Joler and Matteo Pasquinelli (Joler, V.
& Pasquinelli, M., 2020) summarized the biases and limitations of learning
algorithms in order to break the unfulfillable promises we associate with
this technology. They pointed out that data collection itself is neither
technically nor socially neutral. Input data is already biased at the collec-
tion stage where old and conservative taxonomies can cause a distorted
world view, reduce social diversity, and exacerbate social hierarchies.
In addition, they identified machine-algorithmic bias, often known as
statistical or model bias. During this process, the dataset is contorted
by information compression, dimensionality reduction, and statistical
diffractions, resulting in the loss of a substantial amount of variety that
is useless to these algorithms. However, the reduction also occurs when
the algorithm is trained on the data - pattern extraction based on clas-
Figure 4: An example of
a generated image using the Midjourney sification occurs. It can be said that in the process of machine learning,
engine, shared on Facebook, screenshot, the world is compressed into a statistical model, in Dan McQuillan’s words,
Barbora Trnková, 2022
346 347