Page 83 - AC-2-3

Arts & Communication Computer vision in tactical AI art

(since 2016) uses CV pareidolia to make a generative for each category. The classification dataset was created in
11
81
typeface from alphabet shapes found in aerial imagery. the project’s initial phase, which invited visitors to assign
It features detailed documentation, a font catalog, and one of these categories to the images in the archive of the
an interactive word processor where visitors can enter Photographers Gallery where the work was later installed.
text and choose the font size, line spacing, different font Similarly, the Decisive Mirror (2019) analyzes and
86
classes, and source locations for typesetting. Related works classifies visitors’ faces based on unconventional “traits”
include Shinseungback Kimyonghun’s Cloud Face (2012), (predefined classification categories), such as “aliveness,”
Onformative Studio’s Google Faces (2013), and Driessens “imaginariness,” or “one-of-themness,” to underline the
and Verstappen’s Pareidolia (2019). Although they arbitrariness and randomness that creep into the ML
prioritize the decorative and amusing aspects of CV over profiling categories (Figure 10). It is worth noting that in
its critique, these works hint at the shadier sides of errors these projects, Schmieg already addressed the experiential
in automated vision systems. revelation of AI’s limitations and flaws to challenge
Unlike the biases in AI-powered CV, the risky “creative the presumptive objectivity and interpretive validity of
87
biases” and idiosyncrasies are desirable in tactical art, computational processes – the same poetic punctum that
where they can catalyze conceptual cogency and expressive garnered acclaim for Paglen and Crawford’s latter ImageNet
economy. For instance, Jennifer Gradecki and Derek Roulette (2019, discussed above). 74
Curry’s Boogaloo Bias (2021) forsakes Lozano-Hemmer’s The statistical nature of CV and the related ambiguities
83
solemn tone in Level of Confidence (2015, discussed in the of ML are sometimes utilized to address the questionable
previous section) to cast a sarcastic look at the biases intersections of AI and creativity. In Adam Basanta’s
66
and errors in CV translation processes and the impact of installation, All We’d Ever Need Is One Another (2018),
88
datasets and accuracy thresholds on false positives in police custom software randomizes the settings of two mutually
surveillance and arrest policies (Figure 9). Trained on the facing flatbed scanners so that in every scanning cycle,
actors’ faces from the movie Breakin’ 2: Electric Boogaloo each captures a slightly altered mix of the facing scanner’s
(1984, directed by Sam Firstenberg), the facial recognition light and its own unfocused scanning light reflected
algorithm in this work brute-forces the generation of off the facing scanner’s glass plate. The perceptual
leads to “identify” members of the Boogaloo Bois anti- hashing algorithms use parameters such as aspect ratio,
law enforcement militia in their live video feeds and composition, shape, and color distribution to compare
protest footage. “Brute-forcing” in ML facial recognition each new scan to an image database scraped from
adopts the methods of police investigations in the US,
which occasionally compensate for the absence of high- freely accessible online artwork repositories. When the
quality suspects’ images by substituting artist sketches with comparison value between the scan and the most similar
filtered social media photographs, computer composites, database image exceeds 83%, the software declares a
or celebrity images that resemble suspects. Hence, unlike “match,” selects the scan for printing, and labels it according
regular facial recognition algorithms, tuned to minimize to the database image metadata. After it printed one of
the matching uncertainty level, Boogaloo Bias is optimized the scans and labeled it “85.81%_match: Amel Chamandy
to return the highest number of suspects. ‘Your World without Paper,’ 2009,” Canadian artist Amel
Chamandy initiated a legal action over the intellectual
In several landmark projects bolstered by neat and property rights against Basanta because of the reference
striking humor, Sebastian Schmieg expands the problem to her photograph. However, the “85.81%_match” is not
space of image recognition accuracy and normalization for sale, and Basanta does not use it for direct commercial
by emphasizing the fundamental but insufficiently gains by any other means. He consistently applied the
investigated philosophical dimensions of AI ontology functional logic of ML to leverage the open-endedness
and epistemology. He introduces deliberately reduced of creative work and disturb the entrenched notions of
84
and unconventional (seemingly absurd) taxonomies into agency, authorship, originality, and intellectual property
the CV classification libraries used in interactive setups crystalized in copyright laws. 89
to process visitors’ uploaded or captured pictures. For
example, online visitors of the Decisive Camera (2017 – This line of practices places the conceptual, technical,
2018) can upload an image that will be classified within and broader sociopolitical problems of AI engineering and
85
a taxonomic space of only four categories (Problem, application firmly within the human context, even in cases
Solution, Past, and Future) with a probability percentage when their expressive cogency and professional ethics are
debatable. They point to the reflections of human nature
11 Pareidolia is a tendency to recognize patterns within random and the (political) rationales encoded into CV software
visual data. 82 and hardware architectures, which modulate both the

Volume 2 Issue 3 (2024) 10 doi: 10.36922/ac.2282