Algorithmic Empathy:
Toward a Critique of Aesthetic AI
Hannes Bajohr
University of Basel
ABSTRACT: With artificial intelligence making inroads into the arts, a
critique of aesthetic AI still needs to be written. To this end, this article
argues that first, one must do away with the “Promethean anxiety”
that assesses machine-created works by the standards of human-made
ones, and second, one must turn to the technical substrate of such
works for criteria of aesthetic critique. The article takes digital literature as an example and suggests a distinction between the “sequential
paradigm” of linear algorithms and the “connectionist paradigm” of
neural networks. Such media-specificity finds its aesthetic correlate in
the medium-specificity of text and image.
Promethean Anxiety; or, Creativity as the Last Differentia
Let me start with an observation. It was recorded in 1942 by German philosopher Günther Anders. Having escaped the Nazis and
living in California at the time, Anders brought with him the distanced sensibility of the European exile who, not unlike his fellow
émigré Theodor W. Adorno, understood America, and California in
particular, as the intensified expression of life in capitalist modernity. In a journal entry, which would later become the first chapter
of his book The Obsolescence of Human Beings, he described a visit to
a technology exhibition in which a friend acted rather curiously: as
if he were ashamed to be a human and not a machine. This, Anders
noted, was a novel phenomenon, “an entirely new pudendum …; a
form of shame that did not exist in the past. I will provisionally call
it ‘Promethean shame.’” This was to denote “the shame felt when
Configurations, 2022, 30:203–31 © 2022 by Johns Hopkins University
Press and the Society for Literature, Science, and the Arts.
203
204
CONFIGURATIONS
confronted by the ‘humiliatingly’ high quality of fabricated things.”1
In face of the perfection, reliability, and repeatability of modern
machines and mass-produced objects, Anders contended, humans
feel themselves to be deficient: unfinished, unreliable, trapped in
fragile bodies—confronted with the flaw of having been born rather
than produced. The embarrassment of the builder in the face of the
built is only the first sign of the looming obsolescence of the human. For Anders, this was connected to the atom bomb no less than
to the Taylorization of the production of goods, to mass fabrication
and—worst of all for the European aesthete—to television and its
all-pervasive reach.
One may find Anders’s analysis a tad too apocalyptic, but it is
nevertheless heuristically useful: angels and animals—the cosmologically superior and the ontologically inferior—no longer form humans’ basis of comparison. In a secular society and one in which the
domination of nature is total, Anders held, the machine and the serialized product become the new foils for human self-understanding.
Nevertheless, “shame” might not be the right word for what asserts
itself rather as a worry. It may be more useful to speak of a Promethean
anxiety: the fear of losing the status of maker and a reversal of the
hierarchy of human and machine.
The current discussion about artificial intelligence (AI) and creativity seems to be an especially pertinent example of this anxiety,
and one in which the human-machine comparison thrives. But if
reasoning power was formerly the differentia that distinguished humans from machines, today it is the arts that have become the most
recent frontier of such human-machine comparisons, and a powerful source of Promethean anxiety. While in 1968, in a description of
the exhibition Cybernetic Serendipity, computer-generated art could
be touted simply as “creative forms engendered by technology,”2
that is, as subordinate to the control exerted by their human creators, this clear relationship is put into doubt today.
One field that has made particularly large strides in the past deI wish to thank Michel Chaouli, Julia Pelta Feldman, Annette Gilbert, Markus Krajewski, Colin Lang, Christina Vagt, and two anonymous reviewers for their comments on
various versions of this paper; I am also indebted to the discussants at events at University of California, Santa Barbara, the University of Chicago, Technical University
Braunschweig, and Free University Berlin (both Germany), where I had the opportunity to present versions of this paper.
1. Günther Anders, “On Promethean Shame,” in Christopher John Müller, Prometheanism: Technology, Digital Culture and Human Obsolescence (London: Rowman & Littlefield,
2016), p. 30.
2. “Cybernetic Serendipity,” Magazine of the Institute of Contemporary Arts 5 (1968), p. 2.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
205
cade is machine learning3—especially artificial neural networks used
for creating artworks. To take just the most prominent—and most
contested—example, art group Obvious’s Edmond de Belamy, an inkjet-on-canvas print billed as the first “AI generated painting,” was
sold at Christie’s in 2018 for $432,500.4 Although machine-produced
art is much older,5 the fact that an artificial neural net was involved
in the production and even figured as the artist (its formula being
the signature at the bottom right of the painting) gave this work
the character of a caesura. And while there are more sophisticated
works of AI art—one may think of Trevor Paglen’s series Adversarially Evolved Hallucinations (2017) or Hito Steyerl’s installation Power
Plants (2019)—the staggering selling price and the work’s utilization of traditional attributes of painting, down to the gilded picture
frame, induced, as Ian Bogost called it, an “AI gold rush” in the visual
arts; since then, we have seen many more works like this enter the
market.6
In the textual arts, machine learning has seen a similar popularity. So-called “large language models”—like those developed by the
think tank Open AI, GPT-2 (2019) and GPT-3 (2020)7—are able to
produce surprisingly human-like texts, running coherently across
several paragraphs. An OpenAI blog entry introducing GPT-3 included an example in which the model was tasked to continue a
prompt that included characters from Lord of the Rings; the result
3. For easy-to-follow introductions to this technology, see Ethem Alpaydin, Machine
Learning: The New AI (Cambridge, MA: MIT Press, 2016); John D. Kelleher, Deep Learning
(Cambridge, MA: MIT Press, 2019); and Melanie Mitchell, Artificial Intelligence: A Guide
for Thinking Humans (New York: Farrar, Straus, and Giroux, 2019).
4. Obvious Collective, Edmond de Belamy, GAN algorithm, inkjet on canvas, Obvious
Collective website, 2018, https://obvious-art.com/portfolio/edmond-de-belamy.
5. See for an overview Grant D. Taylor, When the Machine Made Art: The Troubled History
of Computer Art (London: Bloomsbury, 2014).
6. Ian Bogost, “The AI-Art Gold Rush Is Here,” Atlantic (March 6, 2019), https://www
.theatlantic.com/technology/archive/2019/03/ai-created-art-invades-chelsea-gallery
-scene/584134/. See for contemporary artistic engagements with AI: Joanna Zylinska,
AI Art: Machine Visions and Warped Dreams (London: Open Humanities Press, 2020).
7. These are no longer the largest models, but their relative ease of use as well as the
integration of GPT-3 into a pay-for-use service have made them the de facto standard
for the nonprofessional use of natural language generation; open source initiatives like
GPT-NeoX (EleutherAI) have, at least as of this writing, garnered far fewer users; see
https://www.eleuther.ai. On the political and ethical problems of such language models, see Emily Bender et al., “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” in FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (New York: Association for Computing Machinery, 2021),
pp. 610–23.
206
CONFIGURATIONS
is, in structure and tone, fantasy fiction.8 GPT-3, released in 2020, is
two orders of magnitude larger than its predecessor. The paper that
announced its launch prompted GPT-3 to “compose a poem in the
style of Wallace Stevens with the title ‘Shadows on the Way.’”9
The success of models like these has, on the side of enthusiasts
of computational creativity, fueled expectations for machine learning to create complex, coherent textual works, and not least literary
ones. But for most everyone else, both Edmond de Belamy and the
GPT models suggest the question: Will machines replace artists? It
is an exclamation of Promethean anxiety, the anxiety that humans
might cede their status as creators to machines. What is more, it
identifies art-making as the differentiating element by which humans still could triumph over machines but are at risk of no longer
doing so.
This anxiety even comes to the fore in writers that appear to have
a positive outlook on computer-generated art and literature, and it
does so in the specific aesthetic and anthropological categories they
employ. In his 2019 book The Creativity Code, Marcus du Sautoy, professor of mathematics at Oxford University, undercuts his apparent
enthusiasm about the possibilities of machinic art by his oft-repeated
conviction that it is “creativity that makes us human.” Du Sautoy
goes so far as to posit a biologically hard-wired “creative urge”—a
classic anthropological differentia—that he holds up against the en8. The text begins: “The orcs’ response was a deafening onslaught of claws, claws, and
claws; even Elrond was forced to retreat. ‘You are in good hands, dwarf,’ said Gimli,
who had been among the first to charge at the orcs; it took only two words before their
opponents were reduced to a blood-soaked quagmire, and the dwarf took his first kill
of the night. The battle lasted for hours until two of the largest Orcs attempted to overwhelm Aragorn. When they finally stopped, they lay defeated and lifeless for miles and
miles.” Alec Radford et al., “Better Language Models and Their Implications,” OpenAI
Blog, Feb. 14, 2019, https://openai.com/blog/better-language-models. Among the most
detailed investigations of the poetic potential of GPT-2 and 3 are the articles of pseudonymous author Gwern Branwen; see Gwern Branwen, “GPT-2 Neural Network Poetry,” Oct. 29, 2019, https://www.gwern.net/GPT-2; Gwern Branwen, “GPT-3 Creative
Fiction,” June 1, 2021, https://www.gwern.net/GPT-3; and Gwern Branwen, “On
GPT-3: Meta-Learning, Scaling, Implications, and Deep Theory,” June 3, 2021, https://
www.gwern.net/newsletter/2020/05#gpt-3.
9. One of the poems prompted this way reads: “I must have shadows on the way / If I
am to walk I must have / Each step taken slowly and alone / To have it ready made //
And I must think in lines of grey / To have dim thoughts to be my guide / Must look
on blue and green / And never let my eye forget / That color is my friend / And purple
must surround me too // The yellow of the sun is no more / Intrusive than the bluish
snow / That falls on all of us. I must have / Grey thoughts and blue thoughts walk with
me / If I am to go away at all.” Tom B. Brown et al., “Language Models Are Few-Shot
Learners,” ArXiv (May 28, 2020), http://arxiv.org/abs/2005.14165, p. 49.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
207
croachment of the machines into art, music, and literature. He starts
his book with the question, “Can machines be creative?” and ends it
with the defiant proclamation: “Creativity is about humans asserting
they are not machines.”10
That art is the test case for human-machine difference is also posited by Arthur I. Miller’s book The Artist in the Machine (2019). Although Miller is more open to nonhuman aesthetics than du Sautoy,
his rhetoric nevertheless constantly returns to the anthropological
comparison he claims as only one option among many. His question
is whether machines must have not only reason or consciousness
to make art, but also emotions,11 which are then expressed in their
products. In AI art, Miller contends, computers “exhibit not only
their creativity but their inner lives.” This rhetoric of interiority and
expression hints at a very specific, post-romantic idea of art-making.
It is not surprising that Miller liberally employs the word “genius” to
describe both human and machine artists.12
Miller and du Sautoy are examples from popular literature, but
for an increasing number of engineers and scholars trying to operationalize and simulate art-making, creativity is chosen as the decisive criterion of art, which inherently evinces artists’ expressions
and intentions.13 Often, it is defined either via evolutionary biology
or a neuroscientific description of brain functions, which both lend
10. Marcus du Sautoy, The Creativity Code: How AI Is Learning to Write, Paint, and Think
(London: Fourth Estate, 2019), pp. 297, 302.
11. This is reminiscent of Geoffrey Jefferson’s demand, dismissively cited by Alan Turin: “Not until a machine can write a sonnet or compose a concerto because of
thoughts and emotions felt, and not by the chance fall of symbols, could we agree that
machine equals brain—that is, not only write it but know that it had written it. No
mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure
at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by
its mistakes, be charmed by sex, be angry or depressed when it cannot get what it
wants.” Geoffrey Jefferson, “The Mind of Mechanical Man,” British Medical Journal
(June 25, 1949): 1005–100, at p. 1110, as cited in Alan Turing, “Computing Machine
and Intelligence,” Mind 59:236 (1950): 433–60, at pp. 445–46.
12. Arthur I. Miller, The Artist in the Machine: The World of AI-Powered Creativity (Cambridge, MA: MIT Press, 2019), p. 54.
13. The literature on computational creativity is vast, and it may suffice to point to the
work of Margaret Boden, who is often referred to in this field. See Margaret Boden, The
Creative Mind: Myths and Mechanisms, 2nd ed. (London: Routledge, 2004), as well as
Margaret Boden, “Computer Models of Creativity,” AI Magazine 30:3 (2009): 23–43.
However, the problem for all authors who—like Miller and du Sautoy—take their inspiration from the philosophy of creativity lies in the fact that they may conflate creativity, which is not domain-specific, with art itself, a specific social field. This levels the
difference that must exist between an artwork, a clever invention, and a particularly
disruptive business strategy.
208
CONFIGURATIONS
themselves, at least in principle, to the operationalization in computers.14 But these strongly anthropocentric categories ignore any
serious contemporary aesthetic theory that is not a neuro-aesthetics,
and their conception of art is flagrantly out of date. It is telling that
the non-digital artworks in du Sautoy’s book are no more recent than
the 1950s, when avant-garde art movements like abstract expressionism celebrated the spontaneous, creative genius one last time.15
What is more, these approaches are insufficient to fulfill the need
they themselves raise: a critique of aesthetic AI. They are too laden
with Promethean anxiety to capture what is specific to the aesthetic
use of AI. Instead, they tend to work by a logic of transference, first
from human to machine, and then from old media to new. Edmond
de Belamy is the best example here: the old medium of painting in a
new media guise, created not by humans but (ostensibly) produced
by a machine. But it may be more interesting, and more productive,
to investigate aesthetic approaches beyond the foil of the human,
and to explore the affordances of the new medium instead of simply
replicating old ones.
Eschewing talk about conscious machines or any specifically human creative urge, I instead want to look at the way these works
work, and which technological and aesthetic structures they implement. Though many of my remarks are applicable to the arts as a
whole, my focus is on digital literature. Digital literature is useful as
a test case for a critique of aesthetic AI because it has a certain history. Against the more “traditional” types of digital literature, the
novelty of neural network–based texts is thrown into sharp relief.
The “traditional” type I want to provisionally call the sequential, the
new type the connectionist paradigm. While the rule-based sequential
paradigm of digital literature can look back on a rich critical apparatus, the “non-transparent” connectionist paradigm is still undertheorized. In what follows, I offer some reflections on the differences
between these paradigms, and hint at what we should keep in mind
while developing a critique of aesthetic AI that eschews the pitfalls
14. Denis Dutton, The Art Instinct: Beauty, Pleasure, and Human Evolution (London:
Bloomsbury, 2009) argues for an evolutionary approach to art, while Anna Abraham,
The Neuroscience of Creativity (Cambridge: Cambridge U. Press, 2018) argues for a neuroscientific discussion of art and creativity. For a critique of such models of creativity,
see Hannes Bajohr, “No Experiments: No Experiments: On Artistic Artificial Intelligence and Literary Writing,” CounterText 8:2 (2022), forthcoming.
15. The one notable exception is du Sautoy’s discussion of Gerhard Richter’s permutative work 4900 Farben (2007). While he reflects on the uses of mathematics for art, he
in no way engages with this work as an example of an inexpressive, conceptual art
practice. Du Sautoy, Creativity Code (above, n. 10), pp. 92–93.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
209
of Promethean anxiety and its human-machine transference of aesthetic categories.
Two Types of Digital Literature: Sequential and Connectionist
Digital, or electronic, literature is a wide-ranging, many-faceted field.
It contains such a large variety of genres and technologies—from
hypertext-novels to codeworks to kinetic literature—that it is hard
to offer a definition that goes beyond its very basic characteristics. In
the formulation of the Electronic Literature Organization, the term
refers “to works with an important literary aspect that takes advantage of the capabilities and contexts provided by the stand-alone
or networked computer.”16 Literary scholar Jessica Pressman noted
that many more recent works of digital literature consciously align
themselves again with the modernist tradition.17 Among the genres
and traditions of digital literature, the most inherently modernist is
also the oldest.18 I refer to it as generative literature, and it is this genre
I want to focus on here. At its most basic, generative literature denotes the automatic production of text according to predetermined
parameters, usually following a combinatory, sometimes aleatory
logic, and it emphasizes the production rather than the reception
of the work (unlike, say, hypertext literature). Scott Rettberg, in his
2019 book Electronic Literature, highlights the generative tradition’s
connection to Dada and surrealism, to Oulipo as well as Fluxus.19
I would add conceptual art, particularly in the vein of Sol LeWitt
and Lawrence Weiner, as a further important reference, since here,
too, the formulation of a concept and its execution into a work are
distinct from one another, and one may see the relation between
concept and work echoed in that between code and output.20
Yet it is not only its age and the historic lineage, but also something about the use of the underlying technology that gives generative literature a special status among the many varieties of digital
literature. It seems to deliberately reflect its underlying technology.
16. Katherine Hayles, Electronic Literature: New Horizons for the Literary (Notre Dame:
U. Notre Dame Press, 2009), p. 3.
17. Jessica Pressman, Digital Modernism: Making It New in New Media (Oxford: Oxford
U. Press, 2014).
18. See Florian Cramer, Words Made Flesh: Code, Culture, Imagination (Rotterdam: Piet
Zwart Institute, 2005).
19. Scott Rettberg, Electronic Literature (Cambridge: Polity, 2019).
20. I make this point in Hannes Bajohr, “Das Reskilling der Literatur,” in Code und
Konzept: Literatur und das Digitale, ed. Hannes Bajohr (Berlin: Frohmann, 2016),
pp. 7–21.
210
CONFIGURATIONS
This is visible already in one of the first examples of generative
literature: Theo Lutz’s “Stochastische Texte” (Stochastic texts), which
he wrote—or rather, generated—in 1959, one year after Anders’s
book on Promethean shame was published. “Stochastische Texte”
is the output of an algorithm that combined elements from a predetermined vocabulary taken from Franz Kafka’s Castle.21 Each line
contains statements that are connected by conjunctions or separated
by a period, such as “NOT EVERY LOOK IS NEAR. NO TOWN IS
LATE,” or “A CASTLE IS FREE AND EVERY FARMER IS FAR,” or “EVERY STRANGER IS FAR. A DAY IS LATE,” and so on (figure 1).
Lutz’s “Stochastische Texte” belongs to what I would like to call
the sequential paradigm within the genre of generative literature: it is
executed as a sequence of rule-steps, and its identity is encoded in
its production much more than in its reception. A colleague of Lutz,
while not providing the program code, sketched the program flowchart in a later article, and the sequential and step-wise nature is obvious here (figure 2). Instead of hoping to recreate intuition, genius,
or expression, the logic of the machine itself—that is, the logic of
deterministically executed rule steps—becomes aesthetically normative in “Stochastische Texte.” One could sense in this an “algorithmic empathy”—a non-anthropocentric Einfühlung in the sense of
hermeneutical understanding that is aimed not at the psychological
states of the artists but at comprehending the process of the work’s
material production.22
For Lutz’s text, we only have an abstract description of the individual steps; the code he used is not (yet) available.23 For much of
21. See Kurt Beals, “‘Do the New Poets Think? It’s Possible’: Computer Poetry and Cyborg Subjectivity,” Configurations 26:2 (2018): 149–77, as well as Barbara Büscher, Christoph Hoffmann, and Hans-Christian von Herrmann, eds., Ästhetik als Programm. Max
Bense / Daten und Streuungen (Berlin: diaphanes, 2004).
22. I use the term “empathy” strictly as a descriptor of a hermeneutic movement generating insight, which need not be anthropological in any way. It goes back to Wilhelm
Dilthey’s notion of hermeneutics as “Nacherleben” (re-experiencing) of experiences of
the other as well as Edmund Husserl’s notion of intersubjective understanding as “Einfühlung” (empathy). See Wilhelm Dilthey, “The Understanding of Other Persons and
Their Expression of Life,” Descriptive Psychology and Historical Understanding, trans. Richard M. Zaner and Kenneth L. Heiges (The Hague: Nijhoff, 1977), p. 133; and Edmund
Husserl, Ideas Pertaining to a Pure Phenomenology and a Phenomenological Philosophy, Volume 1: General Introduction to a Pure Phenomenology, trans. Fred Kerten (Den Haag: Nijhoff, 1983), pp. 6, 79.
23. German computational historian Toni Bernhart is in the process of retrieving the
original programming code for “Stochastische Texte.” A first preview is given in Toni
Bernhart, “Beiwerk als Werk: Stochastische Texte von Theo Lutz,” editio 34 (2020): 180–
206; see also Toni Bernhart/Sandra Richter, “Frühe digitale Poesie: Christopher Strachey
und Theo Lutz,” Informatik Spektrum 44 (2021): 11–18
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
211
Figure 1: Theo Lutz, “Stochastische Texte,” augenblick 4:1 (1959): 3–9.
contemporary digital literature, this is fortunately not the case. A
more recent and more complex example of a sequential work that
inspires algorithmic empathy is Nick Montfort’s 2014 Megawatt.24
It refers not only to its own structural make-up but also to that of a
modernist classic: it is both an interpretation and an appropriation
of Samuel Beckett’s novel Watt.25
Written between 1942 and 1944 but published in 1958, Watt depicts the titular Mr. Watt’s entry into the household of Mr. Knott as
the latter’s servant. However, it is not the fabula but the linguistic
structure, the textual surface, that is most characteristic of this novel.
In addition to the deliberately unidiomatic English, the extremely
repetitive passages stand out—its “geometric audacity,”26 as T. W.
McCormack called it—which since Watt’s publication have been interpreted as a failure of language and a critique of the insurmountable hyperrationality of modernity.27 Take for instance this passage,
24. Nick Montfort, Megawatt (Cambridge, MA: Bad Quarto, 2014).
25. Samuel Beckett, Watt (New York: Grove Press, 1970).
26. W. J. McCormack, “Seeing Darkly: Notes on T. W. Adorno and Samuel Beckett,”
Hermathena 141 (1986): 22–44, at p. 24.
27. See Linda Ben-Zvi, “Samuel Beckett, Fritz Mauthner, and the Limits of Language,”
PMLA 95:2 (March 1980): 183–200, at p. 183; Shane Weller, “Humanity in Ruins: Samuel Beckett,” in Language and Negativity in European Modernism (Cambridge: Cambridge
212
Figure 2: Rul Gunzenhäuser,
“Zur Synthese von Texten mit
Hilfe programmgesteuerter
Ziffernrechenanlagen,” MTW
10:4 (1963): 4.
CONFIGURATIONS
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
213
in which Watt cannot follow a conversation partner because he is
distracted by voices in his head:
Now these voices, sometimes they sang only, and sometimes they cried only,
and sometimes they stated only, and sometimes they murmured only, and
sometimes they sang and cried, and sometimes they sang and stated, and
sometimes they sang and murmured, and sometimes they cried and stated,
and sometimes they cried and murmured, and sometimes they stated and
murmured, and sometimes they sang and cried and stated, and sometimes
they sang and cried and murmured, and sometimes they cried and stated and
murmured, and sometimes they sang and cried and stated and murmured, all
together, at the same time, as now, to mention only these four kinds of voices,
for there were others. And sometimes Watt understood all, and sometimes he
understood much, and sometimes he understood little, and sometimes he understood nothing, as now.28
A recent interpretation of Watt by Amanda M. Dennis speaks of these
repetitions as “obsessive loops.” “Certain passages make language
appear to ‘glitch,’ as if it were a malfunctioning computer program
or electronic device.”29 When one takes a closer look at Megawatt,
Nick Montfort’s text based on Watt, one begins to doubt whether the
metaphor of the glitch is appropriate. Indeed, Megawatt shows that
the “obsessive loops” are not glitches, not errors in Beckett’s program, but on the contrary represent its most consistent execution.
In fact, these repetitive, list-like loops seem to follow an immanent
rule—an algorithm.
Taking a closer look at this passage from Watt allows us to infer
the production principle of what Hugh Kenner has called Beckett’s
“Cartesian sentences.”30 The first sentence applies a simple text generation rule: the permutation of combinatorial possibilities from a
finite set of elements. The “voices” can take on four possible states—
“sang,” “cried,” “stated,” “murmured”—either individually or in
various combinations, and Beckett cycles through all of them. Then,
Watts’s understanding is assigned the values “all,” “much,” “little,”
and “nothing,” one after the other; here, the verbs are not permutated, but only listed. Programmatically speaking, the sentences reU. Press, 2018), pp. 90–125.
28. Beckett, Watt (above, n. 25), p. 29.
29. Amanda M. Dennis, “Glitches in Logic in Beckett’s Watt: Toward a Sensory Poetics,”
Journal of Modern Literature 38:2 (2015): 103–16, at p. 104.
30. Hugh Kenner, The Mechanic Muse (New York: Oxford U. Press, 1987), p. 91. It was
Kenner who first tried to recreate parts of Watt in the programming language PASCAL;
unlike Montfort, however, he did not expand on Beckett.
214
CONFIGURATIONS
semble a function that assigns a value to a variable, and it could be
generated automatically with the same result by a script.
This is exactly what Montfort did in Megawatt. It is in fact a reconstruction and an extension of Beckett’s novel in one. Montfort
selected passages with such “obsessive loops” from the original, and
recreated them in the programming language Python. In the first
chapter, titled “The Voices,” he turns to the passage just discussed,
and generates it. But the script goes further:
Watt heard voices. Now these voices, sometimes they sang only, and sometimes they
cried only, and sometimes they stated only, and sometimes they murmured only, and
sometimes they babbled only, and sometimes they chattered only,
and sometimes they ranted only, and sometimes they whispered
only, and sometimes they sang and cried, and sometimes they sang and stated, and
sometimes they sang and murmured, and sometimes they sang and babbled, and sometimes they sang and chattered, and sometimes they
sang and ranted, and sometimes they sang and whispered, and sometimes they cried and stated, and sometimes they cried and murmured, and sometimes they cried and babbled, and sometimes they cried and chattered, and sometimes they cried and ranted, and sometimes they
cried and whispered, and sometimes they stated and murmured, and sometimes they stated and babbled, and sometimes they stated and
chattered, and sometimes they stated and ranted, and sometimes
they stated and whispered, and sometimes they murmured and
babbled, and sometimes they murmured and chattered, and sometimes they murmured and ranted, and sometimes they murmured
and whispered, and sometimes they babbled and chattered, and
sometimes they babbled and ranted, and sometimes they babbled
and whispered, and sometimes they chattered and ranted, and
sometimes they chattered and whispered, and sometimes they
ranted and whispered, and sometimes they sang and cried and stated, and
sometimes they sang and cried and murmured. . . . And sometimes Watt understood
all, and sometimes he understood most, and sometimes he understood
much, and sometimes he understood half, and sometimes he understood
little, and sometimes he understood less, and sometimes he understood bits, and sometimes he understood nothing, as now.31
31. Montfort, Megawatt (above, n. 24), pp. 1, 7. Montfort reconstructs Beckett’s “algorithm” with such precision that, as in Watt, commutatively occuring combinations do
not appear twice. Expressing this with numerical placeholders: after listing the basic
elements 1, 2, 3, 4, Beckett combines 1 and 2, 1 and 3, and 1 and 4 (note that no element is combined with itself). Yet instead of proceeding with 2 and 1—which is already
covered by 1 and 2—he directly goes on to 2 and 3. So does Montfort with his expanded eight elements. This is important insofar as Beckett does not seem to follow his
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
215
Because Beckett admits that there are more voices (“for there were
others,” as it says at the end of the first sentence quoted above),
and because Montfort knows that in a permutation series the number of possibilities per element increases exponentially, he adds four
verbs to Beckett’s four: “babbled,” “chattered,” “ranted,” and “whispered.” Likewise, Watt can now additionally understand “most,”
“half,” “less,” and “bits.” Montfort’s own contribution consists of
the first three words, the merely expository first sentence (“Watt
heard voices”), and the eight additional words. Both Beckett’s text
and the extensions, however, are generated purely by the code. It
outputs what Beckett actually wrote (the italic text), and what he
would have written, according to his own rules, if he had expanded
his set of elements (the boldface text).
This can be seen very clearly in the Python source code of the
program, which is printed in the appendix of the book:32
1 #### THE VOICES
2 text.append('\n# I\n\n')
3 def combine(num, words):
ÀQDO >@
LIQXP!DQGOHQ ZRUGV ! QXP
LIQXP
ÀQDO ÀQDO>>ZRUGV>@@@
HOVH
ÀQDO ÀQDO>>ZRUGV>@@
FIRUFLQFRPELQH QXP²ZRUGV>@ @
ÀQDO ÀQDOFRPELQH QXPZRUGV>@
UHWXUQÀQDO
,Q:DWWWKHYRLFHV > VDQJ FULHG VWDWHG PXUPXUHG @
$QG:DWWXQGHUVWRRG > DOO PXFK OLWWOH QRWKLQJ @
self-imposed rules strictly in other passages of Watt, which would make their programmatic reconstruction more difficult. (I owe this latter information to Robert Stockhammer.)
32. Ibid., pp. 242–43. The source code, however, is not simply added to the book;
rather, it generates the book as well as the appendix containing itself. This recursive
structure is a frequent characteristic of digital literature.
216
CONFIGURATIONS
+HUHWKHYRLFHVGLGHLJKWWKLQJVDQGWKHUHDUHHLJKWOHYHOV
YRLFHV > VDQJ FULHG VWDWHG PXUPXUHG EDEEOHG FKDWWHUHG UDQWHG ZKLVSHUHG @
XQGHUVWRRG > DOO PRVW PXFK KDOI OLWWOH OHVV
ELWV QRWKLQJ @
SDUD
SUHIDFH DQGVRPHWLPHVWKH\
IRUQXPLQUDQJH OHQ YRLFHV
IRUZRUGBOLVWLQFRPELQH QXPYRLFHV
SDUD SDUDSUHIDFH DQG MRLQ ZRUGBOLVW
LIOHQ ZRUGBOLVW
SDUD
SDUD SDUD RQO\
:DWWKHDUGYRLFHV1RZWKHVHYRLFHV SDUD>@
DOOWRJHWKHUDWWKHVDPHWLPHDVQRZWRPHQWLRQ
RQO\WKHVH VSHOOHGBRXW>OHQ YRLFHV @ NLQGVRIYRLFHVIRU
WKHUHZHUHRWKHUV$QGVRPHWLPHV:DWWXQGHUVWRRG
DQGVRPHWLPHVKHXQGHUVWRRG MRLQ XQGHUVWRRG DVQRZ
26 text.append(para)
After defining the function combine in lines 3–12—a subroutine that
in the end assembles the final text—Montfort shows how Beckett’s
own text can be understood as a set of elements of a list variable
(sometimes also called an array), that is, a single variable that contains a series of items. Here in line 13, the variable is called voices,
and its values are “sang,” “cried,” “stated,” “murmured”—exactly
the verbs that are permutated in Watt. But because there is a pound
sign in front of this line, the Python interpreter recognizes that the
line is merely a comment that should not be executed and ignores it.
Beckett’s concept is still present in the code, but has been, as it were,
switched off.
Instead, line 16—an executable line—contains the new list variable, this time extended by Montfort. In addition to the original four
verbs, it also contains the four additional ones: “babbled,” “chattered,” “ranted,” and “whispered.” The same happens for the vari-
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
217
able understood—first, Montfort lists the four original elements in a
comment in line 14, then he lists his extended set in line 17.
The rest of this short code section assembles these elements. The
empty variable para is defined in line 18—it will be assigned the
finished text at the end. Line 19 defines the variable preface, which
contains the regularly recurring statement “and sometimes they.” A
doubly nested loop follows in lines 20 to 23: it cycles through the
list variable voices and adds the words “and sometimes they” stored
in preface. Finally, the first sentence (the one with the voices) is
completed in line 25, and the second sentence (the one about understanding) gets added to it. In the second sentence, the elements are
not permuted; instead, the values stored in the variable understood
are simply listed. The result is the new, extended text—which I cannot show here completely, because it has grown exponentially and
is now 27 pages long.
Megawatt is a form of algorithmic empathy that is not a copy, but
a reconstruction. But while Megawatt is to Watt what Jorge Luis Borges’s Pierre Menard is to Cervantes’s Don Quixote—a reenactment—because it is not only reconstructive but also productive, it is also (if
hyperbole is admissible for a moment) what Joyce’s Ulysses is to the
Odyssey—an expansion beyond the original. Megawatt is thus not
only interesting as a literary product, an adaptation of an existing
text; it also actually produces knowledge about Beckett’s text and carries
out a hermeneutic movement—albeit a distinctly non-anthropocentric
one. It begins with the reconstruction of the original, whereby the
immanent rule from Beckett’s original is made explicit but switched
off as comment lines; and it proceeds to the extrapolation of these
rules, which are now made explicit, and their extension. This extension serves as a proof for the comprehension of Beckett’s principle.
The fact that this form of reconstruction is possible thus supports Jessica Pressman’s thesis that digital literature returns to the operations
of the historical avant-gardes, but implements them—as “digital
modernism”—with more appropriate means and more consistently.33 Moreover, Montfort’s book also suggests that Beckett’s Watt is
itself algorithmic, a proto-digital literature. In that Megawatt not only
emulates Watt, but also in a sense explodes it, not only imitates,
but also exaggerates it, it also highlights those parts of Watt that are
most apt for digital exploration, and does so in a hermeneutically
profitable way.
Megawatt is a recent example of the sequential paradigm as the
oldest type of generative and in fact digital literature as such. I have
33. Pressman, Digital Modernism (above, n. 17), 1–2.
218
CONFIGURATIONS
spent some time discussing its code to illustrate exactly how well—
by inspecting the source code—we can get a sense of its inner workings: each step of its sequence is laid out in front of us.34
In contradistinction to the sequential paradigm, I would like to
call the newest type of generative art the connectionist paradigm. Here
we turn to works in the mold of Edmond de Belamy as well as the
text generators GPT-2 and GPT-3. By “connectionist,” I refer to deep
neural nets as the most widespread machine learning technology.35
Neural nets follow, at least on a very basic and simplified level, the
logic of the network of connections between neurons and synapses
in the brain. (Incidentally, the first neural net also goes back to the
time of Anders and Lutz, when in 1958 Frank Rosenblatt created
the “perceptron”—modeled on the optical nerve rather than on the
brain itself—which was capable of learning and recognizing basic
patterns.36) At its most abstract, a neural net is made up of three main
elements: the input layer, one or more hidden layers, and the output
layer. In Rosenblatt’s model, there was only one hidden layer, but
modern deep neural networks are composed of a multitude of hidden layers made up of neurons and connected by synapses, whose
“weights” define the effect on the next neuron. The goal of a neural
net is to create a function that fits the input data onto a desired output; the resultant model can be used to create outputs that resemble
the inputs. The central point, however, is that a neural net cannot
be explicitly programmed in the strict sense. Rather, neural nets learn
implicitly by a repeated process of comparing input and output and
adjusting for the errors in each iteration. Thus, there is no code we
could inspect, only a list of numbers representing the structure of
the network and their weighted connections; such a list, however,
34. See Mark C. Marino, Critical Code Studies (Cambridge, MA: MIT Press, 2020), chap.
2 for a model of code interpretation.
35. The term “connectionist” in this context goes back to the pioneering study Parallel
Distributed Processing, which made neural networks—after first attempts in the 1960s—
accessible for mathematics and information science from the 1980s onward; see David
E. Rumelhard, James McClelland, and Geoffrey Hinton, “The Appeal of Parallel Distributed Processing,” in David E. Rumelhard, James McClelland, and PDP Working Group,
eds., Parallel Distributed Processing. Explanations in the Microstructure of Cognition, vol. 1,
Foundations (Cambridge, MA: MIT Press, 1986), p. 43. The earliest use of the term appears in Donald O. Hebb, The Organization of Behavior: A Neuropsychological Theory (New
York: Wiley, 1949), 58.
36. See Frank Rosenblatt, “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain,” Psychological Review 65:6 (1958): 386–408; Nils J.
Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements (Cambridge: Cambridge U. Press, 2010), pp. 64–74.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
219
is incredibly difficult to interpret. This is the famous “black box”
problem of neural nets.37
Edmond de Belamy is an example of the connectionist paradigm:
trained on a dataset of 15,000 portraits from the fourteenth to the
nineteenth century, the neural net produced an output that statistically resembles the works of the training set.38 Since the basic operation is to fit an input onto an output, neural nets have so far mostly
been used for re-producing the stylistic characteristics of the training
set—in this, they are not unlike Megawatt—but without the possibility of explicitly defining the rules by which this happens. And yet,
repetition is in the very nature of neural nets, so that their designers
must make an effort to avoid the phenomenon of “overfitting,” in
which not similar but the exact same output is repeated.39 Usually,
this is done either by introducing noise or by reducing the completeness of the training set. In Edmond de Belamy’s case, it seems that the
training was aborted before the resemblance to the inputs became
too strong, which gives the portrait its spectral quality.
In AI literature, we can observe similar effects, which are brought
about by the failure of proper semantic understanding on the part
of the model. Almost canonical already is Sunspring from 2016 by
Ross Goodwin, an AI-generated film script that was subsequently
professionally produced. Goodwin trained a neural net called Benjamin using over 300 science fiction film scripts, and had it output a
new one. While GPT-3’s proprietary model can produce impressively
coherent text, most homebrewed models still remain restricted by
limited network sizes and training sets. Likewise, in its juxtaposition of incongruent elements, Sunspring tends toward the absurd,
with stage directions like: “He picks up a light screen and fights the
security force of the particles of a transmission on his face.”40 As with
37. Davide Castelvecchi, “The Black Box of AI,” Nature 538 (Oct. 6, 2016): 20–23. Note
that the “opacity” of neural nets also sets them apart from Markov-chains, which can
be expressed in diagrammatic and explicit form, even though they operate probabilistically; in my separation between the sequential and connectionist paradigm, Markovchains would possibly constitute a third realm.
38. See Jan Løhmann Stephensen, “Towards a Philosophy of Post-Creative Practices?
Reading Obvious’ ‘Portrait of Edmond de Belamy,’” in Proceedings of POM Beirut (2019):
21–30, doi:10.14236/ewic/POM19.4.
39. Kelleher, Deep Learning (above, n. 3), p. 20.
40. “INT. SHIP
We see H pull a book from a shelf, flip through it while speaking, and then put it back.
H
220
CONFIGURATIONS
most works of neural net literature, we should assume that a good
deal of manual editing went into this process—but we cannot know
for sure, as there is no code as in the case of Megawatt that we could
study. It remains not only as obscure as the proverbial black box, but
also as nontransparent as the mind of the genius of old.
Toward a Critique of Aesthetic AI
There are an important number of differences between the sequential paradigm of generative literature that employs linear algorithms,
and the connectionist paradigm that is based on neural nets; these
differences may allow us to approach a critique of aesthetic AI that
does not simply compare them to human works.
The first difference is that a classic algorithm needs explicitly
stated procedural rules, while a neural net learns by example and
its rules of generation are not immediately visible. While Montfort
could select the number of words and their possible position in a
sentence, no such choices informed the production of Sunspring’s
script. Rather, it is generated via the neural net’s training process,
which is based on an input data set. The first paradigm functions
top-down, the second bottom-up; for one, explicit rules stand at the
beginning, for the other, implicit rules (the statistical model) are generated by the end. The classic algorithm functions deterministically,
where an identical initial state always produces an identical final
state; neural nets, however, work by statistical induction, which is
fuzzy—and it is so by design, as they adhere to the principle of “approximate computing,” which puts a premium not on the precision
of its results but the efficiency of processing large masses of data.41
In a future with mass unemployment, young people are forced to sell blood. That’s the
first thing I can do.
H2
You should see the boys and shut up. I was the one who was going to be a hundred
years old.
H
I saw him again. The way you were sent to me ... that was a big honest idea. I am not
a bright light.
C
Well, I have to go to the skull. I don’t know.
He picks up a light screen and fights the security force of the particles of a transmission
on his face.” Ross Goodwin (Benjamin), Sunspring, 2016, https://www.docdroid.net
/lCZ2fPA/sunspring-final-pdf.
41. See Weiqiang Lu, Fabrizio Lombardi, and Michael Schulte, “Approximate Computing: From Circuits to Applications,” Proceedings of the IEEE 108:12 (2020): 2103–17.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
221
A neural net would have a much harder time reconstructing Watt in
the way Megawatt did.
From this, the second point follows: for the sequential paradigm,
explicit rules and the deterministic process allow for a higher degree of transparency. Most obviously, the code itself is readable, but
maybe more importantly, it is also easy to infer the underlying rules
by running the program a couple of times and observing the output.
This is much harder when it comes to neural nets, whose inner workings may not be impossible to retrace—“explainable AI” is working
on this42—but, as complex statistical models, cannot simply be reduced to explicitly stated rules. Likewise, observing the output may
be able to give some clue as to the internal process, but will not allow
for the same precision of inference.
Third, this problem is exacerbated, for while linear algorithms
draw a stark distinction between program and data, between procedural rules and items in a database, the “knowledge” in a neural
net is not localized in some particular place. Rather, data and “program” are distributed throughout the whole system as a statistical
dependency. While Montfort could still build on lists of words, Sunspring—using an LSTM-RNN type of network—is character based, so
that no actual words are encoded in the model, just the likelihood of
one character succeeding the next.43 Instead of proceeding according
to atomistic elements that assemble wholes from parts, neural nets
have much stronger emergent properties that, put metaphorically,
work according to a Gestalt logic.44 Here, wholes are not simply re42. Wojciech Samek and Klaus-Robert Müller, “Towards Explainable Artificial Intelligence,” in Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, eds.
Wojciech Samek et al. (Cham: Springer, 2019), pp. 5–22. However, despite all the progress made toward opening the black box—through “membership attacks” and other
methods—no way has yet been found to translate their processes back into rule steps;
it is not unlikely that this is impossible in principle. The point, then, is thus not that
the black box is completely opaque, but that no amount of light can afford us the same
mode of clarity a sequential algorithm allows for. It is fitting that neural nets as objects
of research are always determined by an external perspective—they are not investigated
differently than brain structures or star clusters, that is, with an eye to explanation.
Traditional code, on the other hand, invites a hermeneutic, an internal perspective, and
is geared toward understanding. Verum ipsum factum, the principle that makes the difference between naturalism and hermeneutics, also applies here: Human-written code
will always be read differently than a machine-made weight model—all the imaging
techniques applied to neural nets notwithstanding.
43. See the influential blog post: Andrew Karpathy, “The Unreasonable Effectiveness of
Recurrent Neural Networks,” Andrej Karpathy Blog, May 21, 2015, https://karpathy
.github.io/2015/05/21/rnn-effectiveness.
44. See Hannes Bajohr, “The ‘Gestalt’ of AI: Beyond Atomism and Holism,” Interface
Critique 3:1 (2021): 13–35.
222
CONFIGURATIONS
ducible to their parts, but the training process allows the neural net
to learn the overall shape of something like a painting in the style
of nineteenth-century impressionism, or the overall shape of something like a science fiction film.45
Lastly, and this is a somewhat controversial point introduced by
German media theorist Andreas Sudmann: a linear algorithm, with
its if-then-else conditions that can be diagrammed in a flowchart,
follows the digital logic of discrete states, of on and off, zero and
one, tertium non datur. It is true that in neural nets the “neurons” in
each layer are also either firing or not, but the weights that inhibit or
amplify their activation are described through floating-point numbers “in an approximately analog, a quasi-analog way,” as Sudmann
puts it. If the connectionist paradigm is quasi-analog, it truly stands
in the most extreme contrast to the sequential paradigm.46 One does
not have to follow Sudmann to this extreme, but what is clear is
that there is a radical difference in the technical substance of both
systems. This technical difference, I believe, must translate into a
difference in the aesthetic theorization of such systems.
One approach to such an aesthetic critique of aesthetic AI would
be to investigate in which way the sequential and the connectionist paradigm relate to one of the oldest aesthetic concepts, that of
mimesis. Both Megawatt and Sunspring follow a logic of imitation,
but they do so in radically different ways. The former could be said
to adhere to what German philosopher Hans Blumenberg has called
imitation as construction—that is, the approximation of an existing
state through the inference of the rules that bring it about. The latter then would rather enact the notion of imitatio naturae, the mere
repetition of the real, without such procedural insight. (“Nature” in
this case would only describe the dataset, not the traditional notion
of nature as world or cosmos.) For Blumenberg, both are distinctly
connected to the question of the new. Construction indicates the
possibility of going beyond the given by understanding the rules
of its generation, as Megawatt demonstrates, and is thus decidedly
modern—while the imitatio naturae relies on the world as a binding
stock of things to represent, and belongs, Blumenberg holds, to an
45. This is very well illustrated with regard to poetry in Boris Orekhov and Frank
Fischer, “Neural Reading: Insights from the Analysis of Poetry Generated by Artificial
Neural Networks,” Orbis Litterarum 75:5 (2020): 230–46.
46. Andreas Sudmann, “Szenarien des Postdigitalen: Deep Learning als MedienRevolution,” in Machine Learning – Medien, Infrastrukturen und Technologien der künstlichen Intelligenz, eds. Christoph Engemann and Andreas Sudmann (Bielefeld: Transcript, 2018),
pp. 55–73, p. 66.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
223
ancient aesthetic.47 While I do not want to indicate that neural nets
are somehow aesthetically premodern, I believe that the question of
novelty, and particularly the interplay between novelty and imitation, needs to be posed in relation to this new technology.
Instead of pursuing this path, however, I shall here focus on another possibility in comparing the sequential and the connectionist
paradigms. It confronts the consequences of this distinction for media theory and both medium- and media-specific analysis—focusing
at once on artistic medium and technical media.
It is not a new observation—made by, among others, Rosalind
Krauss, Florian Cramer, and Alan Liu—that the concept of “medium”
traverses several disciplines that use it in distinct ways.48 Its two main
disciplines are art history—with “medium“ as the singular and more
often “mediums“ as the plural—and media theory, including the digital humanities—with “medium“ in the singular and “media” in the
plural (although, as Alan Liu has noted, increasingly “media” is also
used in the singular here).49
The first use, in the meaning of an artistic medium such as painting or sculpture, goes back to the eighteenth century, but its importance in the twentieth century is largely due to the influential
art critic Clement Greenberg. Introducing the term “medium specificity,” he argued for the internal differentiation between different
mediums.50 In the 1940s, Greenberg took up an idea of Gotthold
Ephraim Lessing’s, who in his essay Laocoon had already advocated
for the separation of the visual arts from literature according to their
inherent structural logic. While literature in its linear textuality is
inherently temporal, a series in time, and thus most apt to represent action, the visual arts deal with contiguous things in space, that
47. Hans Blumenberg, “‘Imitation of Nature’: Toward a Prehistory of the Idea of the
Creative Being,” in History, Metaphors, Fables: A Hans Blumenberg Reader, ed. Hannes
Bajohr, Florian Fuchs, and Joe Paul Kroll (Ithaca, NY: Cornell U. Press, 2020), pp. 316–
57; Hans Blumenberg, “Paul Valérys möglicher Leonardo da Vinci: Vortrag in der Akademie der Künste in Berlin am 21. April 1966,” Forschungen zu Paul Valéry/Recherches
Valéryennes 25 (2012): 193–227.
48. Rosalind Krauss, “A Voyage on the North Sea”: Art in the Age of the Post-Medium Condition (London: Thames & Hudson, 1999); Florian Cramer, “Nach dem Koitus oder nach
dem Tod? Zur Begriffsverwirrung von ‘Postdigital’, ‘Post-Internet’ und ‘Post-Media,’”
Kunstforum International 242 (2016): 54–67; Alan Liu, Friending the Past: The Sense of
History in the Digital Age (Chicago: U. Chicago Press, 2018).
49. Liu, Friending the Past (above, n. 48), p. 227 n18.
50. Clement Greenberg, “Avantgarde and Kitsch,” in Art and Culture: Critical Essays
(Boston: Beacon, 1989), pp. 3–21; Clement Greenberg, “Towards a Newer Laocoon,” in
The Collected Essays and Criticism, vol. 1, 1939–44 (Chicago: U. Chicago Press, 1986), pp.
23–38.
224
CONFIGURATIONS
is, extension, and thus are better suited for representing objects.51
Greenberg extends this argument to the mediums of the visual arts
themselves, and finds in the most advanced modern art a tendency,
consummated in his own time, towards the separation of painting
and sculpture. For him, as for Lessing, the extent to which a work of
art highlights the specific structural characteristics of its medium is
a measure of its artistic purity. And while Greenberg originally only
wanted to show a process of historical differentiation, medium-specificity eventually took normative rank.52 Thus, if what distinguishes
painting from other mediums is two-dimensionality—flatness—
then those paintings are the purest that are the flattest, i.e., abstract
paintings lacking spatial illusion. Three-dimensionality belongs not
to paintings, but to sculpture.
The second use of the term “medium”—in the meaning of a channel of communication—is mostly connected to a normally unnoticed but determinative carrier of information, as it was introduced
by Marshall McLuhan into media theory. While McLuhan defined
media as human extensions, he nevertheless confined himself to
mass media and electronic media in a narrower sense.53 Contemporary media theory has a tendency to overextend the use of the word
to just about anything that acts as intermediary between two realms.
Because of this, media’s protean nature has fostered the conflation of
mediums and media. But there might be good reason to avoid this
confusion, or at least to insist on the particularity of each media.
In her 2004 essay “Print Is Flat, Code Is Deep,” Katherine Hayles
coined the term “media-specificity”—a nod to Greenberg and his
“medium specificity” already in the title, though Hayles does not
mention his name.54 Media-specific analysis, according to Hayles,
means insisting on the materiality of media. For digital literature, it
entails the acknowledgement that electronic works—in contradistinction to print books—have surface texts, but also underlying code
that shapes those surface texts.
51. Gotthold Emphraim Lessing, Laocoon: An Essay upon the Limits of Painting and Poetry, trans. Ellen Frothingham (Mineola: Dover, 2005), chaps. 15 and 16.
52. For the fact that this normative interpretation is also a result of the reception of his
work—particularly by his pupil Michael Fried, as Greenberg remarked with some annoyance—see Thierry de Duve, Clement Greenberg Between the Lines: Including a Debate
with Clement Greenberg (Chicago: U. Chicago Press, 2010), pp. 147–48. (I thank Colin
Lang for pointing out this passage to me.)
53. Marshall McLuhan, Understanding Media: The Extensions of Man (Cambridge, Mass.:
MIT Press, 1994).
54. N. Katherine Hayles, “Print Is Flat, Code Is Deep: The Importance of Media-Specific
Analysis,” Poetics Today 25:1 (2004): 67–90.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
225
Yet for Hayles the contrastive foil to electronic textuality is still
the printed book—electronic and nonelectronic literature are the
two main operative categories. My distinction between the sequential and the connectionist paradigms indicates, however, that a further internal aesthetic differentiation is necessary, just as Greenberg
extended Lessing’s division between literature and the visual arts to
further subdivide the latter. All approaches that ignore this internal differentiation are not only culpable, in the words of Matthew
Kirschenbaum, of privileging “formal materiality” (logical calculus
and symbol system) over “forensic materiality“ (concrete machine
implementation),55 as is the case with all theories that understand algorithms executed by computers as ultimately identical with manual
rule steps and thus consider, for example, Oulipo processes to be already digital. What is more, if one ignores the difference between the
sequential and the connectionist paradigm because both somehow
happen “with a computer,” one also overlooks that not only the forensic but also the formal materiality deviates from its predecessors:
linear algorithms and neural nets do not even share the same logic
of deterministic rule steps anymore.
Let me give just two examples of the necessity of this further internal subdivision, which shows that the extant theoretical arsenal of
generative literature is exhausted. With the rise of the connectionist
paradigm, it no longer makes sense to speak of what Lev Manovich,
in The Language of New Media, has called the “database logic,” in
which each item has the same significance as any other.56 When
there are no explicitly encoded items anymore that can be accessed
individually, but only statistical dependencies that are distributed
throughout the system, we are confronted not with a database logic
but with something else entirely. Likewise, the distinction between
“texton” and “scripton” introduced by literary scholar Espen Aarseth—that is, a string as it appears in the output, such as on a screen,
and a string as it appears in the code and that may be instantiated
differently—may lose its usefulness if “textons” are no longer to be
located in any code, indeed, if in neural networks there is no code
anymore in the traditional sense.57 The metaphor of “depth” and
55. Matthew Kirschenbaum, Mechanisms: New Media and the Forensic Imagination (Cambridge, MA: MIT Press, 2007), pp. 10–11.
56. Lev Manovich, The Language of New Media (Cambridge, Mass.: MIT Press, 2000),
p. 218.
57. Espen J. Aarseth, Cybertext: Perspectives on Ergotic Literature (Baltimore: Johns Hopkins U. Press, 1997), p. 62. While neural nets still use textual basic elements, they are
single characters of combinations of characters (“byte-pair encoding”). My point is that
such basic elements do not constititute a “database” in any meaningful way.
226
CONFIGURATIONS
“surface” on which Hayles relied, which still implies the possibility of connecting the latter to the former, needs to be reconsidered
in light of the radically different structure of connectionism. Indeed, the connectionist paradigm shatters some of the basic ways
electronic textuality—and digital literature in particular—has been
thought about.
In the remainder of this essay, I concentrate on the implications
this insight has for assessing digital literary works. While Hayles’s
media-specificity forgoes the normative slant of Greenberg’s mediumspecificity, and only describes a way of analysis that takes the particulars of a media into account, I think it might be useful to rekindle
some of that normativity. Megawatt’s significance rests partly in the
way its underlying structure, the linear algorithm, reflects the structure of the resultant text so well. With the connectionist paradigm,
a new form of visual and textual art is emerging, and it is not yet
clear what it might be capable of. But because this is so, the aesthetic
critique of such works may wish to pay special attention to those
that investigate the specificity of their medium, in both senses of
the word.
Medi(a/um)-Specific Category Mistakes
Let me try to give an example of this thought. A reader of this essay may wonder why, in a text about digital literature, I have also
referred so often to visual works. By so doing, I have hinted at the capabilities of the same media—neural nets—working on different mediums—text and images (and here I go back to Lessing rather than to
Greenberg). One can differentiate even further, for not all neural nets
are useful for all mediums. The most basic neural net architectures
for generating images are convolutional neural networks (CNNs),
while recurrent neural networks (RNNs) are used for texts. They work
in different ways due to the structure of what they produce; they are,
in a way, different media generating different mediums (figure 3).
At the most basic level, digital images are continuous in two
dimensions. Their smallest unit is the pixel, with a color value arranged in a matrix that remains static over the data set. The relationship between pixels is based on correlation by proximity. The
closer two pixels are to each other, the likelier it is that they stand
in a meaningful relationship to each other in forming higher-level
wholes. A convolutional neural net uses this logic of continuity in
a bottom-up process to extract features in this pixel matrix by tasking each of the hidden layers with extracting the salient patters of
its previous input. Since this happens progressively between layers,
there is a process of abstraction at work here. The first layer may look
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
227
Figure 3: Convolutional neural network (left), recurrent neural network (right), adapted
from Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans (New York: Farrar,
Straus, and Giroux, 2019).
at a combination of a few pixels, and then pass on the result to the
next layer, which now looks at a combination of a combination of
pixels, and so on. Thus, there is a progression from edges to simple
shapes to objects, and so on.58
Text, on the other hand, requires a different procedure. Unlike
an image formed of pixels, it is not continuous in two dimensions
with equal basic units taking on predetermined values. Rather, text is
continuous in one dimension, and its basic unit is the alphanumeric
character. Ignoring the question of meaning for a moment here, as
well as the fact that the “value” of a character is in no way comparable to the “value” of a pixel, this difference in dimensionality
requires neural nets dealing with text to have a different structure.
Recurrent neural networks need to “remember” previous characters
to build complex statistical models about their likely occurrence,
which is why the neurons of its networks are connected not only to
the next layer but also to themselves (this is the network type used
for Sunspring). A convolutional neural network usually does not deal
with text, while a recurrent one usually is not used for images; here,
medium and media are correlated.
While the reality of neural networks is infinitely more complex—
58. See for this bottom-up process (which resembles the way the optical nerve operates)
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton, “Deep Learning,” Nature 521 (May
2015): 436–44.
228
CONFIGURATIONS
GPT-2 and 3 are based on so-called Transformers that operate in
two directions at the same time, and may require their own medi(a/
um)-specific description59—this dichotomy is nevertheless useful
for offering more finely grained criteria of media-aesthetical judgment. Following this line of thought, digital literature and art can be
discussed along two axes: in terms of their media-specificity—as the
awareness of their technical structures and affordances—but also according to their medium-specificity—as the awareness of the internal
artistic logic of the medium within which they operate. In the sequential paradigm, Megawatt is an example of a parallelism of both:
the structure of the media, the linear algorithm, reflects the structure
of the medium, modernist literature, rather well. But these axes need
not run parallel to each other.
An intelligent illustration of the interplay of the medium/media
axes in the connectionist paradigm is Allison Parrish’s unpronounceable Ahe Thd Yearidy Ti Isa (2019).60 It operates by a deliberate confusion: it treats text as image, reverses the appropriate neural net architectures, and plays with the asemic effects this technological and
semiotic category mistake engenders. Parrish used a specific type of
convolutional neural network called generative adversarial network
(GAN) that has been extremely successful in generating images.61 Its
architecture splits the production and the assessment of its output
into two separate processes. While the “generator” generates images,
the “discriminator” is tasked with judging how close these images
come to the expected output. In this case, the GAN was fed bitmap
images of words. Herein lies the category mistake: bitmaps of words
are human-readable, but not machine-readable; they do not register as text. Thus, the processable information of the image is not
identical to the information that the depicted word represents: its
59. See Ashish Vaswani et al., “Attention is all you need,” ArXiv (June 12, 2017),
https://arxiv.org/abs/1706.037623762; for an instructive discussion of the technical
and political implications of Transformer architectures, see Dieuwertje Luitse and
Wiebke Denkena, “The Great Transformer: Examining the Role of Large Language
Models in the Political Economy of AI,” Big Data & Society 8:2 (2021): 1–14. There is
some evidence that Transformers no longer distinguish between image and text in a
meaningful way, and thus present an entirely new case.
60. Allison Parrish, “Ahe Thd Yearidy Ti Isa (asemic GAN-generated novel),” Github,
Nov. 30, 2019, https://github.com/NaNoGenMo/2019/issues/144. (The novel was an
entry in 2019’s National Novel Generation Month).
61. Ian J. Goodfellow et al., “Generative Adversarial Networks,” Advances in Neural Information Processing Systems (2014): 2672–80. Incidentally, Edmond de Belamy, which
also used a GAN architecture, takes its title from a tongue-in-cheek translation of Ian
Goodfellow’s name.
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
229
Figure 4: Allison Parrish, “Ahe Thd Yearidy Ti Isa (asemic GAN-generated novel),” Github,
Nov. 30, 2019.
technical materiality is separated from its signifying function. The
GAN treats words as images, and that means, not differently from images; thus, the discriminator cannot, like an RNN would, compare
a string of discrete characters, but only statistical distributions of
pixel values. The result looks like text to the discriminator, but lacks
any semantic or even symbolic value, so that Parrish can speak of its
product as an “asemic novel” (figure 4). It represents a nonhuman
type of reading—a probabilistic reading: text-as-images seen through
the eyes of a machine. And in a final twist, as if to comment on the
futility of the whole process, Parrish uses the “correct” image-to-text
process. After all, the book does have a title—Ahe Thd Yearidy Ti Isa.
To create it, Parrish ran the title “image” through a character recognition algorithm that converts bitmaps into text—properly this time,
and even if the result is still nonsensical, this nonsense is now indeed
machine-readable (figure 5).
Ahe Thd Yearidy Ti Isa is a much more interesting use of neural nets
with a much more complex notion of mimesis than either Sunspring,
with its easy absurdism, or Edmond de Belamy, with its naive imitationism. Its game of multiple confusions and conversions draws
attention to the difference of text and image as mediums by highlighting the media used for their processing. As asemic writing, Parrish’s work operates at the border between literature and the visual
arts, and deals in non-semantic but text-like structures. It is mediumspecific precisely in refusing to carry meaning, and media-specific in
230
CONFIGURATIONS
Figure 5: Allison Parrish, “Ahe Thd Yearidy Ti Isa (asemic GANgenerated novel),” Github, Nov. 30, 2019, title page.
reflecting this refusal on a technical level by using a convolutional
neural network where a recurrent neural network would have been
appropriate. This breaks the clear parallelism of Megawatt, and does
so pointedly. One could thus speak of a technologized “ostranenie,”
Viktor Shklovsky’s notion of defamiliarization or “enstranging”
turned towards the underlying generating structure of an artwork.62
In willfully confusing standard procedures, Parrish’s work allows the
only type of “algorithmic empathy” neural nets still allow—not laying bare the underlying concept, but at least offering a glimpse at the
otherwise inscrutable process through tactical, ultimately illuminating category mistakes. Ahe Thd Yearidy Ti Isa, then, does not give into
the Promethean anxiety, but offers a non-anthropocentric use of AI
beyond mere comparison to conventional “human” works.
For a critique of aesthetic AI, which is still a desideratum, this investigation into the inherent possibilities and limitations of a new
medium may offer a normative example. The problem with digital
AI works that simply simulate “human” works is not so much that
they are mere derivatives, simulations of already existing but “analog” schemes. Rather, in insisting on the human comparison, they
restrict from the outset what can be done in this new medium instead of exploring its affordances. In this sense the proposal for both
medium- and media-specificity is meant to be purely corrective. Not
every difference in production needs its own form of criticism—but
where the form of criticism itself remains undeveloped, in that it
views digital works according to the standards of computerized “geniuses,” the concentration on the medium is at least one way to do
justice to the actual novelty of the works.
To be sure, this suggestion has a temporal core and treats works of
this type as pioneering, and fulfilling an avant-garde function. It implies that once this exploration has been exhausted, these artworks
have satisfied their heuristic task—to give way to a new type of literature that can freely make use of the insights gained, and can even
turn away from media- and medium-specificity. Yet in order to get
62. Viktor Shklovsky, “Art as Device,” in Viktor Shklovsky: A Reader, ed. Alexandra Berlina (London: Bloomsbury, 2016), pp. 73–96
Bajohr / Algorithmic Empathy: Toward a Critique of Aesthetic AI
231
to this place, I believe, we do well in taking Hayles’s appeal seriously.
Focusing on the materiality of the connectionist paradigm—even
through paradox and enstranging, as in Parrish’s case—can be an
inspiration both for the analysis as well as the production of contemporary digital literature.