The Grammatizator
“Mr Bohlen,” Adolph Knipe said gravely, “do you realize that at this moment, with your little finger alone, you have it in your power to become the most versatile writer on this continent?” […] He reached up and pulled a switch on the panel. Immediately, the room was filled with a loud humming noise, and a crackling of electric sparks, and the jingle of many, tiny, quickly moving levers; and almost in the same instant, sheets of quarto paper began sliding out from a slot to the right of the control panel and dropping into a basket below. They came out quick, one sheet a second, and in less than half a minute it was all over. The sheets stopped coming. “That’s it!” Adolph Knipe cried. “There’s your story!”
In Roald Dahl’s 1954 satire The Great Automatic Grammatizator, one Adolph Knipe invents a machine which allows him to generate novels and short stories mechanically with the push of a button. Knipe is himself a writer, a fact that he is somewhat embarrassed to admit to his superior when pitching the machine. When asked for the reason for embarking on such a useless endeavor, he calls it the passion, the “creative urge.” And yet, when none of the literary magazines he keeps applying to wants to publish his stories, he makes use of his engineering talents and invents the Grammatizator, which is described as resembling “the instrument panel of some enormous aeroplane,” and ultimately provides the excited initiator with a custom-made story for any type of publication (Dahl 1986).
The thought itself seems to be coherent; why should numbers, as a self-referential system of formal representation, not be able to represent letters, a self-referential system of formal representation? Representation, however, is not the magic ingredient. During the machine’s development, Knipe confronts a problem that was already famously articulated by Ada Lovelace about Charles Babbage’s analytical engine: a machine cannot take us by surprise or produce original thought (Lovelace 2009).
Roald Dahl skillfully brushes over this topic by letting Knipe endow his machine with a memory capable of storing linguistic elements encoded as electrical impulses and arranged for extraction as needed; the details of these extractions are not given any greater description. And still, Dahl’s narrative of course in many ways anticipates contemporary debates surrounding generative artificial intelligence (genAI) and the way its synthetic artifacts propagate through our digital environments.
Avery Slater, in her 2020 contribution to The Oxford Handbook of Ethics of AI, observes that Lovelace’s original objection related not only to what a machine can and cannot do in the context of originality, but also to the ontological significance of that originality (Slater 2020). For Lovelace, mathematics was the instrument through which “the weak mind of man can most effectually read his Creator’s works” (Lovelace 2009, p. 18). It is the machine that, in contrast to us, speaks the true language of the universe, and anything it produces can only be seen as lying closer to that truth than any humanly made creative artifact. While this view on mathematics might not have been unusual at the time, when looking at today’s AI slop images, it is hard to imagine how Shrimp Jesus or a weirdly smooth-looking bulldog in police attire should be in any connection to a greater truth about the universe. If mathematics is the language of nature itself, then its output is a lot less elegant than one might expect.
A (pseudo) novel order
To better understand the computational processes that underpin these generative systems, it is worth revisiting the work of Max Bense, who, together with Abraham Moles, articulated the theory of Information Aesthetics (Bense 1965). Bense and Moles aimed to establish a scientific, objective, and measurable aesthetic theory that moves beyond subjective judgments of taste, and were deeply influenced by the same intellectual currents, including cybernetics, Shannon’s information theory, and semiotics, that shaped the early development of artificial neural networks (Klütsch 2023).
For Bense, the deliberately compositional character of an artwork is fundamental; it presupposes the decomposability (“Zerlegbarkeit”) into elementary aesthetic signs (Bense 1965, p. 44). Drawing on the mathematician George David Birkhoff’s work, Bense formulates aesthetic value as the ratio of the “complexity” (C) and “order” (O) of an artistic object. This positions aesthetic information as fundamentally characterized by the unexpected. For Bense, this unpredictability arises out of the creation of improbable order out of disorder. He contrasts this process with natural phenomena, which, as we know from thermodynamics, move in their physical properties continuously toward a state of entropy (Bense 1965). By contrast, art objects, which are made or produced through deliberate human activity, strive toward creating an order, an “aesthetic negentropy” (Bense 1965, p. 153).
Although rarely quoted anymore, the influence of Information Aesthetics is still clearly visible when one looks at a prevalent definition of creativity within current AI research, put forward by Jürgen Schmidhuber, who has been an influential figure in the field for decades. For him, a system is creative if it is able to recognize or manipulate “novel patterns, that is, data predictable or compressible in hitherto unknown ways,” that is, making order from disorder (Schmidhuber 2010, p. 145).
It is hard to tell what Bense would have made of generative AI systems and the conditions of ordering in which they produce their output. This process, which Martin Zeilinger describes as the dissolution of imitation into unexpected outputs, can indeed lead to results that could pass as original and even creative (Zeilinger 2021, p. 86). The crux of the dynamic, however, becomes apparent when looking closely at the computational processes that underpin generative systems and how they achieve the impression of the unexpected.
Synthetic contingency
For the largest part of their respective pipelines, generative architectures such as GANs or Transformers are completely deterministic in their way of inference. From the moment in which the inference process starts up until the preliminary output, the system would, for any given input, always produce the same output (ceteris paribus) from which the final, compositional output is sampled. There are several parameters and selection processes that decide which token gets picked, but all sampling-based methods are powered by a pseudorandom number generator (PRNG1). If one keeps the seed of the PRNG constant, then the output will remain static. Here too, vocabulary from thermodynamics is used: the variance of the distribution from which the final token is sampled is called the temperature. In user interfaces and popular literature, this variable is often referred to as the creativity of a system.
If in the artistic process, as Bense argues, this unpredictability of information relates to creating order from a chaotic backdrop (Bense 1965), within today’s infrastructure of machine generation it seems to be the other way around: creating here means to sample at random from the most orderly backdrop imaginable2. There is a certain humor in this dynamic that would no doubt not be lost on Roald Dahl readers: structurally excluding any ambiguity in the process of data cleaning, training, retraining, and inference, only to introduce it in the very end via a technical proxy.
Already in ancient Greek times, the goddess of fortune, Tyché, was associated with the arts: “Tyché loves Techné and Techné loves Tyché,” as Aristotle quotes the Greek poet Agathon (Schirren 2008, p. 90). Tyché is fleeting and unpredictable; she will come when one least expects it and stay away when one tries to will her into existence. Even the ancient Greeks, however, already made a distinction between Tyché and what is often translated as spontaneity (“automaton”). In Aristotle’s Physics, Tyché is a subcategory of automaton, which is only applicable to “agents that are capable of good fortune and of moral action generally,” whereas automaton can also be found in animals and inanimate objects (Physics II.4, 196a5–10). Dieter Mersch, in reference to Lacan, phrases the dichotomy between Tyché and automaton in the arts as a struggle for making the unavailable available (Mersch 2008). Tyché, like Kairos too, cannot be summoned at will; it is entirely beyond our control. Automaton, on the other hand, is precisely the controllable equivalent. Like the PRNG, it is reproducible, controllable, and thus available. For Mersch, Tyché provokes the will to appropriate this unexpected (Mersch 2008). In this sense, randomness as employed in the sampling process of genAI architectures becomes the final frontier of unpredictability within a system built on the fantasy of total availability.
Sell them wholesale!
Circling back to Adolph Knipe, it is clear that the invention of the Great Automatic Grammatizator derives from similar motivations. As writer and programmer Martin Paul Eve rightly notes, the “Grammatizator” is not actually about computerized writing or creativity (Eve 2017). The Grammatizator was never meant to create; it was meant to sell stories to literary magazines.
Indeed, we can imagine the Grammatizator’s consistently mediocre output to be quite marketable: cheap, fast, and predictable. Today, in the age of personalized chatbots and AI influencers, we see many of the same characteristics (see Walter 2024). Combined into a conglomerate of AI-powered content farms, these agents are assumed to take up the bigger part of the internet by now (Stanusch et al. 2025). So far, it reveals itself in tiny details, like the morphing between images, or the repeated use of em dashes in written text. But there is a different type of recognizable feature that is harder to pinpoint: a weirdly sterile aesthetic that might, in the end, have its source in the logical and mathematical character of computational aesthetics (see Fazi and Fuller 2016), which finds its completion in the drive to render uncertainty itself operational. If both the cost of production and the cost of distribution of new content become as minimal as they have in an age of genAI and reels, the new is not scarce anymore, and by virtue of the most basic economic principles it thus loses its value.
In this way, Dahl anticipates what Slater phrases as creativity problematizing production (Slater 2020). Turned into consumables, the generation of synthetic content reveals the premise that art is a way of human productivity and thus a form of reproducible labor. This ambition has left us with countless Grammatizators, each optimized not for literature but for attention. What we are left with are synthetic automata at industrial scale, and the flood of their artifacts. Looking at how contingency is formalized and employed in the generation of synthetic content by these automata, its abundance should be read less as the promise of computational originality than as the visible residue of the paradox of controllable creation. Adolph Knipe said it first: “And stories, well, they’re just another product, like carpets and chairs, and no one cares how you produce them so long as you deliver the goods. We’ll sell them wholesale, Mr Bohlen!”
References
Aristotle. (2009). Physics (R. P. Hardie & R. K. Gaye, Trans.). The Internet Classics Archive. http://classics.mit.edu/Aristotle/physics.html
Bense, M. (1965). Aesthetica: Einführung in die neue Aesthetik. AGIS-Verlag.
Dahl, R. (1986). The Great Automatic Grammatizator: Vol. Someone Like You. Penguin.
Eve, M. P. (2017). The Great Automatic Grammatizator: Writing, labour, computers. Critical Quarterly, 59(3), 39–54. https://doi.org/10.1111/criq.12359
Fazi, M. B., & Fuller, M. (2016). Computational Aesthetics. In C. Paul (Ed.), A Companion to Digital Art (1st edn, pp. 281–296). Wiley. https://doi.org/10.1002/9781118475249.ch11
Gitelman, L. (2013). ‘Raw data’ is an oxymoron. MIT press.
Klütsch, C. (2023). Information Aesthetics and the Stuttgart School. In H. B. Higgins & D. Kahn (Eds), Mainframe Experimentalism: Early Computing and the Foundations of the Digital Arts (pp. 65–89). University of California Press.
Lovelace, A. (2009). Sketch of the Analytical Engine. In L. Otis (Ed.), Literature and Science in the Nineteenth Century: An Anthology (pp. 15–19). Oxford University Press.
Mersch, D. (2008). Spiele des Zufalls und der Emergenz. Maske und Kothurn, 54(4), 19–34. https://doi.org/10.7767/muk.2008.54.4.19
Schirren, T. (2008). Techne liebt Tyche und Tyche Techne. In A. Arweiler & M. Möller (Eds), Vom Selbst-Verständnis in Antike und Neuzeit. Walter de Gruyter.
Schmidhuber, J. (2010). Artificial scientists & artists based on the formal theory of creativity. In E. B. Baum, M.
Hutter, & E. Kitzelmann (Eds.), Proceedings of the 3rd Conference on Artificial General Intelligence:
Advances in Intelligent Systems Research (Vol. 10, pp. 145–150). Atlantis Press. https://doi.org/10.2991/agi.2010.32
Stanusch, N., Degeling, M., Romano, S., Çetin, R. B., Schüler, M., & Semenzin, S. (2025). AI-Generated Algorithmic Virality (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2508.01042
Slater, A. (2020). Automating Origination: Perspectives from the Humanities. In M. D. Dubber, F. Pasquale, & S.
Das (Eds), The Oxford Handbook of Ethics of AI (pp. 520–537). Oxford University Press. https://doi.org/10.1093/oxfordhb/9780190067397.013.32
Walter, Y. (2025). Artificial influencers and the dead internet theory. AI & SOCIETY, 40(1), 239–240. https://doi.org/10.1007/s00146-023-01857-0
Zeilinger, M. (2021). Tactical Entanglements: AI Art, Creative Agency, and the Limits of Intellectual Property. meson press.