David M. Berry
The phantasmagorias of space to which the flaneur devotes himself find a counterpart in the phantasmagorias of time to which the gambler is addicted. Gambling converts time into a narcotic (Benjamin 2002: 12).
The disruption that Walter Benjamin identified in the gambling halls in the early 20th century finds unexpected resonance in today's computational interfaces. The rise of "prompt engineering" as a new technical practice to shape the outputs of LLMs reveals an engagement with probability and chance that I claim mirrors the temporal and psychological structures that Benjamin identified in his Arcades Project (Das Passagen-Werk). In Benjamin's short fragment, Notes on a Theory of Gambling, for example, there emerges a figure that I want to explore as of particular relevance for our computational moment: the gambler (Benjamin 2006: 297-298). Benjamin's gambler is not merely a person who bets, but a temporal person haunted by chance, perpetually suspended between possibility and ruin. As Marasco argues,
The critique of the gambler is as much a way to honor the dream-wish harbored in aleatory experience as to decipher its historical shape. It is an attempt to do justice to a structure of desire that is itself produced by structures of power and possibility. Benjamin’s position is refreshingly ambivalent: the allegory of the gambler is an indication of ideology and insurgency, bourgeois mendacity and revolutionary vitality, a conservative desire to keep the present order as is and the radical urge for its undoing. It is both at once. Perhaps this explains some of its distinctive charms (Marasco 2018: 5).
One of the interesting challenges of working with contemporary LLM systems is that due to their reliance on "prompt salting," that is, the way stochastic sampling processes introduce chance elements into generation, such that the outputs are produced semi-randomly.
[1] This makes the production of outputs less a science than a product of chance. Users of systems like GPT-4, Stable Diffusion, and DALL-E find themselves having to develop elaborate strategies for maximising desired outputs whilst acknowledging the unpredictability of generative artificial intelligence. These strategies could, therefore, be said to exhibit the same mixture of systematic analysis and aleatory hope that Benjamin argues characterises gambling psychology, creating a kind of replacement to rational design processes or software engineering and replacing it with "vibe coding," an affective coding practice that manifests as a kind of computational gaming culture organised around the optimisation of prompts against probabilistic systems.
This computational gambling psychology can be more easily observed through what I term "probabilistic media," that is cultural artefacts produced through stochastic computational processes that operate on probability distributions rather than deterministic rules.
[2] Using these LLM systems, users must craft prompts within the constrained interface of text boxes or parameter settings, making decisions about keywords, syntax, and formatting that will determine the generation process. Like Benjamin's gambler, who places bets "at the very last moment—the moment, moreover, when only enough room remains for a purely reflexive move," prompt engineers operate within temporal constraints that compress deliberation to its absolute minimum (Benjamin 2002: 513). The send button functions rather like the dice throw, a moment of commitment that transforms careful preparation into an aleatory outcome. Benjamin's gambler therefore appears to gain renewed theoretical purchase for thinking through these technologies. This temporal dimension reveals itself most acutely in the psychological pressures that prompt engineering generates.
"The particular danger that threatens the gambler," Benjamin writes, "lies in the fateful category of arriving 'too late,' of having 'missed the opportunity'" (quoted in Marasco 2018: 2). This figure operates through what Benjamin calls purely reflexive moves, responses that occur at the threshold where calculation meets catastrophe, where the systematic meets the aleatory. The gambler's relation to time is therefore disruptive; where each wager represents not a step in a progressive sequence but a moment that "kills time" itself, collapsing futurity into the eternal present of the dice roll.
This temporal compression generates what I want to explore through the concept of "prompt anxiety."[3] I argue this is a distinctive psychological condition where users experience each output from the LLM as a wager against unknown algorithmic probabilities. We see online forums dedicated to prompt engineering which often reveal users developing obsessive strategies for controlling outputs that remain uncontrollable. Consider the elaborate "prompt libraries" that have emerged around LLM generation, featuring thousands of tested combinations designed to produce particular aesthetic effects (see Oppenlaender 2022). These libraries can be seen as similar to gambling systems, promising users greater control over outcomes whilst acknowledging that no prompt can guarantee specific results. The psychological mechanisms underlying these practices might therefore reveal deeper structural parallels with gambling behaviour.
This psychological dimension of prompt engineering perhaps exhibits what Richard Hofstadter identified in political discourse as "the paranoid style." This is a tendency toward elaborate interpretative schemes designed to explain apparently random outcomes through hidden systematic patterns (Hofstadter 1966). Users tend to develop complex theories about "hidden biases" in training data, "prompt injection" techniques, and "model alignment" issues that supposedly explain why their carefully crafted prompts produce unexpected outputs. These theories perhaps serve a similar psychological function to gambling superstitions, providing the illusion of systematic control over processes that remain stochastic. This "paranoid" interpretative tendency finds material reinforcement in the technical architecture of LLM systems themselves.
I argue that the technical architecture of these systems actively encourages this paranoid style through what developers call "prompt sensitivity" (Zhuo, J. et al 2024). Prompt sensitivity describes how minor alterations in input, such as adding or removing a single word, can produce dramatically different outputs, transforming realistic portraits into abstract art or factual explanations into creative fiction. This sensitivity creates what users experience as "prompt paranoia," what Best describes as "a new kind of writers block" (Best 2024). Here, every word choice becomes fraught with potential consequences that cannot be predicted in advance.
Users therefore begin to develop elaborate techniques for prompt debugging, by testing multiple variations to identify which elements produce desired effects, creating a form of systematic superstition around text manipulation. Thus configured, the computational interface becomes a site where magical thinking emerges, representing a return to pre-modern epistemologies within ostensibly rational technological systems, where correlation becomes causation and pattern recognition becomes predictive power. This return to magical thinking within computational systems reveals how deeply the gambler's psychology has penetrated our technological interactions, where systematic analysis gives way to superstitious practice. Here, the computational and the ludic meet, revealing how LLM interaction shares the gambler's relationship to chance, calculation, and temporal compression.
The figure of Benjamin's gambler therefore provides a revealing metaphor for understanding the psychological and temporal dimensions of prompt engineering. As users navigate the probabilistic landscape of LLMs, they embody a similar mix of calculation and hope that Benjamin identified in the arcade's gambling halls. As Marasco observes, "with the collapse of rational faith in historical progress and the sinking feeling of domination as destiny, aleatory experience restores vitality to political life and indeterminacy to insurgent forces. It is in his courtship with destruction that the gambler (and the banker) becomes an accidental accomplice to the revolutionary critic" (Marasco 2018: 20). Perhaps, our computational moment may be thereby characterised not by rational design but by a return to chance-based epistemologies, where the gambler's temporality, suspended between possibility and ruin, becomes a new, potentially revolutionary, mode of technological engagement.
** Headline image generated using DALL-E 3 in June 2025. The prompt used was: "An oil painting portrays four people from diverse ethnic backgrounds engaged with digital gambling apps on their devices against a warm, muted beige wall. The individuals—each exhibiting concentrated expressions—are focused on their phones, tablets, laptops, and smartwatches, highlighting meticulous details in skin, hair, and devices, with soft, warm lighting enhancing the scene's depth and emotional balance." Due to the probabilistic way in which these images are generated, future images generated using this prompt are unlikely to be the same as this version.
Notes
[1] I use the term "prompt salting" to describe the stochastic sampling processes built into large language models that introduce chance elements into AI generation, making outputs semi-random rather than deterministic. This means that identical prompts can produce different results, transforming what might otherwise be systematic software engineering into what is sometimes called "vibe coding." This I see as an affective coding practice that resembles computational gaming culture organised around optimising prompts against probabilistic systems. I use the metaphor of "salting" drawing from the technical process of random seeding used in cryptographic functions. In LLMs, contingency is "sprinkled" throughout the generation process, making outcomes unpredictable despite careful prompt writing. I argue that this technical process of prompt salting is actually a strategic design that platforms utilitise, as it creates the psychological conditions necessary for prompt anxiety. Unlike traditional deterministic software where inputs reliably produce predictable outputs, prompt salting ensures that interaction with AI systems produces uncertain outputs.
[2] Elsewhere I use the concept of "stochastic labour" to describe a qualitatively new form of cognitive work that emerges from human interaction with probabilistic artificial intelligence systems (Berry, forthcoming). Unlike forms of labour organised around predictable processes or even post-Fordist flexible specialisation, stochastic labour requires workers to synchronise with the unpredictable rhythms of AI generation. This labour involves developing elaborate psychological and emotional strategies for managing the anxiety, frustration, and cognitive load generated by unreliable algorithmic systems that produce semi-random outputs through, for example, "prompt salting" (stochastic sampling processes) (see footnote 1, above). Workers must, therefore, engage in continuous interpretation and response to algorithmic uncertainty rather than applying their technical skills directly creating hybrid human-machine assemblages where cognitive capacity becomes subordinated to probabilistic systems. I argue that this gestures toward a transformation in the temporal structure of cognitive work itself, moving from standardised procedures to a form of "stochastic temporality" where time consciousness is organised around managing uncertain outcomes. Examples include programmers using AI coding assistants, like Cursor, who must verify and iterate through multiple loops of generated code (so called "software 3.0"), or creative workers structuring entire workflows around the unpredictable rhythms of AI generation (see Trades Union Congress 2025). Stochastic labour intensifies cognitive load whilst systematically hiding the social cooperation that makes algorithmic intelligence possible within platform capitalism.
[3] The concept of "prompt anxiety" I develop in this article is inspired by Bernstein's notion of "Cartesian anxiety" (Bernstein 1983). Bernstein describes Cartesian anxiety as an ontological condition where we oscillate between the search for fixed, permanent foundations for knowledge and the terror of relativistic chaos. With prompt anxiety the user oscillates similarly with determinism and uncertainty. As Bernstein writes, "either there is some support for our being, a fixed foundation for our knowledge, or we cannot escape the forces of darkness that envelop us with madness, with intellectual and moral chaos" (Bernstein 1983. 18). This anxiety, Bernstein argues, underlies contemporary philosophical debates between objectivists and relativists, creating what he terms a "grand and seductive Either/Or" that must be transcended rather than resolved through taking sides (Bernstein 1983, 18). As he argues, "reading the Meditations as a journey of the soul helps us to appreciate that Descartes' search for a foundation or Archimedean point is more than a device to solve metaphysical and epistemological problems. It is the quest for some fixed point, some stable rock upon which we can secure our lives against the vicissitudes that constantly threaten us" (Bernstein 1983, 170. In adapting this framework to computation, "prompt anxiety" I argue that it creates an epistemological condition wherein the fundamental structure of knowledge production through human-AI interaction similarly oscillates between systematic control and aleatory uncertainty, transforming each LLM output into a site where the possibility of rational knowledge confronts the limits of computational determinacy. However, it is crucial to note that prompt anxiety describes not just new ways of knowing, but also new ways of being technological subjects. In trying to transcend this Cartesian anxiety, Bernstein's develops a praxis which involves moving beyond theoretical Either/Or frameworks toward practical engagement with concrete situations. Praxis in relation to prompt anxiety would similarly require practical wisdom about when and how to engage with AI systems, developing computational phronesis. This represents a form of practical wisdom that accepts uncertainty as a condition of AI engagement rather than a problem to be avoided or excluded. For example, it might include developing work patterns that treat AI as a collaborator (or co-writer) rather than a tool to be controlled or "applied."
Bibliography
Benjamin, W. (2002) The Arcades Project. Harvard University Press.
Benjamin, W. (2006) Walter Benjamin: 1927-1930 v. 2, Pt. 1: Selected Writings. Harvard University Press.
Bernstein, R.J. (1983) Beyond Objectivism and Relativism: Science, Hermeneutics, and Praxis. University of Pennsylvania Press.
Berry, D.M. (forthcoming) ‘Prompt Anxiety and the Algorithmic Politics of Uncertainty’.
Best, A. (2024) ‘Prompt Engineering? What I HATE about it…’,
Medium, 9 March. Available at:
https://medium.com/@andrew_best/prompt-engineering-what-i-hate-about-it-7b7dc1ada106 (Accessed: 6 June 2025).
Hofstadter, R. (1966) The paranoid style in American politics, and other essays. Jonathan Cape.
Zhuo, J.
et al. (2024) ‘ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs’, in Y. Al-Onaizan, M. Bansal, and Y.-N. Chen (eds)
Findings of the Association for Computational Linguistics: EMNLP 2024.
Findings 2024, Miami, Florida, USA: Association for Computational Linguistics, pp. 1950–1976. Available at:
https://doi.org/10.18653/v1/2024.findings-emnlp.108.
Marasco, R. (2018) ‘It’s All about the Benjamins: Considerations on the Gambler as a Political Type’,
New German Critique, 45(1 (133)), pp. 1–22. Available at:
https://doi.org/10.1215/0094033X-4269826.
Comments
Post a Comment