In his 2023 e-flux essay From Writing to Prompting: AI as Zeitgeist-Machine, Groys advanced the claim that artificial intelligence is neither a creative agent nor a cognitive tool but the compressed, computable form of a civilization's accumulated cultural production. The large language model does not think. It processes the archive and produces outputs that reflect the statistical structure of that archive. The output is not the product of a mind; it is the product of a culture. This reframing transforms the act of prompting from a technical operation into a diagnostic one: the prompter who writes a query and reads the response is not using a tool — she is interrogating the zeitgeist, asking what the accumulated mass of human thought produces when subjected to specific pressure.
The diagnostic function of AI is, in Groys's analysis, more interesting and more consequential than its productive function. Everyone has noticed that AI can produce. Groys notices that AI can reveal. The machine's outputs are not merely useful artifacts but cultural documents — evidence of the zeitgeist's structure, data points in the ongoing analysis of what a civilization thinks, values, assumes, and systematically ignores. The practical implication is that the most valuable use of AI may not be making things but making visible what the culture has rendered invisible: the dominant patterns, the buried assumptions, the characteristic blind spots of an entire historical moment.
The zeitgeist, however, is not a unified voice. In his 2024 essay on the Sorokin exhibition, Groys insists that it is monstrous: full of ruptures and inner contradictions, carrying dark and violent aspects alongside its visible surface. The machine trained on this monstrous totality reproduces its contradictions. It can produce feminist theory and misogynist trolling, scientific rigor and conspiratorial fantasy, profound insight and confident nonsense — all with the same authoritative, smooth surface. The smoothness conceals the monstrosity. The polish hides the ruptures. The user encountering only the polished surface never sees the monstrous totality from which the output was extracted.
For the concept of novelty, the zeitgeist-machine has radical implications. If AI output expresses a cultural totality rather than an individual creator, then the new — in relation to AI output — is not what differs from prior individual works. It is what differs from the totality itself: what the zeitgeist, in all its accumulated mass, has not yet produced. Finding genuine novelty requires counter-prompting: pushing the machine against the grain of its training, forcing it toward configurations its statistical structure resists, exploiting the gaps between what the culture has produced and what it has not yet imagined.
The concept also transforms the question of AI authorship. The text that emerges from Claude is not Claude's opinion. It is the archive speaking through a specific query. This does not eliminate the human contribution — the prompter selects the query, evaluates the response, takes responsibility for what she circulates. But it relocates the intelligence. The intelligence is in the archive, compressed and made operational. The human's work is diagnostic rather than creative, curatorial rather than productive, interpretive rather than expressive.
Groys developed the zeitgeist-machine concept in a series of lectures and essays between 2022 and 2024, culminating in the e-flux essay that gave the concept its name. The framework draws explicitly on Hegel's Zeitgeist — the spirit of the age as an objective cultural force that exceeds any individual consciousness — and extends it to machines that process culture at a scale no individual can match.
AI is the archive made operational. The model compresses a civilization's accumulated cultural production into a form that can be queried, diagnosed, and analyzed.
Prompting is cultural diagnostics. The query interrogates the zeitgeist; the response reveals the structure of the culture that produced the training corpus.
The zeitgeist is monstrous. It contains contradictions, exclusions, and violence that the smooth surface of the output conceals.
Novelty requires counter-prompting. Genuine newness lies in the gaps the zeitgeist has not yet filled, accessible only by pressure against the grain of the model's statistical tendencies.
The framework has been contested by cognitive scientists who argue that LLMs do perform novel inference rather than mere reproduction of training patterns. Groys's reply is that the dispute is a category error: whether or not the model generates novel outputs, its outputs are always statistical expressions of the archive it was trained on, and the interesting question is what those expressions reveal about the archive.