The Flickering Signifier

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI identifies as one of its central hazards the AI text that “sounds better than it thinks”—prose that carries all the surface markers of insight while concealing errors that only a reader with deep domain knowledge and the discipline to resist fluency can detect. This is the flickering signifier in operation. When a large language model generates a passage connecting two philosophical concepts with apparent elegance, and the connection is subtly but seriously wrong, the error is not a random glitch. It is a structural consequence of the medium: the generation process interpolates across its training data, finds a statistically plausible connection, and produces it with the same confidence it would produce a correct one. The signifier flickered into a configuration that looks like insight and is not.

The silent middle of the AI transition—those who hold exhilaration and loss simultaneously without being able to collapse into either triumphalism or despair—is, in the terms of the flickering signifier, the population most aware of the epistemological condition of AI text. They have encountered enough flickering signifiers to distrust fluency as evidence of authority. They have also encountered enough genuine insight to recognize that the flickering does not render the output worthless. Their silence is the difficulty of maintaining both truths at once in a discourse that rewards certainty.

Origin

Hayles introduced the concept in her foundational 1993 essay “Virtual Bodies and Flickering Signifiers” and developed it across How We Became Posthuman (1999) and subsequent work on electronic literature. The concept extends Ferdinand de Saussure’s structural linguistics into the digital age while departing from it: where the printed signifier is relatively stable across time—a particular configuration of marks that persists—the digital signifier is produced anew on each rendering, a performance rather than a record. In poststructuralism, the signifier was already understood as unstable in meaning; Hayles adds a further dimension of material instability. The AI moment has vindicated the concept beyond what Hayles’s earliest formulations could have anticipated. A language model does not merely display text differently on different screens; it generates different text for different instances of the same prompt, and the generation is governed by statistical processes that produce plausible language rather than true language, coherent language rather than accurate language.

The concept intersects with Hayles’s broader re-embodiment thesis: the reason the flickering signifier matters is precisely because information is not substrate-independent. If information were purely abstract pattern, the stochastic variability of AI generation would be a technical detail rather than an epistemological crisis. Because the substrate shapes the meaning, the specific material properties of the AI substrate—its optimization function, its training distribution, its tendency toward statistical centers—shape the meaning of every output it produces.

Key Ideas

Stochastic production vs. authorial choice. The printed text carries the weight of necessity: the author chose these words, and the choosing excludes alternatives and reveals priorities. The AI-generated text does not carry this weight. The model does not choose in the relevant sense. It generates—selects from a probability distribution—and the output does not reveal anything about the machine’s values or sensibilities, because the machine has parameters rather than values. The builder who imports this output without re-imposing choice produces work that reflects the substrate’s tendencies rather than her own intentions. The craft of AI-augmented work is therefore not production but selection—analogous to the photographer’s selection from the configurations of light the world presents.

Fluency without sincerity. The implicit truth-claim of human authorship—that these words represent considered judgment, that the author believes what she says—is absent from AI text. The model cannot distinguish between what it knows and what it is interpolating, cannot flag its own uncertainty the way a human author qualifies a claim she is unsure of. The confidence is uniform, the fluency consistent, and the decorrelation of fluency from authority means that surface coherence is not evidence of epistemic reliability.

The gravitational pull toward the statistical center. The AI substrate does not amplify all signals equally. It amplifies the signals that align with its training distribution—the conventional thought, the familiar argument, the well-worn connection—while attenuating the genuinely novel. The builder who wants original work must work against the substrate’s tendencies, must insist on the rough and unfamiliar when the smooth and conventional is readily available. Material literacy is the discipline of detecting and counteracting this gravitational pull.

Material literacy as corrective. The reader who possesses material literacy evaluates AI output as what it is: the product of a stochastic generation process. She reads for the characteristic patterns of the substrate—the smoothness that signals interpolation rather than understanding, the confident assertion that signals statistical likelihood rather than epistemic authority, the elegant connection that may reflect structural similarity or may reflect the model’s tendency to find connections where none exist. The parity principle of the extended mind suggests that AI processing counts as cognitive when it functions as cognitive processes would; the flickering signifier insists that it counts as cognitive while remaining epistemically different—and that the difference demands a different mode of reading.