Material, ĠConfig, urations, Ġof, ĠAI, -, Gener, ated, ĠText, s

The Question

When you type a question into ChatGPT and press enter, what actually happens? Not in the hand-wavy sense of "it predicts the next word", but materially, mathematically, step by step. And what does that process mean for how we understand language, meaning, and authorship?

This paper argues that AI-generated texts are neither meaningless statistical artifacts nor uncanny simulations of human writing. They are something else entirely: distinctive material-semiotic practices that emerge from specific computational configurations, and that demand their own theoretical framework.

The Approach

The paper takes GPT-2 as its case study — not because it is the most powerful model, but because its open-source architecture makes genuine technical examination possible, something OpenAI's later models no longer allow. Through close analysis of its computational processes: tokenization, vector embeddings, positional encoding, transformer attention mechanisms, and probabilistic decoding, the paper maps the specific material conditions through which AI text emerges.

This technical analysis is then read through the lens of new materialism, drawing on Coole and Frost's account of matter's inherent agency, Hayles' media-specific criticism, and Bajohr's concept of "dumb meaning", a form of significance that operates below human cognitive processes but is not therefore empty. The paper argues that meaning in AI text generation is not absent but differently configured: emerging through geometric relationships in vector space, probabilistic patterns of token selection, and the dynamic interaction of input with materialized knowledge.

The Intervention

Two dominant positions frame most debates about AI-generated text. The first dismisses it as "stochastic parrots": form without meaning, statistical noise dressed as language. The second overclaims, reading AI systems through Barthesian and Derridean frameworks in ways that flatten the very real differences between human and machine textual production.

This paper rejects both. Instead of asking whether AI-generated text is "real" language, it asks what kind of material object it actually is and what theoretical tools we need to examine it on its own terms. The answer requires staying close to the specific material processes of a specific model at a specific moment, rather than generalizing across "AI" as though it were a stable category.

The paper also takes seriously the material constraints of these systems: how the biases encoded in training data are not merely a data quality problem but a consequence of how social relations become materialized into fixed computational configurations, shaping what can be expressed and foreclosing alternatives at every step of generation.