In A Mathematical Theory of Communication (1948), Claude Shannon drew an explicit boundary:
“The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.”
Shannon was not dismissing meaning. He was defining the limits of his mathematical apparatus. Information, in Shannon's sense, does not measure significance or truth. It measures the probabilistic reduction of uncertainty across an alphabet: the capacity of a physical channel to transmit bits, whether the sequence is a poem, an operational order, or thermal noise.
The Statistical Channel
Large Language Models are engineering implementations of Shannon's framework. Text is broken into tokens, mapped to vectors in high-dimensional space, and weighted by attention matrices. The training objective is statistical: minimizing cross-entropy loss by estimating conditional probabilities over token sequences:
P(wt | w1, w2, ..., wt-1)
The model predicts sequences with high regularity. Because the generated text is syntactically coherent and fluent, readers often assume that semantic reference and understanding exist inside the model. This is an old category error: confusing the efficiency of a transmission channel with the presence of an understanding subject.
Epistemological Distinctions
Three distinctions clarify the boundary between automated statistical syntax and human understanding:
1. Information and Norms
As Georges Canguilhem argued regarding living organisms and technical objects, genuine knowledge is normative. It develops through the physical confrontation of an organism with its environment. Information reduces variance within an established code; knowledge involves setting norms and correcting errors in practice. Calling model deviations "hallucinations" borrows psychological vocabulary for what is actually statistical variance unmoored from perception.
2. Grammatized Memory
Following Bernard Stiegler, technical objects externalize memory into material forms (tertiary retention). Writing, print, and digital databases record symbolic traces. An LLM is an automated engine of externalized text. It parses and recombines traces left by past human writing. The model does not speak; it computes probabilities over an archive.
3. Meaning Arrives with the Reader
Weights and tensors contain parameters, not intentions. A transformer does not refer to objects in the world. Semantic reference occurs when an embodied reader interprets the output, bringing personal context and judgment to the text.
Summary
Treating technical systems as thinking agents is not new; eighteenth-century automatons provoked similar reactions. But projecting intentionality onto next-token predictors obscures how these systems work. An LLM is a tool for statistical language processing. The responsibility for verification, meaning, and truth remains with the human reader.