Large Language Models and the Patterns of Human Language Use

Christoph Durt and Thomas Fuchs’s chapter, published on November 21, 2024, asks how language models can produce text that people find meaningful. Their answer relates three kinds of pattern: numerical regularities learned by a model, regularities in human language use, and the lived experience from which that use arises. The chapter develops a philosophical explanation of meaningful output without taking conversational fluency as evidence of sentience.^[1]

What the three kinds of pattern explain

Statistical patterns concern relations between tokens. Patterns of use concern what people do with language in communicative situations. Patterns of experience concern the situations, expectations and activities in which people speak and interpret. The authors argue that a text corpus carries partial traces of those activities. A model can recombine these traces into new, intelligible text without sharing the experience that gave rise to them.^[1]

Their comparison with a glove makes the relation concrete: accumulated language practices shape the possibilities of expression, while each speaker uses and modifies those possibilities. On their account, speaking also develops an initially incomplete thought; it is not simply the delivery of a fully prepared internal sentence. This is why they resist reducing human meaning either to reference alone or to statistical distribution alone.^[1]

Convincing language and attributed experience

The chapter uses LaMDA’s reported fear of shutdown alongside the familiar frightened computer in 2001: A Space Odyssey to discuss the force of clichés and confessional language. That comparison is an interpretation of why a reader might hear a personal admission, not an identified training-data trace or experiment establishing what caused a particular reply. It gives a linguistic perspective on the ELIZA Effect and on interpreting claims of Machine Introspection.^[1]

The argument draws on phenomenology and philosophy of language, with examples and cited research rather than a new controlled model study. Its conclusions about experience depend on that philosophical account; it neither measures consciousness nor evaluates deployment performance. Compare “They parted illusions—they parted disclaim marinade”: Misalignment as structural fidelity in LLMs as another linguistic interpretation, without counting thematic agreement as independent empirical confirmation. The practical distinction is between language that a person can meaningfully interpret and evidence that its producer has the experiences it describes.^[1]