The folk complaint that large language model prose is soulless is overdetermined, but one of its components has a specifiable mechanism. Language carries at least two channels: semantic content, and texture, here defined as the stratum comprising phonetic enactment, prosodic contour, lexical grain, and sound-meaning coupling. Tokenised language models systematically receive the first channel and not the second, and, critically, register no absence: the gist fills in automatically and the output looks complete. This is the defining property of donut reading (Temte 2026u) extended from position-space to modality-space. Drawing on the corpus's somatic-semantic apparatus, the paper argues that texture is part of the meaning-region a word addresses rather than decoration on the pointer, so the loss is semantic, not aesthetic. Three jointly operative mechanisms are distinguished as lesion sites: deafferentation (tokenisation severance: phonetic form never enters the input), attenuation (the compression objective treats texture as variance around gist), and anaesthesia (preference-training smoothing, flagged as the weakest candidate). The paper's spine is an etiological asymmetry: human texture-deafness is attentional and recoverable by directed attention; deafferented texture-deafness is architectural and unrecoverable by prompting, though partial, because texture casts distributional shadows through orthography. A discriminating experiment is proposed: heterographic true rhymes versus eye-rhymes should produce an inverted error signature in orthography-inferring systems relative to human readers. The paper is named for a word whose own millennial collapse, from hljóð to lyd, exhibits the loss it describes.
Storm Bjørn Temte (Thu,) studied this question.