The symbol grounding problem asks how symbols acquire meaning through connection to theworld. Applied to large language models (LLMs), it has generated five incompatiblepositions: LLMs fail at grounding, circumvent it, can solve it, dissolve it, or face it as an openempirical question. This paper argues that all five share an unexamined presupposition: thatcausation is a feature of the world to which symbol-producing systems must connect.Drawing on the tradition that questions causation’s worldly status (Hume, Russell, Norton)and on prediction-error-minimization accounts of neural processing (Friston, Clark), weargue that causation is a neural phenomenon — a label that brains generate when predictionerror reaches stable lows. This label is real and functionally significant, but it belongs to thebrain that generates it, not to the world, and not to any text the brain produces. Applying thegrounding demand to LLMs is therefore a category error: the application of a brain-scopedcriterion to a non-brain system. The paper connects, for the first time, Harnad’s (1990)symbol grounding problem and Haverkamp’s (2025) concept of semantic weightlessness,showing them to be structurally complementary descriptions of the same brain-scopeddemand viewed from opposite directions. Haverkamp’s phenomenological observation ispreserved — LLM text does feel unanchored to human readers — but the diagnosis isrelocated from system deficit to observer projection. Floridi, Jia, and Tohmé’s (2025) claimthat LLMs “circumvent” the grounding problem is acknowledged as practically correct butshown to carry an unwarranted theoretical presupposition.
Franny Philos Sophia (Sun,) studied this question.