Harnad (1990) put it this way: how can the semantic interpretation of a formal symbol system be made intrinsic to the system, rather than "parasitic" on the meanings in our heads? (1990, p. 335). His illustration is a merry-go-round: learning Chinese as a first language with a Chinese–Chinese dictionary as the only source. Every definition points to other symbols just as opaque, and the ride never stops at what anything means (1990, p. 339). It is a relative of Searle (1980)'s The Chinese Room and of Bender & Koller (2020)'s The Deep-Sea Octopus, but posed as a design problem rather than only as an argument.
Harnad's sketched solution is hybrid: ground elementary symbols bottom-up in nonsymbolic representations—iconic ones, analogs of sensory projections, and categorical ones, detectors of invariant features that connectionism might learn—and build complex symbols on top of them (1990, p. 335). In a footnote he adds the flip side: the "hermeneutic hall of mirrors," in which programmers project onto the system the meanings in their own heads (1990, p. 339) (see ELIZA effect).
The problem was posed for GOFAI, but it returns with language models: their vectors are not discrete symbols, yet they can be just as untethered from the world (Millière & Buckner, 2024a, sec. 3.3). There are three responses. For skeptics, the problem stands: Floridi, Jia, et al. (2025) argue that LLMs do not solve it but circumvent it, operating on content that humans have already grounded, in a kind of epistemic parasitism (2025, sec. 8). Externalists reply that reference can be inherited through causal chains of use, as in Putnam and Kripke, and that training data place the model in such a chain; Mollo and Millière add that RLHF might supply an extralinguistic standard of correctness (Millière & Buckner, 2024a, sec. 3.3). Finally, systems trained end to end on images and actions, like Gato, may deserve different treatment (Millière & Buckner, 2024b, sec. 3.1.2) (see multimodality).
The underlying question is what grounding requires: perceptual contact, causal history, or norms of correctness? Each answer draws a different line between form and meaning.