Symbol Grounding
The problem of how symbols get meaning if they are defined only by other symbols — the sharpest form of the objection that a text-trained system cannot understand anything.
Reviewed July 16, 2026Contested
When not to use it
- As a practical criterion. It does not predict whether a system will perform a task reliably.
- As a settled refutation. The premise that meaning requires external reference is exactly what the other side denies.
Reach for something else instead
- Conceptual role semantics, which locates meaning in inferential relations rather than reference.
- Behavioural evaluation, which sidesteps the question by measuring task performance.
- Multimodal grounding, which supplies a causal channel to the world and may relocate rather than solve the regress.
Sources & further reading
- Harnad (1990), The Symbol Grounding Problem — the original formulation.
- Searle (1980), Minds, Brains, and Programs — the Chinese Room, which this sharpens.
- Bender & Koller (2020), Climbing towards NLU — the octopus, updating the argument for text-trained systems.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Treating it as an empirical claim. The core disagreement is about what meaning is.
- Assuming the argument is against capability. It is about what kind of thing could have been learned.
- Reading interpretability results as settling it, when both accounts predict the same findings.
At a glance
FieldFoundations
OriginHarnad 1990
SharpensSearle's Chinese Room
Updated byBender & Koller's octopus
Statusunresolved, partly definitional
DifficultyFrontier
Where this sits
A destination. 8 concepts lead here, and nothing in the corpus depends on it.
6Levelsteps in
8Needs firstconcepts
0Opens upnothing further
1Areastays here
Learn these firstDistributional Hypothesis
Computed from the prerequisite graph, not assigned. How this works