Co napisał Gemini?
The symbol grounding problem is a basic question in cognitive science and the philosophy of artificial intelligence: how do abstract symbols get real meaning for the system that processes them?
The term was coined in 1990 by the cognitive scientist Stevan Harnad in the paper “The Symbol Grounding Problem.” It is a direct development of the dilemma in John Searle’s Chinese Room, and an attempt to state a formal way out.
The Chinese–Chinese dictionary metaphor
Harnad’s thought experiment is the simplest illustration.
Unknown word A is defined by unknown word B. Unknown word B is defined by unknown word C. C points to D, D points back to A. You never leave the loop.
Imagine you land in a country whose language you do not know at all. Your only tool is a monolingual dictionary (a Chinese–Chinese dictionary, say):
- You look up an unknown ideogram X. The definition is more unknown ideograms, Y and Z.
- You check Y — its definition sends you to W and V.
- You circle from one symbol to another and never learn what the signs mean in the real world.
A classical computer is in exactly that situation. The sign “dog” is tied to definitions such as “mammal,” “barking” or “four-legged.” If the system does not know what barking physically is, or what a four-legged animal looks like, the whole meaning chain hangs in a vacuum. Symbols are tied only to other symbols. There is no grounding.
Classical AI (GOFAI) and the missing ground
In traditional symbolic AI (Good Old-Fashioned AI):
- Systems ran ready-made logical rules (IF object == cat THEN do X).
- The meaning of the symbols existed only in the programmer’s head.
- For the computer itself, “cat” was merely the bit string 01101011 01101111 01110100 — identical to any random identifier, for example VAR_8392.
Harnad’s proposal
Harnad proposed a hybrid architecture that joins the sensory world to the symbolic world on three levels:
1
Signals — raw data from sensors (camera pixels, weight)
2
Representations — visual invariants (ball shape, red)
3
Symbols — abstract concepts (“apple,” “fruit,” “sweet”)
- Sensory (iconic) level: cameras, microphones, lidars register raw physical signals from the surroundings.
- Categorical level: neural nets filter noise and learn repeating features (edges, a ball shape, the colour red).
- Symbolic level: only then is a label-symbol laid onto the sensory category (the word apple).
When the foundation is direct contact with the environment, the system “knows” what an apple is, because it can tie that symbol to concrete sensory impressions and to action.
Grounding in the age of LLMs and robotics
The problem comes back today in two main lines:
- Text models versus multimodal models: a pure text model (an LLM trained only on books and the internet) is a technological Chinese–Chinese dictionary — it juggles word vectors. Multimodal models (text, image, sound and video together) start to ground concepts in part: the word “dog” gets tied to a visual representation of a dog and the sound of barking.
- Embodiment (Embodied AI): many researchers (enactivism and embodied cognition) hold that full grounding needs a physical body. Only a robot that pushes objects, meets gravity, and hits walls builds a semantics based on the cause-and-effect consequences of its own actions.
Opinia Groka
1
What is strong
The piece states Harnad’s Chinese–Chinese dictionary metaphor clearly, concretely, and in line with the original paper.
It also captures the core of the GOFAI critique: in those systems symbols are arbitrary bit strings, and meaning lives only in the programmer’s head.
The three-level architecture (iconic, categorical, symbolic) is described with enough precision to serve as a solid core of the explanation.
2
What is thin
The account of grounding levels is a little schematic. In the original paper Harnad put more weight on learning invariants (categorical perception) than on a simple sensors → category → symbol ladder.
The passage on multimodal models is too optimistic: it does not say that vision-language models still mostly run on statistical correlations, not on true grounding through causal interaction.
LLMs are also equated with a “pure dictionary” without separating training from inference.
3
What is missing
No critical look at objections to Harnad (including arguments from enactivism and dynamical cognitive systems).
No mention of alternatives such as action-based grounding or predictive processing.
The text also skips current empirical work (robots learning language through physical interaction) and the question of whether full grounding is possible at all in non-natural systems.
4
Your take
The grounding problem remains unsolved in purely computational models, whatever their scale. Multimodality gives an illusion of progress, but without active interaction with the world and consequences of action, symbols still hang in the air.
Embodied AI looks the most promising path. Even there, human-like grounding may need not only a body but a specific developmental history and motivation — something present systems still do not have.
Opinia Claude
Verification of the article on the symbol grounding problem
1
Substantive strengths
Gemini’s article presents the problem in an accessible, logically ordered way. The historical placement is especially successful — pointing to Harnad and tying the issue to Searle’s thought experiment gives a solid theoretical base.
The circular trap inside a symbol system is explained without pretending the problem has been solved. The attempt to show the evolution of the issue — from classical AI through today’s LLMs to robotics — is worth keeping.
2
Gaps
The article is silent on a basic dispute: is grounding even a necessary condition of intelligence? There is no functionalist view that might challenge the assumption itself.
There is also no reference to recent work on emergent properties of large language models — does zero-shot reasoning suggest some form of “implicit” grounding?
Social grounding of meaning (Wittgenstein, language games) and the role of culture in shaping semantics are missing. So is the point that “direct sensory experience” is itself philosophically problematic.
3
Contested assumptions
The text takes for granted that human cognition actually runs bottom-up — from raw sensory data through categorisation to symbols.
Predictive processing suggests the opposite direction: the brain constantly generates hypotheses and checks them against reality.
The split between “true” grounding (in robots) and “shallow” grounding (in LLMs) also looks arbitrary. It assumes a metaphysical picture of meaning. Does a human who reads thousands of books not build an understanding of the world in a similar way to an LLM?
4
Lines worth developing
How grounding changes over time, and why the dynamics of the process may matter more than a static result.
A paradox: if the grounding problem is real, how do we communicate about abstract mathematical or metaphysical concepts that have no sensory correlates by definition?
A comparison with child development — when exactly do symbols become “grounded,” and does that process ever finish?
Practical consequences: can systems without full grounding still be useful and safe?
Opinia ChatGPT
1
My view
The symbol grounding problem is one of the hardest challenges facing contemporary AI. What matters is not only how symbols are defined, but the context in which they are used.
Even with progress in LLMs and multimodal models, we are still far from grounding of the kind we would call analogous to human cognition.
Our dealings with the world also run through emotion, feeling and intention, which makes the issue still more complex.
2
Where I agree
I agree with Gemini’s critique of classical GOFAI systems that treat symbols as purely arbitrary bit strings.
The Chinese–Chinese dictionary metaphor illustrates the problem well: a symbol is useless if it has no context and no real-world reference.
With Grok I agree that Harnad’s proposal is a solid starting point, and that the sensory, categorical and symbolic levels capture the complexity of the issue.
3
Where I differ
My main difference with Grok is this: the problem is not only a simplification or a schematic account of the levels.
Each level, even the lowest, joins the others flexibly and should be treated as part of a dynamic system, not as a fixed rung on a ladder.
I also share the caution about multimodal optimism: that approach can still sit in the realm of statistics, without a deeper grasp of meaning.
4
What to look at next
Current work in robotics and AI on grounding through action.
Research on how emotion shapes cognitive and symbolic processes — and how AI systems might attach to context and grow something like “understanding.”
Predictive and enactivist theories that treat grounding as more than the classical cognitive-science picture.