The Classical Grounding Problem
The grounding problem became central in AI and cognitive science once it was noticed that purely formal symbol manipulation does not by itself explain how symbols become about anything at all. Harnad’s classical statement is still the cleanest formulation of that challenge: if a system only moves from symbol to symbol, then meaning risks becoming circular unless some contact with perception, action, categorization, or worldly constraint interrupts the loop [HARNAD-1990].
That formulation remains important because it names a real difficulty. A system can be internally consistent, rule-governed, and formally elegant while still leaving unanswered how its internal tokens acquire more than derivative meaning. But the classical framing also tends to overcompress the issue. It can suggest that cognition depends on one final anchoring relation between symbols and reality, as if the central task were to find the definitive bridge once and for all.
History, Debate, and the Need for a Middle Position
Part of the later debate turned toward embodiment and action. Brooks, in particular, pushed against the idea that intelligence must begin from rich internal representation, and argued that contact with the world is often achieved through situated coupling rather than through semantically complete inner models [BROOKS-1991]. That move helped expose a weakness in representational foundationalism, but it did not eliminate the need for structured reasoning, explicit interpretation, or revisable model construction.
More recent neuro-symbolic work reopens the same issue in a different vocabulary. Hybrid systems try to combine learning and reasoning, but the grounding question returns as soon as learned representations, symbolic objects, and task-specific abstractions are treated as if they had already achieved stable world adequacy [GARCEZ-LAMB-2020]. This is why the grounding problem remains worth keeping, but only if it is reformulated more carefully than in its strongest classical form.
The opinion defended here is that both extremes are misleading. Final grounding is too strong because it assumes a finished ontology and a final relation of reference. Free-floating pragmatism is too weak because it risks treating any locally useful interpretation as enough. What is needed is a middle position in which contact with reality is pursued seriously, but always under fallible and revisable conditions.
Grounding Reconsidered in Meta-Rational Pragmatics
Within meta-rational pragmatics, the grounding problem should neither be dismissed as meaningless nor accepted in its strongest classical form. The classical version of the problem is often framed as if cognition required a final and unambiguous anchoring of symbols in reality, as if the central task were to explain how internal representations become intrinsically connected to the world once and for all. From the perspective of meta-rational pragmatics, this formulation is misleading because it already presupposes exactly what should remain under scrutiny: that symbols are stable primitives, that the world is already partitioned into the correct units of reference, and that a single, fundamental relation of anchoring should exist between them [HARNAD-1990].
Meta-rational pragmatics rejects this strong formulation, not because it denies reality, structure, or rational inquiry, but because it treats all available representations, ontologies, and validation procedures as finite, revisable, and dependent on inductive bias, context, and practical constraints. Meaning is not viewed as something that becomes grounded through a final ontological bridge. Rather, it is progressively stabilized through interpretation, interaction, error correction, coordination, and theory-guided refinement. In this sense, what is usually called grounding is better understood as the achievement of increasingly robust forms of constrained coupling between representational systems and the regularities of the world [BROOKS-1991] [GARCEZ-LAMB-2020].
This position should not be confused with relativism or with an unconstrained pragmatism according to which any locally successful interpretation is sufficient. Meta-rational pragmatics does not reduce truth to convenience, nor does it treat all frames of interpretation as equally acceptable. On the contrary, it seeks rational foundations, but foundations that remain explicitly revisable under pressure from anomaly, contradiction, competing explanations, and empirical resistance. It accepts that we must build systems that reason, learn, and refer under non-ideal conditions, but it refuses to identify any current model, ontology, or semantic scheme with the final structure of reality.
From this perspective, the real issue behind grounding is not whether a system possesses an ultimate and non-interpretive link to the world. The relevant question is whether it can construct and revise representations that remain sufficiently stable, systematically corrigible, empirically constrained, and interoperable across contexts of action, explanation, and further learning. Grounding, in this weaker but more defensible sense, becomes a matter of disciplined stabilization rather than metaphysical anchoring.
This reformulation is especially important in AI. Current systems do not operate without inductive bias, context selection, representational assumptions, and validation regimes. Learning is never pure access to truth, but the formation of usable structure under architectural and data constraints. Reasoning is not a mysterious essence, but a family of regulated transformations over provisional representations. The question, then, is not whether AI systems can escape interpretation and directly grasp reality. No cognitive system, human or artificial, appears to do that. The real question is whether they can support increasingly coherent, revisable, and reality-sensitive interpretive structures without mistaking their own local maps for the territory itself.
Meta-rational pragmatics therefore occupies a middle position. It rejects the fantasy of final grounding, but it also rejects the view that cognition is merely free-floating interpretation. It assumes that rational systems must pursue contact with reality through fallible but disciplined processes of model construction, negotiation of assumptions, empirical correction, and theoretical revision. In this framework, grounding is not abolished. It is transformed from a search for absolute anchoring into an ongoing effort to build representations and inferential structures that are accountable to the world, while remaining aware of their own incompleteness.
The central claim can be stated simply: meta-rational pragmatics is not anti-foundational; it is anti-premature-finality. Applied to grounding, this means that the task is not to discover a perfect and final relation between symbols and reality, but to design cognitive and computational systems capable of progressively improving their contact with reality without erasing the interpretive, biased, and revisable conditions under which such contact is always achieved.
Relation to the Other Articles
This article overlaps most directly with Meta-Rational Pragmatics in Context, where grounding appears inside the broader discussion of reification and neuro-symbolic AI, and with A New Foundational Intuition for Neuro-Symbolic AI, where premature closure and plurality of partially adequate theories are treated as architectural issues. It also connects back to Meta-Rational Pragmatics, because the present argument is ultimately another way of defining why interpretation must remain governed rather than silently absolutized.
References
- [BROOKS-1991] Brooks, R. A. (1991). Intelligence without Representation.
- [GARCEZ-LAMB-2020] d'Avila Garcez, A., & Lamb, L. C. (2020). Neurosymbolic AI: The 3rd Wave.
- [HARNAD-1990] Harnad, S. (1990). The Symbol Grounding Problem.