What Makes Semantics Semantic?

A classical definition of semantics comes from philosophy of language and logic. In this tradition, semantics concerns the relationship between symbols and what they stand for. Whereas syntax specifies the formal rules governing how symbols are combined, semantics specifies their meaning by assigning them reference and truth conditions. In formal semantics, the meaning of a sentence is defined by the conditions under which it would be true in a model of the world. In model-theoretic semantics, an interpretation function maps symbols to objects, properties, or relations within a domain, and sentences are evaluated relative to that mapping. More broadly, classical semantics treats meaning as a relation between representations and the world: words refer to objects, predicates denote properties, and statements express propositions that can be true or false. Under this view, semantics is fundamentally about correspondence—how internal or linguistic structures map onto states of affairs in an external domain.

In artificial intelligence and neuroscience, however, the term semantics is rarely framed in terms of truth conditions or formal reference. Instead, it is used operationally. When researchers speak of “semantic representations,” they refer to internal variables or neural activity patterns that capture task-relevant structure in the input space. In machine learning, semantics often corresponds to latent features that support generalization—representations that remain stable across variations in noise or context while preserving distinctions that matter for prediction or decision-making. In neuroscience, semantic coding refers to neural activity that correlates systematically with behaviorally meaningful variables such as spatial location, object identity, or decision variables, rather than merely reflecting raw sensory input. In both fields, semantics is defined not by formal truth, but by functional role: a representation is semantic if it organizes information in a way that supports reliable inference, prediction, or action under task constraints.

Despite their differences, classical semantics and modern scientific usage share a common structural principle. In both cases, semantics involves lawful correspondence between internal organization and external structure. In philosophy, this correspondence is expressed as reference and truth conditions. In AI and neuroscience, it appears as reliable tracking of behaviorally or environmentally relevant variation. Whether described as denotation, latent structure, neural tuning, or invariant features, the core idea is the same: semantics is not mere internal complexity, but internal structure that stands in a stable relation to something beyond itself. The remaining question, then, is not whether semantics involves correspondence, but what kind of correspondence is sufficient for a physical system to possess it.

Consider a simple organism that must distinguish between two types of berries—one nutritious and one poisonous. From an external perspective, the berries are objectively different and have different consequences for survival. Yet if the organism’s sensory system produces exactly the same internal state for both berries, it cannot behave differently toward them. For that organism, the distinction does not exist. Although the world contains a difference, that difference is not represented in a way that can guide behavior. In this sense, there is no semantics about the two berries for that system.

Now consider the opposite case. Suppose the berries are actually identical, but internal noise causes the organism to produce different internal representations when encountering them. Although its internal states vary, these variations do not correspond to any real external difference. Behavioral differences, if any, would be random and not systematically related to the world. Again, there is no semantics—because semantics requires that internal differences track real external differences in a stable and law-governed way.

These cases illustrate a first principle: semantics requires structured correspondence. If external differences do not induce distinguishable internal states, meaning cannot arise. If internal differences do not reflect external structure, meaning is illusory. Semantics emerges only when internal variation reliably tracks external variation in a way that can guide behavior.

But correspondence alone is not sufficient. We must ask when such correspondence actually matters. Internal states may track external variation, yet those distinctions might have no consequence for the system’s achievable outcomes. Semantics requires more than alignment; it requires leverage. A distinction becomes meaningful only if exploiting it expands what the system can reliably achieve. If distinguishing two environmental states does not alter achievable performance under available policies, the distinction carries no usable meaning. Meaning exists where structured differences can be turned into reliable advantage under the system’s constraints.

Under this definition, two conditions must be satisfied. First, external variation must evoke distinguishable internal states. Second, acting differently on the basis of those distinctions must lead to systematically different outcomes. Representation alone is insufficient; the distinction must matter. What makes it matter is an external selection mechanism—environmental constraints, task demands, survival pressure, or an outcome functional—that assigns differential consequences to behavior. Without such selection, internal distinctions remain inert.

There is, however, a complementary situation. Suppose distinct external states initially evoke the same internal representation, leading the system to take the same action. If that same action produces systematically different outcomes depending on the true external state, the mismatch between expectation and consequence reveals that the representation is insufficient. Through interaction, the system can reorganize its internal structure to differentiate what it previously collapsed. Here semantics arises not from passive encoding but from outcome-level inconsistency under identical action. The world forces refinement by enforcing consequence. Meaning emerges within the closed loop of representation, action, and externally imposed outcome differences.

This leads to a structural constraint:

Impossibility I: No Semantics without External Constraints

No learning system can acquire semantics solely through internal statistical adaptation in the absence of a selection mechanism.

Internal statistical adaptation can extract correlations, compress high-dimensional data, and model complex structure. But without an external mechanism that assigns differential consequence, there is no principled way to distinguish meaningful structure from arbitrary structure. Statistics can generate organization; only selection can stabilize meaning.

This perspective clarifies the role of contemporary large language models. During pretraining, LLMs optimize next-token prediction over vast corpora, extracting enormous statistical regularities. This process builds highly structured representations. Yet next-token prediction alone does not impose task-level consequence. It does not determine which distinctions are behaviorally meaningful; it merely models distributional structure. Usable semantics emerges only when external selection mechanisms—supervised fine-tuning, reinforcement learning from human feedback, instruction tuning, or downstream objectives—assign differential consequences to outputs. These procedures reward some behaviors and penalize others, stabilizing particular internal distinctions because exploiting them improves expected outcomes. Statistical pretraining builds capacity; externally enforced selection converts capacity into meaning.

This reasoning raises a deeper question: where do stable selection mechanisms originate? In artificial systems, they are typically imposed externally—through labels, reward signals, or human feedback. In biological organisms, however, selection is grounded in physical coupling with the environment. Actions consume energy, incur risk, and affect survival. Errors have material consequences. Embodiment provides a continuous selection mechanism: the world enforces differential outcomes on behavior.

If semantics requires externally enforced consequence, then intelligence requires a channel through which consequence feeds back into the system. Embodiment, in this broader sense, is not about having a physical body but about being embedded in a causal loop where different behaviors lead to different enforced outcomes. Whether through physical interaction, reinforcement signals, or human feedback, such a channel stabilizes distinctions and converts structure into usable meaning. Without it, a system may accumulate statistical regularities but cannot determine what truly matters. Intelligence, therefore, is not merely the capacity to compute—it is the capacity to stand in relation to consequence.