Can Artificial Intelligence Possess Meaning Without the Pressure to Survive? A Relational Account of Meaning, Grounding, Normativity, and the Limits of Large Language Models

Abstract

Meaning is routinely conflated with information, representation, linguistic competence, or behavioral success. This Perspective develops a relational account in which meaning is the directional causal value that a difference acquires relative to a system's organization, goals, and possible futures. The account is used to analyze large language models (LLMs) and future artificial agents. Human language is treated as a selective, socially stabilized compression of a meaning-world already shaped by human perception, embodiment, action, culture, and value. Text-trained LLMs therefore learn an indirect record of reality: a world already filtered and compressed by human beings. This does not reduce LLMs to memorization; linguistic traces can support the reconstruction of latent spatial, temporal, causal, and conceptual structure. However, semantic competence, grounded functional meaning, and first-person meaning remain distinct. We introduce the concept of meaning ownership to identify who or what ultimately bears the consequences to which a meaning points. Biological survival pressure provides a powerful route to intrinsic normativity because living systems must maintain themselves under conditions of vulnerability, but biological life is not necessary for functional meaning. Artificial agents may acquire operational significance through embodiment, self-maintenance, persistent memory, causal intervention, endogenous goal revision, and social answerability. Whether these conditions can generate first-person concern or consciousness remains unresolved. The absence of autonomous meaning need not limit instrumental superintelligence; instead, it may produce a dangerous asymmetry between extreme competence in achieving goals and dependence on externally supplied ends.