Consciousness, normativity, and intrinsicmeaning in large language models

Even if large language models (LLMs) were conscious, that would not give them intrinsic semantic intentionality—the ability to have content that is normatively assessable as true or false. The author introduces a hieroglyph-copying thought experiment to show that consciousness plus goal-directed behavior does not guarantee such semantic content. LLMs, which operate through probabilistic optimization, lack epistemic commitment or truth-directed representation. Their outputs appear meaningful only because humans project meaning onto them, not because the models internally ground semantic content.

Study at a glance

Abstract

Recent advances in large language models (LLMs), together with renewed speculation about artificial consciousness, have intensified debates concerning meaning and semantic intentionality in artificial systems. This paper advances a restricted and conditional thesis: even if LLMs were to possess some form of consciousness, this would not suffice to establish intrinsic semantic intentionality. To clarify this claim, I introduce a hieroglyph-copying thought experiment showing that consciousness, even when accompanied by goal-directed or procedural intentionality, does not guarantee normatively assessable semantic content. Drawing on this result, I argue that contemporary LLM architectures—despite their ability to generate contextually appropriate linguistic outputs—operate through probabilistic optimization processes that do not by themselves ground epistemic commitment or truth-directed representation. The argument does not deny the sophistication of LLM representations nor attempt to resolve competing theories of meaning. Rather, it isolates and defends an insufficiency thesis: consciousness alone does not ground intrinsic semantics. The apparent meaningfulness of LLM outputs is therefore best understood as dependent on human interpretive projection rather than as an internally grounded semantic achievement.