The problem of how words connect to things in the world, rather than only to other words. A system trained purely on text has definitions all the way down, which is either a fundamental limitation or a philosophical worry with no practical consequence, depending on who you ask.
Technical view
Harnad’s formulation asks how symbols acquire meaning without an infinite regress of definitions. Multimodal training partially addresses it by linking language to perception, and embodied learning links it to action. Evidence that text-only models develop representations correlating with physical properties such as colour and spatial relations suggests substantial grounding is recoverable from language alone, though the point at which correlation becomes grounding is exactly what is disputed.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.