Philosophers argue deep learning systems understand things, but in a broken way

A philosopher proposes that AI systems can actually achieve understanding by building internal models that track real patterns and make reliable predictions, not just mimicking training data. But there's a catch: this understanding is fragmented and messy — the AI's internal logic doesn't match how the world actually works, it can't explain itself, and it doesn't connect different domains the way human understanding does.

This paper cuts through the meaningless debate about whether AI 'really understands' anything by proposing a testable definition: if a system builds an accurate internal model and uses it to predict reliably, it understands something. The practical implication is clearer: we should stop asking 'does this AI understand' and start asking 'what specifically did it understand, and how broken is that understanding in ways that matter.' For deployment decisions, knowing an AI has fractured understanding of its domain is more useful than pretending it either does or doesn't understand at all.