The case that they do not: form is not meaning
The sharpest skeptical argument comes from the linguists Emily Bender and Alexander Koller. In a 2020 paper they imagine a hyper-intelligent octopus that taps an undersea cable and listens to two people texting. It learns to predict what comes next so well that it can impersonate one of them, yet it has never seen the ocean, a coconut, or a bear. When one texter faces a real emergency and asks how to build a weapon from sticks, the octopus, fluent in the conversation's patterns but ignorant of what any word points to, has nothing useful to offer. Their claim: a system trained only on the form of language has no path to its meaning.
Bender and colleagues sharpened this in 2021 with the phrase 'stochastic parrot': a system that stitches together word sequences by probability without reference to meaning, while readers project understanding onto the fluent output. On this view a language model manipulates symbols expertly but grasps nothing behind them.
The case that they do, at least in part: world models
The counterevidence is that predicting text well seems to require building structure that goes beyond surface statistics. In a striking 2023 study, Kenneth Li and colleagues trained a model only on move sequences from the board game Othello. It was never shown a board or told the rules, yet probes revealed it had constructed an internal representation of the board's actual state, and editing that internal representation changed its predictions exactly as you would expect if it were really using it. Something functionally like a world model can emerge from pure next-word training.
Larger models also show abilities that smaller ones lack, from multi-step arithmetic to translation, though researchers debate how sharp and genuine these 'emergent' jumps really are. Reviewing the whole dispute for the journal PNAS in 2023, Melanie Mitchell and David Krakauer found the research community genuinely split, and suggested that 'understand' may name several different things that people and machines possess in different mixtures.
So who is right?
The most defensible reading is that both sides see something true. Modern models clearly understand in a functional sense: they track context, infer intent, and generalize to new problems in ways that pure lookup cannot explain. But their meaning is learned from how words relate to other words, not from a body moving through a world, from hunger, or from a shared life, so it stays partial and ungrounded compared with ours.
And competence with language is not the same as consciousness. A system can use words aptly with no inner experience of what they mean. 'Do they understand?' turns out to be less a yes-or-no question than an invitation to state precisely which kind of understanding you have in mind.