Geoffrey Hinton, the computer scientist often called the godfather of AI, argues that large language models genuinely understand language rather than piecing it together by imitation.
Large language models are the systems behind chatbots such as ChatGPT, trained on vast quantities of text to predict what word comes next.
Hinton, who shared the 2018 Turing Award and won the 2024 Nobel Prize in Physics for his work on neural networks, has made the case in a series of public lectures.
He rejects the common charge that the models are glorified autocomplete, stitching stored text together on a statistical basis.
That idea, he says, is nonsense.
Meaning without symbols
Hinton's argument runs against a long-held assumption in linguistics and philosophy: that real comprehension requires translating language into an internal system of symbols and logical rules.
He contends that neural networks build meaning through structure instead.
In his account, words behave as building blocks that combine in layered patterns, capturing context, relationships and nuance as they interact across the network.
Hinton has likened the process to high-dimensional Lego, with each word a shape that locks together with others to form meaning.
Meaning, on this view, is not a set of fixed definitions filed away in the model.
It emerges from the interaction of those building blocks, learned from data rather than programmed as rules.
A different kind of understanding
The claim reframes what the models are doing, casting their capability as a distinct form of understanding derived from structure rather than symbols.
It also puts Hinton at odds with critics who hold that a system trained only on text cannot grasp meaning, because it lacks any grounding in the physical world.
That objection, associated with the linguist Emily Bender among others, treats fluency as no proof of comprehension.
Hinton's position is the reverse: that human beings understand language in much the same way the models do.
He has drawn the comparison further, arguing that the errors models make when they fabricate information resemble the way human memory reconstructs events rather than replaying them.
The stakes are not academic for Hinton.
He left Google in 2023 to speak more freely about the risks of the technology, and has warned that systems able to match or exceed human intelligence could prove difficult to control.
If the models understand in the way he describes, that prospect moves closer.
