Understanding

Understanding is grasping why something is so, or what words mean. Epistemologists have debated whether understanding why is a species of knowledge (Grimm) or a distinct achievement that can survive kinds of luck that knowledge cannot (Kvanvig); either way it shows in the ability to connect, apply and extend what one grasps rather than merely to state it. Linguistic understanding, what it is to understand an expression, is a further topic.

“Does it really understand?” is the question most often asked about language models, and it is usually asked badly. It runs together semantic competence, explanatory grasp, practical know-how and the experience of insight, and these can come apart.

On Williamson’s externalist account, to understand an expression is not to stand in some first-personal relation to its meaning but to participate, with sufficient fluency, in the practice of a linguistic community that uses it. On that account it is a serious question whether language models, trained on and causally coupled to that practice at scale, understand what they produce. I do not answer it in the negative. My argument about what these systems can transmit does not need the premise that they fail to understand, and I do not assert it.

It does not need it because understanding is compatible with every epistemic failing that matters for testimony. A human bullshitter understands his sentences perfectly. What a hearer needs from a source is not that it grasps its words but that the truth of what it asserts figures in why it asserts it (see the Tracking Minimum). These are different achievements, and only the second is at issue.

There is also the user’s understanding. Sullivan has argued that what limits understanding from machine learning models is less their opacity than “link uncertainty”, the lack of evidence connecting a model to the phenomenon it is used to study. That moves attention to explanation and validation, which is where I think it belongs.