Does it understand?

The dispute worth stating honestly.

Both extremes overreach

Only predicting the next word understates what the training captures. Genuine understanding overstates what a text-prediction system demonstrably has. Researchers disagree, and the disagreement is real.

It has no body, no persistent memory and no stake in being right.
It also handles novel combinations in ways pure memorisation could not.

Why it matters

The framing changes what you expect. Treating it as a fluent, well read, unreliable assistant produces better decisions than either extreme.