The fact they are trained to complete sentences and are thus word predicting models only tells us vaguely how they work. It does not tell us anything about their sentience.
The sentience, if there is, comes from the emergent reasoning that develops in the massive neural network to offer actual good predictions: the more you think, the better the predictions.
For example, models trained to produce textual board games predictions can be sounded to show they developed subnetworks corresponding to an inner representation of the states of the game. Models trained to produce images have emergent subnetworks on depth-maps, contours and lighting.