Loading slide
Now we can see hallucination all the way down. Earlier in this module we met it from one side: the model was trained to sound confident, so it sounds confident even when it should not. This chapter shows the other side, the deeper reason it cannot simply check itself.
A database can tell you "no such record." It has an index; it can look and report a miss. A model storing its knowledge as a smeared-out pattern across billions of weights has no index to consult. It cannot glance inside and confirm a fact is really in there before it speaks. It just generates the most probable continuation. And because the training text was mostly true, the most probable continuation is usually true, which is exactly what makes the misses so dangerous. The false answer arrives in the same fluent, confident shape as the true ones.
This is hallucination, and putting both halves together, the trained-in confidence and the missing self-check, shows why it is structural rather than a bug. The model has no concept of true versus false to begin with. It is predicting what would plausibly follow, and sometimes what would plausibly follow is wrong. There is no inner gauge of certainty to consult, because nothing in the training ever built one.
So this is not a defect a later version simply removes. It falls straight out of the three things this chapter laid bare: knowledge as a pattern with no readable index, an edge the model cannot feel, and a clock frozen at training time. The same distributed learning that lets the model know so much is what stops it from knowing what it does not know.