The model has exactly one source of knowledge: the patterns in its training text. If a topic is thin or missing there, the model has nothing else to draw on — but it still answers.
Thin patterns get filled in
Frequently covered topics leave strong, consistent patterns, so answers tend to be close to the truth. Rarely covered topics leave weak patterns. The model still needs a next word, so it borrows the shape of a nearby answer — a familiar-sounding name, a typical date, a plausible title — and the result reads like a real answer even though the details were never in the text.
Frozen in time
The training text was collected at one point and then stopped changing. Anything that happened or was corrected afterward is absent. The model has no clock, so it cannot tell that its information is out of date. It reports the old version with the same confidence as the current one.
It helps to drop the idea of a failed lookup. The model is not searching for a fact and coming back empty. It is doing its normal job — continuing the text — in a place where the patterns are weak. From the outside, a thin-pattern answer and a well-supported one look identical.