Comment by zamadatix

Another way to look at it is the LLM is, by definition, what's expected to be probable based on the training data and this, by the same definition, is extremely unlikely data to run across. With high uncertainty comes the need to verify until it can level out as "really surprising" instead of "plausible sounding error".