Hallucination
Also known as: confabulation, invented detail
A model stating something that is not in the footage, fluently and with no sign of doubt, which is what makes it hard to catch.
Updated
What it means
A hallucination is an answer that sounds exactly like the correct ones and happens to be false. Asked what is in a clip, a model may report a second rider who was never there.
It may equally report a jump in a stretch where nothing left the ground. The phrasing carries no hint that this answer is different in kind from a true one, which is why it survives casual review.
It is not lying and not a bug in the ordinary sense. A Large language model produces plausible continuations, and on unfamiliar footage the most plausible description is sometimes of a scene that was not recorded.
How it works in practice
The defence is cross checking against things that can be measured. A claimed impact should coincide with a motion spike and an audio peak, so Motion detection and Audio event detection become the arbiters.
Claims that fail the check are dropped rather than ranked lower, because a single invented moment in a short highlight is far more damaging than a missed real one.
Keeping the question narrow helps as well. Asking which of these five stretches is strongest invites fewer inventions than asking what happened during this hour of footage.
What to watch out for
Confidence is no guide. A Confidence score attached to an invented detection can be as high as one attached to a real detection, so filtering by certainty alone does not remove the problem at all.
Unusual footage raises the rate. Underwater clips, heavy fisheye and near darkness all sit outside what most models have seen, and those are the clips a person is least able to verify quickly by eye.
Trusting the explanation is a related trap. A model asked why it chose a moment will produce a reason whether or not that reason drove the choice, so the justification is not evidence.
How RawClip handles it
Anything the model asserts about a clip is checked against measured motion and sound before it can influence the cut, so an invented jump never becomes an edit point. Unsupported claims are discarded rather than merely ranked lower.