Does anyone have a link to the actual study? The article doesn’t seem to have it.

Maybe aerodynamics for boomerang?

I guess this is the same reason Ai clocks are stuck on ten to two. In case of the clocks it is because ten to twolooks nice on a clock so most images will have it. Ai doesn’t understand how time corresponds to how a clock looks, just that something that looks like a clock has two lines there.

So since most depictions of wine glasses are half full, when you ask for a wine glass it will give you one that is half full. Fill level being somewhat irrelevant as it wasn’t part of the training data