We've got lots of datasets at this point in history - it ends up being pretty reasonable to formulate a hypothesis, propose a change to the ML system, then see how it does on a host of datasets. Most papers focus on one or two datasets, getting some initial improvement going from train-to-test, but the successful ideas end up transferring to a wide range of contexts. The extension to other dataseets + contexts constitutes the prediction of performance on other systems.

And sometimes (often?) techniques do fail to transfer! IME, it's only the simplest ideas which do transfer to new contexts well - there's a lot of dark heat in pushing Imagenet another hundredth of a point, which you realize when you try to take incremental 'new SOTA' papers and apply them to audio problems.

But the things which /do/ transfer constitute real advances. And that's (drumroll) science! Sometimes things look good initially and get falsified. Medicine is still science, even if drugs get disqualified in phase iii trials...