I think pictures, for now at least, while neural net research is still at the current stage.
I feel like letters and numbers should be easy for OCR systems.
I guess picture where you need to select particular object from the group. But IMO they should change type of object also for new request as on some sites every time it asks for same type.
Better in what sense? Lately I have been more annoyed as a user by the picture ones, because they seem ridiculous when it’s not even clear to a human which boxes should satisfy the question.