Can foundation models label data like humans?