Image search does not classify image contents. It uses the site text for ranking of the images on that site.
Do a google (not image) search for "white people" and you'll see that this phrase is mostly used in pages that are in fact about racism and therefore likely to contain images of black people.
It most certainly does, and has been for at least a year or two, or possibly even a bit longer (can't remember when I first noticed this behaviour).
You can e.g. do a search along the lines of site:<domain of online clothing store> <hair style/hair colour/…>, and at least for the most common and recognisable kinds of hair styles, it will actually return relatively reasonable image results, even though online shops most certainly don't have the habit of annotating the hair styles worn by their models on their product pages.
Along the same lines, Google is now also in the habit of OCRing any text content it can find in images and indexing that for search, too.
It's true that it'll still also take the text surrounding the image into account, but it's no longer true that image search is only based on that.
