It's great for spooks too. Now they just have to exfiltrate the keywords describing the images instead of the images themselves.
The difference between whether someone has 4 or 5 fingers, whether they're holding a gun versus a random object, or whether they're mixed-race or caucasian, all seem like they would be pretty important things. Likewise, car number plates, signage in the photo that help identify where it is, metadata of the image itself (often more useful than the image), are all incredibly important. All of those are things that AI is absolutely terrible at lmao.
I'm not sure how a bunch of probably-correct keywords that miss a lot of the important aforementioned details in the image is more useful than the image itself, or it's metadata. Both of which would be lost. My point applies equally with respect to image classification, too.