So apparently, it's a cloud only service -- you have to upload all images to Microsoft to use it.

Also to be able to use the service you have to apply as a business and be approved manually.

This would have been more useful as a software that runs directly on your server and checks at the time of upload (or to run against existing local user-content) instead of requiring re-uploads of the content to their cloud, not to mention the privacy implications.

Another puzzling thing is that they say MS doesn't retain the images and that they're converting the content to hashes immediately and comparing against known abuse hashes. Wouldn't it be more efficient for everyone if the hash were generated locally and only the hash sent up to their cloud?

I agree that sending the hash is the best implementation if it works as you have described, but that is a system that would be trivial to circumvent (make the .png into a .jpg for example). Maybe they convert all the images to a similar file type and then hash it?

The hash they use is more complicated than that:

  converts images into a common black-and-white format 
  and uniform size, then divides the image into squares 
  and assigns a numerical value that represents the 
  unique shading found within each square. Together, 
  those numerical values represent the "PhotoDNA 
  signature" or "hash" of an image

- https://www.microsoft.com/en-us/PhotoDNA/FAQ

What I was suggesting was that the hash generation tool they use be distributed so the hash can be generated locally before sending to MS. This way MS can continue to keep all known hashes to themselves while making it easier for businesses to bulk check images without huge bandwidth charges or privacy issues.

It appears we are in agreement. The only reason I can think of for them doing this is maybe they want to keep the PhotoDNA stuff private? Imagine the overhead on their end as well. The best way might be downloading the hashes to each server and comparing them like a virus scanner.