If you post this image on Twitter, you're immediately suspended for 12 hours

You'd presume abuse image prevention would use hashes that survive slight fuzzing.

https://inhope.org/EN/articles/what-is-image-hashing?locale=...

Spoken like a person who does not consider the sheer volume Twitter (and companies like it) have to deal with.

The company would go bankrupt if they hired enough people to handle all of the manual posts. They also can't just leave abusive images up... so this is the compromise.

I’ve always wondered that for hash table we always check that the key matches the target because of hash collision. Should the same be done here? For image hash matches, either run though an image similarity check or, if original image is not available (it’s so bad it can’t be stored), a secondary hash used to double check?