Usernames aren't random enough to be safe as a simple MD5. Perhaps with a strong bcrypt, but similar to PIN codes, it might be better to give partial information like "is the second character an ...", assuming nobody else made similar statements. Or give the first ~two hex characters of the hash, so that it would match 1/(16²)rd of the usernames. I'm sure there's also a clever way for a zero-knowledge proof here, probably something with diffie-hellman using the name as your random integer or something, but I'm too sick to think about this stuff right now. Privately sharing data publicly is hard.