The output of MD5 is binary: a sequence of 128 bits, commonly encoded as 16 bytes (technically, 16 octets, but let's use the common convention of bytes being octets).
Humans don't read bits or bytes. They read characters. There are numerous code pages which tell how to encode characters as bytes, and, similarly, to decode bytes into characters. For almost all of them (because of ASCII), the low-value bytes (0 to 31) are "control characters", hence not really representable as characters. So nobody really reads MD5 output directly. If someone is "reading" the hash values, then these values are most probably encoded into characters using one of the few common conventions for that. The two most prevalent conventions are hexadecimal and Base64.
With hexadecimal, there are only digits, and letters 'a' to 'f' (traditionally lowercase for hash values). You won't get "SALT ME!" in an hexadecimal output...
With Base64, encoding uses all 26 unaccentuated latin letters (both lowercase and uppercase), digits, and the '+' and '/' signs. You could thus hope for "SaltMe" or "SALTME". Now that is doable, because each character in Base64 encodes 6 bits, so a 6-letter output corresponds to 36 bits only. Looking for a password which yields either "SaltMe" or "SALTME" will be done in (on average) 235 tries, i.e. within a few minutes or hours with some decently optimized code.
Note, though, that someone who actually spends some time to read Base64-encoded hash values probably has some, let's say, "social issues", and as such might not react the way you hope.
And it is done: When hashing with MD5 then Base64-encoding the result:
infjfieq
yields: SALTMEnBrODYbFY0c/tf+Q==
lakvqagi
yields: SaltMe+neeRdUB6h99kOFQ==