BIP39 is just a representation of a binary string with the assumption that remembering a random string of words is easier than remembering a random string of letters. However, for any levels of significant entropy, they both are similarly “impossible” to remember (and painful to type).
A BIP39 word is 11 bits, a letter from the [A-Za-z0-9] set is a little over 5.95 bits, so requesting a BIP39 word instead of 2 alphanumeric characters is not any better, just more complicated.
Also, I don’t see why I should type 12 random words (for 132 bit entropy) when I can come up with something more secure (larger vocabulary, numerics, punctuation, UTF-8 characters, capitals, etc.) in passphrase of 8-9 words that I can actually remember. Zxcvbn can measure password strength pretty well, so I don’t see why should we both restrict our choices and complicate our implementation for no additional benefit.