This is why he's incorporated the avatar - in the example he provides you wouldn't see the "boot" avatar if you had missed a letter, the boot is part of the address and should be communicated verbally when giving the address verbally.
So to check whether the address is valid I need to use a smartphone app. If I cannot do it online and only after someone gives me the address I will know it is broken, but will not know what is the correct address.
Bingo, error detection is functionally distinct from error recovery. In this context, error detection helps prevent some misdeliveries, but can't help ensure additional correct deliveries in the face of error.
By analogy, ECC memory can typically detect two bit errors but can only correct single bit errors. That's because the functionality used to detect two bit errors in ECC memory is insufficient to recover both bits.
The concept of an avatar doesn't translate across cultures. And the visual hash isn't an "avatar" as typically used on the internet. Furthermore, the images may be incomprehensible in some cultures.
Keeping a minimal edit distance between code words in the corpus would enable error detection and correction without the burden of sending another piece of info.
