Just looking at the happy human/sorry robot image, it doesn't work as well if the image is large (at least for me), the outline of the decoy message starts to take visual priority over the real message especially if you don't know what it is. With no additional context, I definitely read sorry robot on first glance with the image at phone's width and didn't realise that's not the actual message until I read the description. Seems to work when the letters are smaller though.

I feel like this can be easily cracked if you downscale it down to only a few pixels per letter. Then the decoy letters get filtered out and the human letters lose their blur

Story checks out in Gemini

Gemini says:

The text in the image reads:

TAJS GBNFBNCB JS

WKJIYBN JN BBCDY FONT

…but man does it hurt my eyes to read

Wait, do people think robots use OCR to spy on them? They have access to the actual unicode symbols. They know that an A is an A regardless if it's displayed as Arial, Times New Roman or Wing Dings. Is there a subculture somewhere that people only communicate through text as images?

This sounds awful for accessibility tools such as screen readers.

I know this has good intentions but this is pretty much an accessibility nightmare