Agreed. A few year back the devs looking for quick fixes would go over to StackOverflow and just copy answers without reading explanations. This caused the same type of problems that OP is talking about. That said, the ease of AI might be making things even worse.
As far as I know, the apps are not intercepting the text messages for passcodes. The messages have a specific format and a hash to indicate which app they are targeting. It is up to the messages app to read the message and to forward the code. This design should not need to give the apps any access to your messages.
It’s much more complicated than this. Given that models have been shown to spit out verbatim copies of some training material, it can be argued that the weights do in fact encode the material, just in some obfuscated way. Additionally, it can be argued that the output of the model is a derivative copy of the original work regardless of whether the original work can be “found inside” the model weights, just by the nature of the process. As of now, there is no precedent that I know of on whether this constitutes redistribution of copyrighted material.