That's true. But, it's also completely possible to make a machine learning algorithms to separate "real text" (text meant for human to human communication) from text encoded images since they differ in significant ways. Granted, you could always try to make a text encoding format which resembles real text, but I'm fairly sure the machine learning algorithm could be constructed in such a way to make its usage unfeasible.
Yes, which is why you keep humans in the loop: you design a system in which the will of humans who want the system to be used only for text prevails and which is resistant to attack by humans with contradictory agendas.
Yeah. You'd have to be more clever than that. There's a lot of interesting work on meaning extraction and text analysis. I wonder if you could have some kind of information density threshold.
It would be possible to encrypt text in such a way that things could be said about its information density but not about its meaning, too, though that would permit steganography to be used. But it would raise the technical bar for using the system for this purpose so high that it would probably drive away all the chickenboners.
There's also a dumb way: length limits. That would force binary data to be divided up into a huge number of posts, making it an annoying medium for file trading. Plotting the Iranian revolution would not require >1mb posts to a forum.