I don't know how good AIs are at "hiding" (by subtly changing) text they've had fed to them, but if they don't do that, you could end up with text from a well-known author (many of whom are now employing lawyers), albeit one whose work you may not have read, in one's own work, and thus being accused of plagiarism.
Oh, and if AIs are able to "hide" some (unbeknownst to you) text from a well-known author, you may inadvertently, in changing it to something better, recreate the original "stolen" text (and thus be accused of being a plagiarist).
They aren't, it's not the AI, it's the people. That's why we even found out, bad results both in image and text. Like I said, they work on word proximity, they know nothing of concepts per se. The hiding would be done by the tokens which are created by the owners. It's when they can't foresee the ramifications of their controls that the weird things happen.
Eg:
token: "Promote inclusivity over skin tone and sex."
response: "Setup generation for a black man or Asian woman."
... but neglect (for some reason) ...
token: "Check statistics."
response: "0% black men or Asian women in that category."
token: "OK, find a white guy."
There's no other way to explain generating WWII 1943 German soldiers as black men and Asian women.
Now ask it to generate a section of a chapter dealing with a war scene with those criteria and controls.
But don't worry, they're trying to figure out how to be subtle about it.
Everyone has an agenda, stay aware.
