You could potentially introduce an occasional prompt and check the answer against a few LLMs to see if they match.
If I ask even just GPT 3.5 something like: "Who was John Adams?" - it'll give me a slightly varied answer pretty much every single time (even if I modify its settings to make it less creative with responses, it'll still usually vary a bit).
Here is a simple API hit on gpt-3.5-turbo-16k-0613
Output 1) John Adams was an American statesman, lawyer, diplomat, and Founding Father who served as the second President of the United States from 1797 to 1801. He was one of the key figures in the American Revolution and played a crucial role in drafting the Declaration of Independence. Adams also served as the first Vice President under George Washington. He was known for his strong advocacy of republicanism and his belief in a strong central government. Adams was a prolific writer and his letters and writings provide valuable insights into the early years of the United States.
Output 2) John Adams was an American statesman, lawyer, diplomat, and Founding Father who served as the second President of the United States from 1797 to 1801. He was one of the key figures in the American Revolution and played a crucial role in drafting the Declaration of Independence. Adams was also a strong advocate for the separation of powers and a strong central government. He was known for his intellect and commitment to public service.
And so on.
so the only way one could see the output of the AI was to read the QR code which resulted in the intput & the output?
The open source variations in the wild aren't going to necessarily follow any such rules. So even if eg a government wanted to lock it down that way, the mass of AI content spilling over the walls anyway would void the effort.
I doubt the genie is going back into the bottle or can be easily controlled. We'll come up with some basic ways to discern AI content, one of which will be a government regulation (or standards body) stipulating all AI content must/should be labeled as such (with movie like rating gradiants: entirely AI (E-AI), partially AI (P-AI), no AI (N-AI); or something like that). If we're lucky it'll just be a content industry standard that will be developed, without forced government labeling. QR codes or the equivalent could play a role there (QR code next to the content label; end-point data includes where it came from, when), although someone has to host a permanent repository for the end-point data.
Like, what??