You could solve that issue using differential privacy (DP). I've applied DP successfully to some deep learning problems, but I'm not sure if it would work for C

You could solve that issue using differential privacy (DP). I've applied DP successfully to some deep learning problems, but I'm not sure if it would work for ChatGPT as it effectively limits the number of training steps / epochs you can do with any given sample. It would be interesting to check that though, ChatGPT has an enormous amount of training data so the sample rate for each training step is probably miniscule, which makes applying DP much easier. In the end what matters is mostly how often a single text source is used in the training. In each training step, DP would add random noise to the gradients to mask the effect of any specific input value. That works really well but if we repeat it too much there's the risk an adversary could average out the noise (though from my experience that is extremely difficult even without DP). I guess that would also answer the question whether ChatGPT or other models are parroting existing text to some degree, as differential privacy would ensure that the effect of a single text source on the resulting model is indistinguishable from random noise.