Food for tought: What happens when we add a layer of indirection in the training? Imagine we have a model but don't train it with the copyrighted data. But we

Food for tought: What happens when we add a layer of indirection in the training? Imagine we have a model but don't train it with the copyrighted data. But we train if with every review ever made public about it. If we have a large number of reviews the model will be able to give pretty good responses about the topic and may even be able to reproduce some part of it. Is this copyright infringment? And what if we add another layer? Now we just train the model with the comments people leave about the reviews.