You are basically circling the important point in your comment: the AI looks at a sentence in the book, tries to predict what’s next, then it gets handed the right solution, and it then backpropagates and “learns” from it.
If you look at the data, it never stores the original work, but it stores a few numbers here and there (really small amount of data compared to the book) to improve at this sentence next time.
In it’s technical form, these are weights that the AI uses on previous text to derive at new text. Now you would need to argue that these weights completely cover every work in so much as they could replicate the original work in its expression, and not just it’s idea. That is a tough argument to make.
Opinion: I agree there needs to be some negotiation going on. But making AI developers ask everyone for permission to train on their work is not practical. 90% won’t respond, 5% will respond to late, and the rest of the work might even be bad for training. Couldn’t we just do a burden shift, where every book defines under their copyright a license for usage in AI training? That sounds to me like the easiest practical solution. That would mean if Sarah Silverman makes it public she doesn’t want her books used for AI training, then it’s illegal to do so now.
You are basically circling the important point in your comment: the AI looks at a sentence in the book, tries to predict what’s next, then it gets handed the right solution, and it then backpropagates and “learns” from it.
If you look at the data, it never stores the original work, but it stores a few numbers here and there (really small amount of data compared to the book) to improve at this sentence next time.
In it’s technical form, these are weights that the AI uses on previous text to derive at new text. Now you would need to argue that these weights completely cover every work in so much as they could replicate the original work in its expression, and not just it’s idea. That is a tough argument to make.
Opinion: I agree there needs to be some negotiation going on. But making AI developers ask everyone for permission to train on their work is not practical. 90% won’t respond, 5% will respond to late, and the rest of the work might even be bad for training. Couldn’t we just do a burden shift, where every book defines under their copyright a license for usage in AI training? That sounds to me like the easiest practical solution. That would mean if Sarah Silverman makes it public she doesn’t want her books used for AI training, then it’s illegal to do so now.