NEW YORK (AP) — John Grisham, Jodi Picoult and George R.R. Martin are among 17 authors suing OpenAI for “systematic theft on a mass scale,” the latest in a wave of legal action by writers concerned that artificial intelligence programs are using their copyrighted works without permission.
(I think you’re arguing from an ethical standpoint whereas OP was arguing legally, but anyway…)
No, that shouldn’t happen. If an AI were ever able to recite back its training data verbatim, that AI would be overfitting. It happens by accident sometimes early on in development when your training data is too small and your model is too big, but it’s an error, and is something to be avoided and corrected.
The whole point of training is to get it to a point where it can’t recite back any of its training data. In order for that to happen, the AI is forced to sort of generalize and abstract (sorry for anthropomorphizing) its training data. That’s the only way to get it to be able to generate something new, which is the whole point of the endeavour.
Long story short, if an AI could recite back an entire book, by definition it could not be an AI, and it wouldn’t resemble any of the popular LLMs we have now like ChatGPT. (But you may see snippets and pastiches and watermarks show up)