Allowing these systems to utilize copyrighted materials without compensating the copyright holder is a legal loophole the size of the sun. Then it becomes a perfectly legal defense to write "an AI" that slurps up copyrighted material and produces results that are "based on" that material.
My issue is that the copyrighted work is fundamentally integrated into the algorithm. Say you wrote a program that would input a copyrighted work, then prints out the contents, but with some words replaced with synonyms. Would the output of that be okay? What about if phrases were replaced? How about if entire sections are added, rearranged, and/or removed? At some point, the results become so mangled as to not be recognizable, but the fact that a copyrighted work serves as an input to the program is still an issue, IMHO.
A major difference between humans and computers is that a human can't (generally) recall hundreds of pages of text and regurgitate it on demand. So a human who has read a book only retains a fraction of the contents, and even then, generally they don't retain all of the information accurately.
My point is not to say that contemporary "AI" systems never emit things that are arguably copyright infringement. It's more just to say that there's nothing about being an AI that makes it necessary that its output be copyright infringing.
The only point I can see making is that for some reason I doubt openai actually paid the author for the contents of the book, but maybe they did.
Maybe they borrowed ebooks one by one from some digital library and ingested them that way for free.
Much more "scalable" i.e. easy to get raw training dataset. Move fast and break things
But the comparison could just as easily be written, "One is a computer program, literally designed to learn from copyrighted works, the other is a human being that learns from copyrighted works"
If ChatGPT is parroting entire paragraphs of copyrighted work in its answers, then "pilfer" is probably the right word. Or maybe "plagiarize". But if it's training on the information gleaned from the work, and using that training to synthesize the answers, isn't that a lot more like how a learned person applies their knowledge?
I haven't yet made up my mind completely on how IP should interact with AI. I see good arguments for and against. But the answer is far from obvious!
A human can reason, make decisions, act in best interest, judge context of a situation, apply it to new scenarios, judge emotion, and create a new derivative work.
The computer program is literally taking someones work and adding fancy way to search through the other person's work, without adding any value.
We need to stop fantasizing these fancy toys are actually any sort of intelligence.
That strikes me as a rather extraordinary claim. It seems obvious to me that contemporary AI's are "some" sort of intelligence, albeit with open questions around "what kind of intelligence?" and "how intelligent are they?".
What if a person bought a copy of the book, read it and learned the material, and then provided advice to people based on the material. I think most people would say that that is fair. So why is it less fair for the AI to do what is effectively the same thing? Is it simply a matter of scale? Or is there something more fundamental at work?