AI has already hurt plenty of people though; we already have examples of biased "AI policing", artists and writers are getting their work stolen left and right, and I've heard tell of several instances of seriously vulnerable code making it to production and revealing folks' private information due to bugs introduced by AI assistants.
This regulation is almost certainly too far reaching and written by bureaucrats rather than experts, but the need for regulation is already there.
We need regulation that holds people accountable. I see too many 1%ers claiming they were not responsible for something their algorithm did.
In old sci-fi, a computer was the legal agent of its operator/owner and anything that computer did was 100% the fault of its owner/operator.
If we simply say "you are responsible for what your AI does," then the problem is solved.
We could solve the problem if not being able to bring corporations to justice at the same time. Forxe them to beck e proprietorships. Every owner/partner is culpable for crimes committed by that proprietorship.
Did a cop arrest an obviously innocent person? It's the cop's fault.
Did my web site plagiarize a painting? Then it's my fault.
Did a startup release venerable software? Then the crook that exploited the bug is responsible but so it the entity that released the software.
If you read a bunch of books and then write your own book, is that book a derivative of the books you have read throughout your life? I doubt too many artists have never experienced nor been influenced by other works in their medium prior to making their works, but that's not what we are referring to by derivative work. Typically derivative work means something like a movie based on a book, or a translation, or a parody, etc where something of the new work is clearly coming from a previous work. While an AI could certainly produce a derivative work, for example if you only trained an AI on a specific artist's style to produce new works that specifically emulated that style, at some point the link between a new work and those that influenced it is so tenuous that the new work is original.
>If you read a bunch of books and then write your own book, is that book a derivative of the books you have read throughout your life? I doubt too many artists have never experienced nor been influenced by other works
An LLM isn't an artist and I'm not a piece of software. But if you think the LLM output other to be copyrightable in its own right, who should hold the copyright? The person who wrote the prompt? The person who picked which entropy source the model used? The owner of the GPU that ran inference? The last person to provide input for fine-tuning the network?
> An LLM isn't an artist and I'm not a piece of software.
Hard disagree on both counts, but for the sake of argument let's go with your assumption.
> But if you think the LLM output other to be copyrightable in its own right, who should hold the copyright?
The LLM, being a piece of software, should logically be treated like any other piece of software used to create a work by a human artist. It is a tool. Let's say there is a song composed for a synth keyboard. Should the copyright go to the person who designed the keyboard? The person who owns the equipment that made the keyboard? The people who recorded the samples that each key stroke references? The person who curated the samples for the keyboard? The person who owns the keyboard? To the person who figured out the sequence of keystrokes that produces the song? Or to the person who hits those key strokes when the song is being recorded?
Well the person who designed the keyboard gets the ip of the keyboard's design. The person who owns the production hardware does not inherently get any ip. The artists who created all the works sampled get the ip to their specific samples. The curator would not get any ip for the collection but they might if they did something transformational, like passing those samples through a filter. The keyboard owner does not inherently get any ip. The composer gets the ip to the composition in general. The artist who played the song for the recording gets the ip for that specific recording of the song. If I were having a conversation with someone, I would say that the person who figured out the sequence of keystrokes is the person who holds the copyright.
Likewise, the creators of the LLM would hold the ip related to its design, but a specific output of the LLM would belong to the person who thought of a way to make it produce that specific output - ie the prompt writer. Those who made the inputs would own the ip to the part of the inputs which remained untransformed in the final work. Curating input for the network training would be equivalent to curating numbers for a phone book - the collection itself is not copyrightable but some transformational work (like cleaning the data in a specific way) would be. Owning the hardware the LLM was trained on would not inherently grant any ip rights.
Of course there is some grey area where it might genuinely be unclear who holds the rights to what (which has always been an issue with intellectual property laws), but the idea that just because it's sometimes unclear who contributed what to a work means no work was created at all is obviously folly. Others might have a different opinion than I on how the existing framework should be tweaked to accommodate this new technology, but it's definitely not a major departure.
> I've heard tell of several instances of seriously vulnerable code making it to production and revealing folks' private information due to bugs introduced by AI assistants.
By this standard, we should ban all integrated development environments because, like the AIs you mention, they allow programmers to produce hugely buggy code.
> artists and writers are getting their work stolen left and right
Please show me a few examples where a work was copied by an AI and the copy is so good it violates copyright law.
All I have seen is someone used a pirated dataset to train AIs. Suing over that is like suing Seagate because someone tested a prototype hard disk by storing pirated books on it.
> AI has already hurt plenty of people though; we already have examples of biased "AI policing"
Cops hurt innocent people. That's not the fault of the Chevrolet they are driving or the AI they are using.
Criminals will seek to obfuscate blame for their crimes to avoid getting caught, so police will frequently get the wrong person. Libertarians would rather leave it to lynch mobs, while statists want to fund better tools for the police. No solution is perfect.
This regulation is almost certainly too far reaching and written by bureaucrats rather than experts, but the need for regulation is already there.