[flagged]
If you want to learn to draw in the style of Moebius then go for it. If you want to train an AI on Moebius' work, you should ask permission from Moebius' estate, since they stand to lose revenue by your model.
I'm not talking about infringement.
That's the entire point. It's in the US constitution, stating that exact reason.
Starting a legal precedent that copyrights spreads like fungus and encompasses any thought related to the copyrighted work seems like a horrible idea, one that can only lead to a dystopian future.
I think that is a good argument, but conflates is and ought, and I have two counters:
copyright owners can dictate how their work is used (with some exceptions), and if that use hurts the copyright owner, the owner should have the right to forbid it.
the intent of copyright is to reward and encourage creators and creativity. If a script kiddie can just train a model and duplicate the hard-won aesthetic work of Molly Crabapple or Ralph Steadman or anyone at all, and either dilute the value of it or actually profit from it, what is the incentive for creators to create new work at all?
Consider a poet who publishes poetry in some unique meter, or has some other unique stylistic structure for which they are well known... should they be allowed to sell copies of their poems that can be used for reading only, but prevents usage of those stylistic devices by other authors?
I'm going to assume that we agree the answer is "no, the author should not be able to prevent those uses" at least for human consumers of their works. This is how art has always developed... even though that use "hurts the copyright owner" by diluting the market for works with that style, the owner does NOT have the right to forbid it.
Now, let's say that same poet drew a lot of inspiration from a bunch of out-of-copyright poets. Let's also say that I train an AI model on the poet's inspirations, but NOT on the poet's work directly. Then I ask the AI to write a poem in the style of the poet's inspirations, and to include the unique stylistic device for which that poet is famous. In your world, is this OK?
I don't think this is a fair analogy. Unfortunately, analogy breaks down because the technology is unprecedented. So, to answer directly, no the poet cannot copyright the unique meter, but, no, machine learning is not that.
If you need an analogy, think copy machine, not human learning. An LLM can only regurgitate that which it has seen before. Absent the poet, the other poets can still make other poetry, but the LLM literally cannot make poetry that it has not seen before. If it produces a poem with that unique meter then it definitely copied that poet, and was not "inspired by" the poetry. If you wrote poetry inspired by EE Cummings your process for doing that would be very different from an LLM's, which would programmatically use his material.
Full disclosure: I think I probably disagree with you on some points you've made in this thread, but I'm not going for any gotchas right now, I am just trying to map the contours of what you think is OK and not OK. We're all sort of flying blind on this stuff, so getting a sense of what others are thinking is really important in my mind. Appreciate the engagement.
Like if I read their books and then write better ones in their style, undermining their profits?
again, for the people in the back.
> from reading your books or seeing your art if they obtain a legal copy.
You're arguing against something I didn't say
We should still obviously do it. More of a thing that people want is usually good.
The whole point of the phrase was to describe a hypothetical situation with no side effects to avoid sideways arguments about "but actually here's some bad things that would happen if you did that unrelated to the central argument".
So let me try again: in my view, you shouldn't think of policies or laws impacting real people by ever placing yourself outside of the reality in an ideal case, because the hard bit is not conjuring ideals but finding a way of making them happen.
It's always a lot messier where the rubber meets the road. People have already died and suffered because ideals (specifically around asset ownership) that weren't quite thought through, but caught on. Take communism as an example.
Part of my point is that such "implementation details" are not as unrelated to the central argument as they seem. This is very different from the software world where it might be ok to assume that in 2 years we'll have the computations be 10x as fast and work out a solution backwards from there.
???
They are not creating anything unique or otherwise. It’s not remotely comparable.
There is a distinct lack of capitalist incentive to spend your time developing a novel style only to have your style replicated for anyone to use, based on your work, and without permission.
In terms of automation used to destroy incentives for original creators, based on original creators’ works, photocopying machines and generative models are in the same quadrant.
Please say so if that’s your point of view.
Because the argument is the same either way.
Either we protect creative works from being devalued by automation, to maintain an incentive to create, or we don’t.
My entire career has been selling copyrighted works for good money. And I have been able to confidently do that for many years because of copyright protection.
So “it just delays the inevitable” means what? I have to give my money back later?
The inevitable is that your industry will change and the distinction between what is an "Artist" and an "Operator" will change even more. If most of your clients come to you for the end product you are most likely an "Operator". ie. "Do X like I want it to be done." if you are an "Artist" clients come to you because you are either in the Zeitgeist or people like your way of thinking and the process behind your art. The end product is collaborative.
If you are an "Operator" you will have problems in the near future if you are an "Artist" you will be fine. It's like in VFX and the mark the writers strike made on the industry the "Aritsts" are all fine because the studios want to retain them and the "Operators" are left on the street.
And I agree, requiring permission to use copyrighted works in model building will give legs to the current paradigm, but only delay the change.
However, AI is going to upset everyone’s apple carts, so anything that allows changes to happen more smoothly, less disruptively, is probably with the effort.
AI is likely to devalue all human labor, except for the provenance value of creations (creator, history, associations) and a preference for the human element (many personal services, or services with a personal touch element).
Imagine what I could accomplish if I was a tech startup flush with cash unencumbered by the law. Surely a planetary hostile takeover would only be a few years and existential gambits away.
And lots of things that aren’t generally protectable like new artistic styles and techniques are co-opted all the time by other artists.
The second screenshot is a diff that compares the LLM output with the original book.
It also works with Harry Potter, but the API behaves weirdly and after producing correct output in the beginning really quickly, suddenly hangs. You can continue generating the correct words by doing more API calls, but it only does a few words at a time before stopping. It clearly knows the right content, but doesn't want to send it all at once.
I think there is some output filtering for "big" authors and stuff that is too famous that they filter in order to avoid getting sued.
I wrote more details about the weird Harry Potter behaviour here: https://news.ycombinator.com/item?id=37614697
And just because something could do something doesn't make it a violation.
How many bytes do you think a million compressed books takes? Consider that the way these models are trained is basically completing the next symbol based on the previous words, which is how most compressors are made.
I wonder if this might be because Slavboj might be a monist/physicalist, and you might be a dualist[2]? If that's the case we'd all argue until we're blue in the face, if we don't at least recognize this underlying difference. For the record, since I've studied biology, I'm probably closest to some form of mechanism[3] (due to the rejection of vis vitalis [4] in early 20th c. )
Of course, you could also just be a very skeptical monist mindful of the kluger hans effect[5] and working from there.
Let me know which (if any), maybe we can still find middle ground!
[1] https://en.wikipedia.org/wiki/Physicalism
[2] https://en.wikipedia.org/wiki/Mind%E2%80%93body_dualism
[3] https://en.wikipedia.org/wiki/Mechanism_(philosophy)
This is all becoming akin to the 'but on a computer' for patents.
It's exactly the issue here. The capability doesn't matter, what matters is the scale of it.
Would you be ok with your car billing you every single time you exceed the speed limit because "there already are speed cameras out there" ?
It's like saying I can kill people with a swiss army knife so I should be able to own a nuclear bomb since it's _literally_ the same thing, besides the scale but that's a detail right :)
One solution to this would be to say this:
* The copying done while training a neural net is "fair use", similar to copying the DVD into RAM and onto the screen while watching it.
* The resulting neural net is a derivative work of all copyrighted works used during its training: No copying of the derived work is allowed without permission of all copyright holders
* The output of the neural net is subject to normal copyright laws for humans: i.e., it's only a violation of copyright if it's obviously a copy.
Basically, you can either train every instance of a neural network separately (like you have to do with humans) or get a license for all your training data.