A clearly defined set of rules
As someone who works in a copyright-adjacent field... calling intellectual property laws "clearly defined" is a
bit of a stretch, since (at least in the US) the actual
intent behind the law has been folded, spindled, and mutilated over time. The subject of AI art also falls into pretty weird territory here, since it's a case that the people who drafted the laws in the first place
very much didn't consider.
To go into a bit of
theory for a second, let me trot out the WEMI model. Basically, any piece of media has four levels of abstraction:
- The Work is the most abstract level — this is the intangible "idea" of, say, Dracula.
- Once you have a Work, you or someone else can produce an Expression of it. This would be, say, a particular adaptation or translation of Dracula.
- Once you have an Expression, you also have Manifestations of it. This would be a particular print run of the French translation of Dracula.
- Individual instances of a Manifestation are referred to as Items.
The idea behind copyright is that the only part of this process that can actually be
directly monetized is the sale of individual items (since none of the other levels are actually something
physical) and that every level takes time and effort to produce. To avoid a situation where an artist/author would have to produce all four levels
by themselves before seeing any money, we ended up with a system where you can offer someone else the privilege of being able to produce a manifestation or expression of your work in exchange for a cut of their profits.
Copyright laws basically just protect that entire ladder, with three added quirks — we have the
public domain, where ownership of a given work defaults to the public after a certain period that serves as a combination "take your time producing nice art" and "here's your retirement fund", and
fair use exceptions, because the dividing line between a novel work and an expression of an existing work is
way fuzzier than you might assume, and we needed to nail that down for legal purposes. Finally, you are allowed to, just like
any contract, define the terms of the agreement when you offer people privileges
any way you want, as long as the end results are legal (since a contract that contractually binds you to do something illegal is not actually contractually binding). This has led to the current system of "work-for-hire" (where you produce a work and then sell the right to produce expressions for a single flat sum instead of licensing it for a gradual payoff) becoming quite popular (read: corporations realized that they stood to gain more money that way, and
their lawyers are better than
your lawyers).
It's actually fairly simple
in intent (though it can get
rather complicated once you get to works produced by multiple people). A few issues have arisen since the original laws were drafted, however:
- While corporations can't actually hold ownership of a work, they can hold exclusive rights to produce expressions of that work. This has resulted in Disney continually lobbying for copyright extension up until a few years ago, since that meant that they held the copyright for Mickey (the ability to produce copies of old Mickey Mouse comics and films) and a trademark on Mickey (the right to use Mickey Mouse as part of their branding and corporate identity — this is arguably the actual reason they'd be suing you for using Mickey Mouse in your web comic). This was not an intended use-case of copyright, and has basically turned it from a way for starving artists to feed themselves to a way to fill rich people's pockets.
- Digital media completely upends the economic structure copyright is built on, since it means that the manifestation stage goes from being one of the most expensive steps (since you need to set up the factory to produce the stuff, and then need to pay to produce each item) to being the cheapest (since copying a piece of digital information is, for all intents and purposes, free). This is why we've seen a resurgence in commissions and patronage (where you're paying the artist directly for an expression of their work), and why software companies bend over backwards to give you licenses to use their products instead of selling you something (since, legally, you can make duplicates of items you actually own...)
- Procedurally generated content (of which AI art is a specific instance) is something that feels like it should be a violation of copyright, but isn't actually a violation of copyright... because it raises questions of who/what can produce works.
Actually, let me break out of the list format to address that more fully. The
reason why AI art doesn't inherently violate copyright is because, as far as I can tell, the
real debate is over whether or not an AI (a non-sapient process) can produce a novel work, or if the results of such a process
must be expressions of an existing work, however mutilated (aka
exactly what fair use laws were designed to cover). What adds to the trickiness here is that the software
producing these pieces are actually black boxes — you can't actually open up GPT-3 to see how it ticks, since it's all a collection of weighted nodes on a graph measuring patterns.
I actually don't really have a concrete stance on whether or not I approve of or hate procgen art. I
do think that it's conceptually thorny territory, since it raises the possibility of having works that, legally speaking, came out of nowhere, and that you
can use it to violate copyright if you do so with intent... but I don't think that it
has to be used that way.
...
The hilarious thing about the attribution argument, by the way, is that you really
can't do anything better than "we trained our AI on this collection of information" — the only information the AI actually has available to it when you make your request is a set of probabilities. It'd be like feeding a Markov chain the works of Shakespeare, making it produce a sonnet, and then expecting it to have attribution information for
which sonnets it pulled imagery from.
Oh, yeah, and it's also really funny to me that people get
super passionate about this topic when it comes to pretty pictures, but consider it a novelty when applied to written works (because
boy howdy is language hard), a selling point when applied to level designs (given the popularity of roguelikes and roguelites), and bitch about computers taking over a
completely different part of the process for music (where most of the grumbling has to do with artificial enhancements of an
expression of music (autotune and post-processing) rather than composition). Just goes to show how
complicated this topic is, huh?
EDIT: If you're wondering how I know this... I went to library school for a graduate degree. While there is a specific exception in American copyright law to allow for the legal existence of libraries, we
do need to know this shit, because people can and
will ask us about it.
I'd prefer it if people
didn't ask me to rattle off other "fun facts" about intellectual property law, though, since this post took me a couple hours and gave me a
bit of a headache.
