The test for non-literal copyright infringement is "substantial similarity." If, after filtering out irrelevant and non-copyrightable elements, the allegedly-infringing work is substantially the same as the original work, then it infringes. If it infringes, then two common defenses are independent creation and fair use.
In your hypothetical, the AI-generated work would infringe the original because you stated it would be substantially the same as the copyrighted work. You can't claim independent creation because the algorithm was dependent on the original work and you controlled the output of the algorithm to be exactly like the original work. Fair use is pretty much a non-starter, so I'll skip that analysis.
So, no, you couldn't use an AI to launder copyrighted works into the public domain.
If you are using it to mix a snippet of code (from a sufficiently large code base) into a large code base of your own, then you are just remixing. That is not infringement. In music, there are entire genres based on remixing. You could even take it a step further and ask yourself: what is not a remix?
Point is songwriters absolutely get litigious over reproducing small portions of their IP.
complex.com/music/majority-sisqo-thong-song-publishing-owned-by-writer-livin-la-vida-loca
huffpost.com/entry/katy-perry-dark-horse-lawsuit-payment_n_5d43d825e4b0acb57fca3ff2
factmag.com/2016/06/25/sampling-hip-hop-copyright/
What's more surprising is to see copyleft advocates positioned so strongly in favour of giving copyright that kind of reach. I think that in a different context, some of the cases you refer to would be used by these same copyleft supporters as examples of why copyright needs to be more weakly enforced, not more strongly.
You have to filter out any non-copyrightable elements before you do the substantial similarity analysis. For code, that means removing non-expressive elements like arithmetic or boolean expressions, looping, recursion, conditionals, etc. APIs are not copyrightable under the recent Supreme Court holding in Google v. Oracle.
How much of your code is actually left after filtration?
The tests exist as a way of determining if someone did the action of copying which is the important thing at the core of it. And in this case the facts aren’t really in dispute, it’s whether what GH is doing counts as copying.
If you had a really really good memory and remembered almost exactly how your former company implemented something and when faced with a similar problem unknowingly produced similar code. It’s iffy as to whether this is copying. Because it really probably isn’t — you’re allowed to learn from copyrighted works — but the courts aren’t omniscient and when presented with the code they very well may rule that copying was more likely than not.
But we are omniscient in this case so we don’t really need the tests. Is what GH does more like copying or learning? This isn’t something that can be determined purely from the output of the tool.
We could remove the non-copyrightable parts of text works too. Just take out all the basic building blocks of language like verbs, nouns, stop words...
But perhaps there could be a way to make something that automatically converts these "non-expressive elements" into copyrightable elements?
Output would probably be maddening to figure out though.
The AI part isn't about independent creation but about figuring out what to copy.
Yes, this is what is pretty interesting to me. I said in a previous comment that I have a really good OS generating AI. It asks you your favorite color and outputs a disk image you can use as an installer.
Right now it just happens to output a cracked version of Windows if you answer "blue". Who can know how that happened? It's a black box after all. Seems useful though, since Microsoft is loudly saying that if I distributed this it would have no license problems at all.
If when you crack open Copilot it’s determined that it’s not actually learning and boils down to storing and regurgitating snippets of code no matter how few or many layers of indirection it’s still infringement.
What your AI actually does is underneath all the indirection is what’s important.
Some would say that pretty much _all_ learning involves "storing" and "regurgitating".
Aside: my daughter started writing out her name at kindergarten a couple of years ago. One of the staff seemed a bit dismissive of this, and claimed what was happening wasn't really "writing" it was "merely memorising the shapes of letters and then reproducing them in the right order". <rolls eyes>
I didn't bother to argue...
We know what it's doing, it's doing very well understood statistical inference techniques to derive outputs.
i = i + 1
That not an existing aboutme page. You can go to davidcelis' website and verify that it's completely different.
Copilot just picked a random person and linked to their social media accounts. You can search any large quote within that about me on Google and not find a match, it is unique.
The only two examples of generating large sections of copyrighted work are the quake floating point hack and the zen of python. Both those examples are commonly known and copied and talked about, to the point that they have wikipedia pages.
In fact, that's a crucial selling point for the product.
I think the big question is, if copilot ends up copying significant portions a GPL work, not just tiny snippets, is the resulting work infringing, and if so, who is liable?
I have almost no sense of how often code is infringed currently and how often anyone does anything about it. I have a feeling that we live in a world with constant infringement, basically no one cares, and no one does anything about it. And I would assume, the status quo will maintain its current course with this new tool. But again, I'm giving zero factual evidence, it's just a feeling from not seeing our hearing almost any news about open source code infringment.
But there is also caselaw (involving George Harrison IIRC) on "unconscious copying", where having heard a piece is suggestive that it was not an independent creation, despite not being deliberately copied. So, training on a corpus that includes a specific piece is arguably a case of that.
There"s an interesting question of whether a model is just a sophisticated statistical compression of a corpus, or whether it is a thing in itself. I would say, if it finds patterns that are disproportionately simpler than the corpus, it has found "something".
But another view is of creation as involving a side-channel or one-time pad, in that music is created by a human and heard by a human, who have common information that has never been present in music before (e.g. specific aspects of common neurophysiology, auditory anatomy, exact heartbeat waveform, new sounds/rhythms in the world, new speech patterns, associations between existing melodic fragments and words/emotions/visuals/status etc). In this sense, truly new music is discovery of the Human Music Processing System, which ultimately involves the whole human and their social and physical experience.
If it were intelligent, it’d be given the Lagrange specifications and then I’d be able to say “Write me an open world video game based around gang culture and I’d like it to run on the raspberry pie zero.”
Valid point - most hilariously poor example to back that point up in recent memory.
I.e. if you buy a piano or a guitar, you could play and record copyrighted music on it. That's not piano's or guitar's fault though, it's yours.
Suppose a model was trained solely on a single Beatles album. It could only spit out that album. That would be clear infringement, wouldn't it?
No. It’s not. A Markov chain has some very specific properties that are absolutely not fulfilled by GPT-3 models.
Just say “stochastic” if you want a buzzword. Stop appropriating Markov chains.
A Markov chain is a model where: (1) there is a state, (2) the probabilities of the next state depend only on the current state.
That could describe anything from a wet fart to a deterministic computer to GPT-3 to the human mind to quantum mechanics (the real one, not a simulation).
If I unplug the Internet and all USB devices, my computer is a Markov chain with at least several trillion bits of state, so 2^(several trillion) possible states. And there is one next state, which has a probability of 1, and all other possible states have a probability of 0. That's a Markov chain.
GPT models choose the next word probabilistically, with the probabilities chosen by feeding the previous N words into a neural network. That sounds like a Markov chain to me!
What properties does a gpt-3 model have that a Markov chain doesn’t? (Other than effectiveness.)
GPT-3 is conditioned on the entire input sequence as well as its own output, which is strictly NON-MARKOVIAN. In fact, the point in saying something is Markovian is exactly that: the state transition probability only depends on the current state.
"Given a prompt, provide a completion" is what a Markov chain does. GPT-3 is exactly the same, in the sense that apples and oranges both satisfy your hunger.
Edit - googling, the history of player pianos vs copyright is interesting
https://slate.com/technology/2014/05/white-smith-music-case-...
https://www.techdirt.com/articles/20100712/18325210185.shtml
The former for making and selling it, the latter for buying and using it.
Just like counterfeit goods.
No?
This analogy train went too far, don't you think? All examples that I've seen on Twitter require quite an intentional manipulation by human for Copilot to produce something copyrighted. It does not recite Linux code by pressing 1 key.
Surely a judge presented with the "complex series of button pushes," otherwise known as playing an instrument, would hold the player accountable for any infringement and not the piano?
These analogies have gone so far off the rails that I can't tell which side this thread is arguing for by now ;)
One end is GitHub's, at the input: Copilot's "database" was initialized from code that GitHub does not have copyright to. The contention at this end is that they are ignoring the licenses that would grant them the right to use that code.* The article, GitHub, and others assert that there's no copyright issue for creating a database of this kind (a machine learning model).
The other end is the the developer taking Copilot's output. The article seems to take the (absurd IMO) position that there's also no copyright implications here, because the output is not copyrightable at all.
*And personally this is the side that concerns me most.
Or maybe it is, but if so it essentially means the end of licensing because it would be trivial to make an AI that can take an input and produce the same output. Or maybe even cp is good enough to strip the source of its original license in that case.
Open source licenses are worth protecting or you break the cycle that helps more software be open.