Copyright and AI-generated works
This paper is based on a study made for the European Commission on the ‘copyrightability’ of AI-assisted work.
This one highlights the variations in definitions of a ‘work’ for copyright purposes between EU member states. So, it would seem that AI generated art, provided it had an original prompt, choice of software, etc., might pass the standard in some member states, but not all of them.
This paper and the next one both highlight the importance of process in dictating what is and isn’t copyrightable. Which is an interesting contrast to my understanding of copyright infringement, which doesn’t care about process at all and is only judged based on outcomes. This leads to the interesting issue where you could have an AI generated work that isn’t protected by copyright but would nonetheless make you guilty of copyright infringement.
This dichotomy is even starker in the US: no copyright protection for machine-generated works but those can still make you guilty of copyright infringement.
Because infringement is a matter of outcomes, not process.
Another way of looking at this dichotomy is if a monkey with a typewriter managed to spontaneously write a near-exact copy of Stephen King’s “Misery”.
The copy wouldn’t have copyright protection (no human author) but you’d still be liable for copyright infringement if you published it.
Same thing with an AI that stochastic parrots it’s way to create something that turns out to copy an existing work.
As regards such AI systems, where users are effectively no more than passive “players”, the user clearly does not have a valid claim to authorship in the AI-assisted output (i.e., in anything beyond its initial prompt) – leaving the developer of the AI system as the only candidate for authorship of the AI-assisted output.
Sounds like developers like Stability AI/Midjourney would be the likely copyright holders in the EU countries where such works qualify for protection.
It’s pretty clear that this is another area where the EU needs to harmonise its rules on AI. I wouldn’t be surprised if, as the field grows, they end up acting pretty quickly to settle on a uniform higher standard that brings it more in line with how everybody believes US law works.
But until then it’s all quite uncertain.
The bar is probably high enough to ensure that very few AI-generated works qualify for copyright protection, but this will probably be one area of focus when it comes to the EU’s efforts to harmonise its rules on AI.
Qualification of AI creations as “works” under EU copyright law: AI-assisted versus AI-generated works
Another paper on the same subject. The answer to “are AI works protected by copyright?” seems to be ‘no’ in the US but EU law is less clear.
According to this, current AI art systems are “AI-assisted” as they require direction (the prompt). But whether that and the user’s choices together are original and creative enough for the end result to qualify as a ‘work’ (a precondition for copyright) is a solid ‘it depends’.
Also, whether that copyright belongs to the prompt author or to the creator of the AI software is another ‘it depends’. The choice of training data and weights is, according to this, possibly a more ‘creative’ decision than the prompt and might have a stronger claim to copyright
In general, this paper describes a pretty high bar that an AI-generated work would need to pass for it to get copyright protection. Most works wouldn’t pass it but there might be edge cases that do.
My impression of both of these papers on the copyrightability of AI works in the EU is that the way Github Copilot works means the code has a decent shot of qualifying for copyright protection there.
There is user choice, redaction and editing, iteration, and then integration into a final creative product. Which makes OpenAI’s and Microsoft’s focus on AI-generated code even smarter, IMO.
Which brings me to…
I have mixed feelings about the use of tools like Github Copilot. There seems to be sufficient evidence that the code you end up with is of a lower quality than without AI help, just accomplished much faster.
The suggestions will always be at least partially flawed, due to the basic approach, which, when combined with automation bias and anchoring bias are likely result in a noticeable bias towards worse outcomes. Noticeable, that is, to everybody except for the coders themselves.
That said, worse code that’s written 40% faster is exactly the sort of tradeoff that the software in general loves to make. The software industry has a horrendously low standards for quality and defects but universally loves to ship the crap faster. By focusing on improving code output, OpenAI might well get Copilot to a sweet spot that would be impossible for the industry to ignore and become the multi-billion dollar business it and Microsoft are hoping for.
So, it looks the right strategic move for OpenAI.
Links and Notes
No, large language models aren’t like disabled people (and it’s problematic to argue that they are) - by Emily M. Bender
I’m honestly saddened that this needs to be said.
Directly relevant to the Copilot issue. One way to mitigate the increase in defects that’s almost certain to result from the use of Copilot is to lean heavily towards choosing more verbose and readable suggestions from the tool.
Another issue with much of the research we’re seeing with AI is that a lot of it is, honestly, bullshit.
This draft paper would imply that a lot of the research on the efficacy of AI/machine-learning approaches is too flawed to be valid.
Despite the furor around ChatGPT at the moment, I’m not excited. I don’t think it will be the doom of academics or even current approaches to assessment and I also don’t think it is (or will be) a very good tool to incorporate into learning activities or any creative work. I think its indicative of some places where we might have new tools to help solve problems and spark creativity. In terms of learning technology we will eventually need to pivot our techniques to incorporate good AI assistance (or assistants), but I don’t think that’s where we are in January 2023.
I think about this all the time—that the actual amount of time spent in doing something creative (writing, designing, making music, whathaveyou) is often buffered by hours and hours on either side by real—sometimes pleasant, sometimes infuriating—boredom.
Flood the zone with bullshit (or facilitate others doing so), then offer paid services to detect said bullshit.
Seeing people adoringly praise AI art on social media is like having that “so, wait, you actually like Adam Sandler comedies?” moment several times a day, every day.
The quality of AI art seems inversely proportional to the amount of praise given to it by the initial poster.
- AI-generated code helps me learn and makes experimenting faster . If you’re going to use Copilot, this is the way to do it. Not including the generated code in the final product also hedges your bets on both the lawsuit and on the question whether AI-generated code qualifies for copyright protection.
To help you best, I’d need to know you . AI business models are going to collide so hard with the GDPR. I don’t see how we’re going to avoid a trainwreck of epic proportions. And fines.
ChatGPT Cites Economics Papers That Do Not Exist – Economist Writing Every Day . Here’s a bunch of economists surprised that ChatGPT consistently makes shit up.
Deno 1.30: Built-in Node modules . I’ve been using Deno quite a bit for a side project and it’s honestly been a joy.
Software and its Discontents, Part 2: An Explosion of Complexity - Kellan Elliott-McCrea: Blog . Good post. Doesn’t touch on the management issue, though. That is, current management tactics are a very bad fit for most software projects.
As ComiXology fades, the industry reacts – and wonders what’s next . An overview of what’s being written about the ComiXology layoffs. My first reaction: “wow, comics people sure are better writers than tech people”
Why Nintendo’s Satoru Iwata refuses to lay off staff - Polygon . From 2013. But still good advice.