> But the latest and greatest software trend–generative AI–is in danger of being swallowed up by copyright law.
About time.
> If the AI industry is to survive, we need a clear legal rule that neural networks, and the outputs they produce, are not presumed to be copies of the data used to train them.
But they are compressed lossy copies of all that data! That's the whole point of noise/denoise functions that neural networks are based upon. The whole mathematical foundation of training a neural network is "teaching" it how to recognize and/or create copies of data stored in the training set.
> Otherwise, the entire industry will be plagued with lawsuits that will stifle innovation and only enrich plaintiff’s lawyers.
When you're willingly breaking already established law en masse for profit in hope no one cares enough, be it copyright law or any other, you're not an "innovator", you're a criminal. The fact that you're a tech giant or a Bay startup doesn't matter in this regard; the only thing that matters is the notable amount of time required for the justice system to catch up with your novel tools for laundering intellectual property.
De minimis is a longstanding defense in copyright law. If you are copying very little from very many works, as is the case when you turn multiple petabytes into a few gigabytes of neural network weights, you are in the clear. The problem arises when models overfit and spit out almost perfect copies of the training data.
Copyright doesn't have an explicit size, but rather uses size as one of many indicators.
For example, I could take a massive 8k video and covert it into a very small 144p youtube video. Am I in the clear simply because the output is tiny compared to the input? Similar I could take a huge studio master copy of a song and convert it to a very small and rather compressed (distorted) mp3.
I partially agree that some of the problem is when perfect copies are spit out by the models, but I do think there is a bigger problem. Copyright is a complex concept that can't be defined exclusively by a single metric like size, and any mathematically definition will in the end be killed if large copyright holders feel threatened by it.
ML models do not supplant the pre-existing work, and provide fundamentally new modalities. Transformative use seems like a slam dunk to me, but I guess we'll see what the Supremes decide in twenty years or so...
I’m unclear about this. Let’s say a movie comes out and I make a YouTube review using brief clips or screenshots from the movie. Since my review is transformative, I should be in the clear (I think?).
But when it comes to market harm, does the tone of my review effect the enforceability of copyright?
As in, if my review is negative it would harm the market for people going to watch the movie vs a positive review right?
Reviews have a distinct "character of use", one of the four cornerstones of fair use exceptions.
A review can be commercial, can cause significant harm to the market, can include substantial amount of the work, and yet the character of use can be significant enough to convince a judge that a exemption should be applied. Since judges historically has come to this conclusion there exist now legal precedence. With precedence we can make some general conclusions which tell us that reviews are in general exempted when using other peoples copyrighted work for the purpose of reviews.
This character of use is very different then if I convert a studio record of a song into mp3 and publish it on p2p sharing site. Judges has historically viewed the character of use in those situation as not being worth giving exemptions.
You're not directly competing with the movie though, your work is a review, not a feature film.
If you were to make a parody movie from the material of the movie itself, directly taking scenes and altering them to your liking but still relying on the viewer recognizing the original in it, you'd have a harder time, I think.
There's a Stable Diffusion example where, having been trained on too many Getty Images pictures stamped with their logo, the system generated new images with Getty Images logos.[1] That's a bit embarrassing. There are code generation examples where copyright notices appeared in the output. A plagiarism detection system to insure that the output is sufficiently different from any single training input ought to be possible.
Yes, agreed, I don't think the problem is with networks that mix tons of input data in a way that doesn't heavily derive from one or a couple of sources. The currently available models do not have overfitting solved, though, and this technological imperfection also has direct practical (and legal) consequences.
How are you using the word “copy”? It doesn’t seem to match the standard meaning. For instance, most people would not consider a brief summary of a movie’s plot to be a “copy” of that movie, or protected under copyright.
If you have an image, then train a neural network on that image, then use the neural network to reconstruct that image in detail, then the NN by definition contains enough information used to reconstruct that image - hence, a copy.
With NNs trained on thousands or millions of data entries, this concept becomes fuzzy in the same way as you described - a short summary likely wouldn't be considered a copy, just like a 64x64 generated thumbnail wouldn't be considered in the same way a 4096x4096 hi-res image.
The thing is, the “good” models can’t reconstruct the image in detail. It’s considered a sign of “overfitting” if you reconstruct the input exactly. Even if you put the exact query that was associated with that image, you’ll get the weighted average (feature-wise) image associated with the query. This applies to all like machine learning models without loss of generality.
Sure, but I could write a program to spew out an unbounded number of images containing random pixels. It could create an image that is identical to a copyrighted image, but if I just keep that image on my hard drive, have I violated copyright? I don't think I would be, but if I started distributing them, yes I would.
> If you have an image, then train a neural network on that image, then use the neural network to reconstruct that image in detail
I haven't seen that happening since the discussion started. Most of the complains I saw aimed at things like "it stole my style" not "it reproduced my art".
It's more, 'this product is profiting from my labor without my consent (ie paying me).'
In music you aren't allowed to use the same notes, even if you played them on a trumpet with a swing beat, while the source was on the piano very staccato.
While we don't have the same vocabulary for art, it's not unreasonable to expect similar protections.
It takes work to create/identify/classify information, both in the economic and physics sense. That work should be allowed the same protections we do other forms of work.
Your example is one where nearly no work was done, thus it doesn't deserve much value. "Let a = the set of all songs" doesn't help me find new songs I like. A songwriter does that work. Another artist that takes and uses and resells that work (without consent), is stealing that work.
To me it's funny that nearly all the problem with the current team of AI generation would be solved if the model generators simply licensed the content they train on. "But that would cost too much" Ok, just use public domain work, "But that wouldn't be as good" Oh so you are saying the work has value, but you are unwilling to pay for it, and instead your scheme is to just take it. That seems like a good definition of stealing - not paying for something that has value.
> Your example is one where nearly no work was done, thus it doesn't deserve much value
You are aware that there is very expensive art out there where the artist did not much work. Like painting a canvas in one colour or throwing an item in the corner of a museum.
According to you, that would not deserve much value but it does have a lot value in reality.
In fact "value" is what somebody else gives to the piece of art.
A prompted AI artwork made by me may have more value to me than all the art in the Louvre.
The discussion here continues to turn around copies when it's not a copy those algorithms generate.
> Another artist that takes and uses and resells that work (without consent), is stealing that work.
Another artist accidentally uses a melody from another song (because it's a finite set) and are sued for all their income is a horrible system. The winners aren't the people producing value, it's the people who got there first and are now profiting off other people's work.
If I grew up under a rock, somehow became a self taught musician, and ended up authoring a song that had recognizable components from Happy Birthday, then even still the author of Happy Birthday, having established that melody so successfully in the public zeitgeist, reasonably should benefit.
This is so common the recording industry itself has established rules for sampling and licensing and covers and what not. Are there some folks out there abusing the system, for sure. But overall its goal is to maximize the value produced by the recording industry, which very much includes the people who 'got their first' who built foundations for future artists. To me, this all seems basically reasonable.
Copyright is supposed to promote the creation of new works. You just described a system where a song written well over 100 years ago is preferred over over a new artist creating a new work.
Honestly, can people stop speaking in absolutes regarding these systems? We (researchers and non-researchers alike) are gradually trying to comprehend exactly how much they generalise and memorise, but this is darn hard work and it is not our fault that several major tech giants decided to deploy and profit from these models long before the scientific and legal landscape was clear. Somepalli et al. (2022) [1] for example is a fairly strong argument against your statement above.
The fact is that these systems are complex, new, and interesting. However, it is not the fault of small-time programmers and artists that modern copyright law is a major, overreaching mess that is now finally greatly affecting what the big corporations want to do. They are getting sued? Cry me a river… Perhaps they will finally stop backing the American-led copyright lobby then?
> is a fairly strong argument against your statement above.
From a quick skim of this paper, they apparently used toy models with a few hundred to a few thousand images in the training set. For the ones with as few as a few thousand training images, they rarely or never saw exact duplicates.
For instance, in their figure 4, they show exact duplicates for the training set with only 300 images (well, duh), and didn't find any exact duplicates for the training set with only 3,000.
I'm not sure I'd call this a "strong argument" when applied to models with millions or billions of images. Quite the contrary. LAION-5B (used in Stable Diffusion) was trained on 5 billion image/caption pairs.
Firstly, thank you for engaging in a discussion. Secondly, I am not an expert in image processing, rather my focus in on language. Thus my intuitions will not work as much in my favour in this domain, although the models do have similarities.
They explore a range of sizes and I do not think it is fair to to only highlight the smallest ones. They do explore a 12M subset of LAION in Section 7 for a model that was trained on 2B images. Yes, it is not an ideal experimental setup to use a subset (they admit this) and far from LAION-5B, but it is a fair stab at this kind of analysis and is likely to lead to further explorations.
Let us return though to your claim, which is what I objected to: “Pretty much none of these systems ‘reconstruct an image in detail’.” I think it is fair to say that this work certainly makes me doubt whether none of these systems (even the larger ones) exhibit behaviour that may limit their generalisability or cross the boundary of what is legally considered derivative work.
You may very well be right that once we scale to billions of images this behaviour is improved (or maybe even disappears), but to the best of my knowledge we do not know if this is the case and we do not know when, how, and why it occurs if it does occur. I remain a firm believer that these kinds of models are the future as there is little evidence that we have reached their limits, but I will continue to caution anyone that talks in absolutes until there is solid evidence to support those claims.
Incorrect. Training images are used to generate a latent manifold which might not contain any of the images in the original training set to within a meaningful delta unless they're massively overrepresented or cliche.
This might be technically correct but doesn't seem to matter much in practice because of how many practical cases of NN outputting copyright-infriging content there are. These include examples in the recent lawsuits that made the rounds on HN. Either even the best specialists behind these NNs cannot make the NNs not contain data that is "massively overrepresented or cliche", or they are unwilling to.
"Many practical cases of NN outputting copyright-infriging content there are"
I have seen very few practical cases of generative models such as copilot and even less of a stable diffusion of reproducing original copyrighted works in exact detail, and the few that I did encounter were instructed torturously to do so, which strikes me as highly contrived.
Or they're focusing on performance to get the tools to the place where they're good enough to start being adopted. Once getting sued is more of a problem than not having a product at all, it's not hard to switch gears. I imagine there are a number of ways to avoid storing "too close" copies in a model that have various tradeoffs, I'm sure they'll be adopted quickly in the face of litigation.
This feels like an argument for communism being the most productive system in theory. Most of the time I feel like I see uninspired material that’s tracing it’s own training data.
When humans do it within a delta, it’s considered copyright infringement.
Machines currently are adept at making copies within a delta, hence articles such as this to limit copyright so the people who operate said machines can profit.
> But they are compressed lossy copies of all that data!
Sounds similar to, for instance, songs written by people who have heard other songs. I wouldn’t expect legal cases concerning AI-generated works to be any simpler than legal cases concerning the difference between a songwriter violating the IP of another songwriter or simply being inspired by another song.
Yeah, I think this is the crux of the issue that people keep glossing over when claiming it's not a true reproduction.
Drawing Coca Cola's logo from memory by hand and slapping it on your product is still copyright infringement, even if it isn't an image Coca Cola has ever produced. In that sense it doesn't matter at all if it's AI or human - the production and subsequent distribution for profit of a copyrighted thing is not allowed, period.
The current set of AIs do exactly this all the time. That's a very clear legal problem.
I meant that _learning_ is not a copyright violation.
I do not see much space for misunderstanding: which relevant box takes instances of input data to output one of them?
Edit:
Make your point explicit, sniper... You can hide in front of the ### consuetude of "silent disagreement", but it remains violently annoying. The article contains "nuances" like "derivative work", but the poster is objecting to the article calling «neural networks» not «copies», as «compressed lossy copies», and I retorted that so is everything you learn, and holding them in the "corpus" is not considered "retaining a copy". If you have objections to that, either you present them, or there is no contribution in shaking your invisible head.
About time.
> If the AI industry is to survive, we need a clear legal rule that neural networks, and the outputs they produce, are not presumed to be copies of the data used to train them.
But they are compressed lossy copies of all that data! That's the whole point of noise/denoise functions that neural networks are based upon. The whole mathematical foundation of training a neural network is "teaching" it how to recognize and/or create copies of data stored in the training set.
> Otherwise, the entire industry will be plagued with lawsuits that will stifle innovation and only enrich plaintiff’s lawyers.
When you're willingly breaking already established law en masse for profit in hope no one cares enough, be it copyright law or any other, you're not an "innovator", you're a criminal. The fact that you're a tech giant or a Bay startup doesn't matter in this regard; the only thing that matters is the notable amount of time required for the justice system to catch up with your novel tools for laundering intellectual property.