What does copyright say about generative models?

https://www.oreilly.com/radar/what-does-copyright-say-about-generative-models/

by BerislavLopac • 4 years ago

65 83 4 years ago

Rochus5• 4 years ago

What was originally intended to protect artists has turned into a rent-seeking game in which artists who can afford lawyers monetize the creativity of artists who can’t.

It's rather a rent-seeking industry; the vast majority of artists benefit only marginally from copyright; an original intent, to give composers an income who would otherwise be out of the monetary loop, is long forgotten; instead, early on it was all about protecting the publishers' business by restricting copies; ironically, composers (or musicians in general) today earn best, on average, when they work as a clerk for a copyright collecting society; I don't think patent and copyright law can be fixed; they can only make it even more complicated and unwieldy, so that it gets even further away from composers and authors, and instead plays into the hands of trolls or monopolists.

fixing copyright law to accommodate works used to train AI systems, and developing AI systems that respect the rights of the people who made the works on which their models were trained

And then also charge each student of art for studying original works with the intent to create new works based on what they have learned to make a living? This idea can be extended in any direction and quickly leads to a system that is rather against an open society where people benefit from each other.

ROTMetro• 4 years ago

The system sure seems to have created a ton of music. I know many people who have benefited from being able to actually earn money from their work and would strongly disagree with you. I respect your opinion, but that is all you have written here, your opinion, not some great truth about copyright.

dcow8• 4 years ago

You present a false assertion: that owning the copyright is a predicate to receiving money for creative works.

jerojero• 4 years ago

Is it the system, or is it the access to technology and education? Tools like the internet allow pretty much anyone to learn so much about how to make music and not only that but being able to publish your music very easily as well as forming communities, meeting other artists, etc.

I'm not sure if it's copyright that has really allowed for music to flourish, I think that's a claim that requires a bit more research behind it.

Rochus5• 4 years ago

It's not just an opinion; I'm a musician myself and had many production contracts in the past; I also studied law and am well familiar with the business and the legal issues.

There are indeed some people who were able to make a good living from the money they received from the copyright collecting society, though they are a small minority; all musicians I know who are members of a copyright collecting society get a few hundered Swiss francs a year and don't live from their music.

Some years ago I analyzed the yearly business report of SUISA, the Swiss copyright collecting society responsible for musicians, and found that of the 120'000'000 CHF payed by the society only one sixth was given to Swiss composers ("Urheber"), and that the probability of any of those composers would get enough money from SUISA to make a decent living is less than 0.1%; for comparison, the probability of being seriously injured or killed in a traffic accident in Switzerland is > 0.2%; the vast majority (> 97%) gets nothing to less than 1900 CHF.

In comparison the administrative cost to run SUISA in the same year was 26 mio CHF (compared to the 20 mio CHF payed to musicians), and the publishers received 54 mio CHF. SUISA payed ~19 mio CHF to their 166 employees; in 2006 357'000 CHF was payed to the director alone.

coldtea6• 4 years ago

The system sure seems to have created a ton of music

Because of it, or in spite of it?

heavyset_go6• 4 years ago

ML isn't the same thing as human students, it's just computational statistics with some impressive results.

Consider it a black box where copyrighted work goes in and similar work pops out. Conflating what is essentially a function approximator with human beings when it comes to law is a stretch.

Rochus5• 4 years ago

a black box where copyrighted work goes in and similar work pops out

If a generated work is similar to an already existing copyrighted work, it is not relevant whether the work was generated by a DNN; the author of the original work can file a claim anyway. On the other hand a work generated by a DNN by itself is not copyrightable, since a human author is required by copyright law. The use of training data for the DNN is fair use according to jurisdiction in comparable cases.

BerislavLopac• 4 years ago

ML isn't the same thing as human students

Are you sure about that? https://www.smbc-comics.com/comic/themes ;)

dcow8• 4 years ago

There is no world imaginable where the concept of owning ideas should be protected. Imagine if chefs had to license recipes…

The arts communities need to take cues from the scientific community where citations are the cultural norm. And as a society we need to figure out how to protect artists by protecting and celebrating the expression of ideas. Not the ideas themselves.

thewataccount• 4 years ago

And as a society we need to figure out how to protect artists by protecting and celebrating the expression of ideas.

Github copilot would (until they manually changed it) spit out the famous fast inverse square function word for word - comments and everything which is more than just the idea (why would you want the comments with them swearing?)

dcow8• 4 years ago

I'll admit, writing something down is where it gets very murky.

On one side, you have music and musicians, where the written form is somewhat of meaningless tool used to produce creative expression, which is the performance of sound on an instrument.

On the other hand you have authors (and recently, software engineers), where the creative expression as most understand it, is in the choice of arrangement of words (or statements).

You can't really resolve the two. And so I lean towards a stance that it is fraught to try and build a flourishing creative society on the idea that simply writing an idea down (or saving a file of data) means you own it and can legally bring a case against someone else who happens to write the same idea down, be it music or a book.

And then there's https://libraryofbabel.info, which contains all possible combinations of characters that every has been, or ever will be, written. Should that be copyrightable?

Anyway, I am of the opinion that Copilot does infringe on copyright as we know it. But I'm also of the opinion that simply giving individuals universal claim to any written text they produce is problematic. And to answer your question, maybe I do want the comments. Maybe they communicate something that the code cannot, swearing be damned.

I'd prefer a culture of citations for written/stored/saved works. And a culture of celebrating performance of the arts, not storage of them.

EMIRELADERO• 4 years ago

There is no world imaginable where the concept of owning ideas should be protected.

Copyright does not apply to ideas, it applies to specific expressions/implementations of those ideas, which is what it seeks to incentivize. The fact that many companies (particularly Disney) sued people who only copied the ideas and not the expressions shows a failure of the legal system, not copyright.

dcow8• 4 years ago

Why, then, is one artist even able to bring a case against another artist for composing a song with a similar riff as one they published? It seems it's not so simple. Perhaps I should have been more specific to say that copyright law as it exists in society today is rather broken so as not to disparage the original intent of copyright as a concept. But, I feel like we're just talking semantics, really. The point remains that the implementation is broken, copyright as we know it.

akira2501• 4 years ago

Copilot itself is a commercial product that is built a body of training data, even though it is completely different from that data. It’s clearly “transformative.”

Is it, though?

The article wrestles with the notion of the gap between "idea" and "expression." To me, I wonder if this is the same gap. The training data is equivalent to the "idea," and the output of using that training data in a particular way is the "expression."

In this view, the result of your training isn't transformative, and it might not even something you can claim copyright over. What is it other than a particular arrangement of facts that have been feed into it? Merely adding weights in a highly dimensional space does not seem "transformative."

This article feels like it's wrestling with the wrong side of the problem.

kmeisthax2• 4 years ago

But how much of a song or a painting can you reproduce?

The reason why fair use is vague is specifically to confuse people who ask these kinds of questions. The Supreme Court needed a tool that artists could use to legally smack down people who republish fragments of other people's work, but didn't want to abolish the 1st Amendment in the process. So basically judges have the final say as to whether or not something is novel creativity or in debt to the original. Any hard-and-fast rule beyond "binding precedent applies" is effectively copyright abolition by degrees.

We lost most of Elizabethan theater because there was no copyright. [..] Without some kind of protection, authors had no interest in publishing at all, let alone publishing accurate texts.

This is a dated example, if only because creative works leave a lot more evidence now than they used to. People today will act to preserve art against the artists own wishes and at great personal risk.

and it’s easy to suspect that the actual payments will be similar to the royalties musicians get from streaming services: microcents per use

Given the amount of data these systems need (read: more than humanity can provide) I'd say microcents is arguably too high. Remember that you can't actually derive a clear chain of value between one particular training set entry and one particular execution of the model. It's all chucked into a blender that runs on almost-linear algebra and calculus. At best you can detect if parts of the image resemble specific training set examples[0] and pay people slightly more if the model regurgitates training set data.

Let's also keep in mind that a good chunk of the licensing system is based on being able to say no to specific users, or write very tailor-made licensing agreements for specific works or conditions. That's still going to be threatened, even if we can pay sub-Spotify-tier royalties every time a model trains itself on your work.

It is easy to imagine an AI system that has been trained on the (many) Open Source and Creative Commons licenses.

Working on it: https://github.com/kmeisthax/PD-Diffusion

The thing is, we already have a good database of reusable, public-domain, no-attribution-necessary images; it's called Wikimedia Commons. I really can't fathom why OpenAI didn't start there, other than just an assumption that they were entitled to larger datasets or a feeling that they could get established before anyone sued.

Even then, OpenAI already tried this with computer code and they're getting sued for it anyway, because they never bothered with attribution in the case of training set regurgitation.

[0] This is possible because part of the prompt guidance process involves a thing called CLIP which can do both image and text classification in the same coordinate system.

Nadya• 4 years ago

Just an FYI but your link 404's. I assume it is a private repo.

kmeisthax2• 4 years ago

Aww dang, I can't believe I forgot to publish it

theGnuMe• 4 years ago

Isn't this just the riff sampling thing? So depending on the output you could be infringing.

Like how Vanilla Ice "stole" Queen's bass line.

https://en.wikipedia.org/wiki/Sampling_(music)

dcow8• 4 years ago

Which is ridiculous. You can't copyright 8 notes of music. This is the festering disease we need to rid ourselves of.

lcnPylGDnU4H9OF2• 4 years ago

Well, now I'm disappointed. I was going to point to this TED talk (https://www.ted.com/talks/damien_riehl_copyrighting_all_the_...) but it appears to have been removed as the video no longer loads for me. However, this (https://www.ted.com/talks/damien_riehl_why_all_melodies_shou...) appears to be an abridged version of the same talk.

The disappointing bit: I say it's an abridged version because I distinctly remember him talking about how he actually claimed copyright for all 8-note melodies under a permissive license, which I can't find in the one that actually loads. He "brute forced" every 8-note melody with a program, saved them to a disk and claimed copyright under (I believe) the MIT license. (I can see legal issues with doing that, of course, so it's not hard to imagine possibly why the video was replaced with a different one.)

an_ko• 4 years ago

Damien Riehl and Noah Rubin have all the melodies on a hard drive. (More info in this Adam Neely video: https://www.youtube.com/watch?v=sfXn_ecH5Rw and the TED talk linked in a nearby comment.) So depending on a court's kinda arbitrary definition of creativity on any given day, all of them may or may not be copyrighted already.

An infringement being accidental also doesn't seem to stop copyright holders from successfully suing people. I think this means you could technically be successfully sued for dropping some bricks on a piano, since no matter where they land, the action constitutes a public performance of a copyrighted work. Fun stuff.

geoelectric3• 4 years ago

Robin Thicke wasn't exactly a sympathy-inspiring celebrity, but the Blurred Lines case was another egregious example.

As someone who grew up in the 80s and 90s, I also really wish we'd gotten the explosion of almost completely sample-based music we were starting to see with bands like the Beastie Boys, the KLF, and Pop Will Eat Itself. The Biz Markie sampling case basically shitcanned an entire nascent subgenre.

kevingadd3• 4 years ago

What's the threshold you can copyright? 9 notes? 16? Why is that the threshold and not 8?

I may not agree that it should be 8, but I don't see any rigorous or well-reasoned model here to explain why "you can copyright 8 notes" is a festering disease when it seems like people are not reasoning through why something should or shouldn't be copyrightable.

avereveard• 4 years ago

You already get protection for characters and stories so it's not like ai will destroy publishing and it's not like existing content is not protected already.

Copyright doesn't protect skill, as it doesn't protect algorithms, because these are tools to create and not finished products.

nonrandomstring3• 4 years ago

I'll grab some popcorn. This could be one of the all time epic battles.

Earlier someone posted, and then deleted, an Ask HN: "Are artists fighting AI Art repeating Metallica versus Napster?"

This is actually an entertaining question, because it brings in the power of the entertainments business.

Where is Napster today? Didn't Metallica win that one? It might not be a great comparison, but what if RIAA, MPAA, Sony and the game industry decide that "generative AI" occupies the same threat space as "piracy"?

It was of course Metallica and an army of ten thousand lawyers and goonies from a vast, wealthy, moribund industry that actually did manage to block the road of progress and frighten the genie back into the bottle.

In fact, the power of the film and music industry to shape technology has been so immense, you have to wonder whether they could do it again over "AI".

Right now I think the entertainments industry is shitting itself over LLM technology, but is split over whether it can gain enough control to allow it on it's own terms, or mobilise to fight it. We haven't yet reached the stage of commodity proliferation. That will be the watershed.

geoelectric3• 4 years ago

The artists didn't exactly win Metallica vs. Napster either.

The end result was a compromise that still funneled some degree of money into the labels, with the artists getting an equally raw deal as before proportionally speaking--but now with a micro share of your $10/mo to Spotify or wherever instead of their share of $10+ for a single album.

I'm not saying the prior business model was sustainable (at least ethically) but at the end of the day, "if you can't beat them, join them" is still one hell of a compromise to make.

nonrandomstring3• 4 years ago

Yes I think you're right. The technology was tamed and brought to heel. It starts out looking "disruptive". What will that look like when Big Media figure out how to take legal control of generative AI and become effective arbiters of all that can be cheaply, mechanically created?

bjelkeman-again• 4 years ago

They took the cake from the pirates and gave it to the privateers. The majority of the musicians get nothing and the cats (labels) and the tech companies get paid.

amelius• 4 years ago

I'll grab some popcorn. This could be one of the all time epic battles.

No, it will be boring, the outcome is clear: those with the deepest pockets will win this battle.

Just look at Disney who had copyright law changed back in the previous century.

nonrandomstring3• 4 years ago

those with the deepest pockets

But you merely beg the question. Which pocket is the deeper? The left pocket of a network of industries that stand to dominate the world via AI? Or the right pocket of the same interests who will be destroyed by it?

jacquesm• 4 years ago

To the artists it didn't really matter whether they were bitten by the cats (the record labels) or by the dogs (the pirates), they ended up losing out anyway. This is because artists are less motivated than either pirates (who have nothing better to do) or record label executives and their cronies (who are protecting their income stream). Artists are a rule are happy to be making enough money to be able to continue to make art. And that makes them the vulnerable party in these transactions. Witness the countless accounts of abuse by record companies of bands and individual artists, including bookkeeping tricks, contractual abuse and so on.

sharperguy3• 4 years ago

I see this going one of two ways:

* States take a laissez faire approach to regulation these networks, we get an open source community of people developing them publicly for anyone to run on their local machines, run small clusters etc.

* States take a heavy handed approach and prevent anyone distributing a model trained on copyrighted works without agreement from the copyright holder, or paying some form of tax that gets redistributed to them. We end up with an oligopoly of ad-supported, data-collecting, heavily censored SAAS models which we cannot download, modify, or run locally.

Guthur• 4 years ago

It should say that the notion of intellectual property is nonsensical.

Imnimo• 4 years ago

I'm not sure I buy this line of reasoning about "inputs" rather than "outputs". If there is some prohibition about using an image as an input, regardless of whether any vestige of that image exists in an output, doesn't that equally prohibit using an image to train neural network that just says whether an image is a cat or a dog? Or how about a network that just tries to denoise a photo from your camera?

XorNot• 4 years ago

Ends up worse then that: 8a707ed2cb8e1aa2e54c1b68d78e76b51d05f0a23c8a8b9d2f7f04aca01b3a96

This is the SHA256 of Captain America - The Winter Soldier. Pulled from my Blu-Ray copy of it. It can only be produced (in the current universe) if you have access to the complete data from the Australia Blu-Ray release of that movie.

So to create that number, every single frame and pixel of the original movie was included in some way.

Is the number itself copyright? After all, every byte of its input definitely was - there's no way to produce it without those bytes.

owisd• 4 years ago

A good rule of thumb is if the derivative work is drawing sales away from the original work. Nobody is buying Captain America - The Winter Soldier to calculate the SHA256 so the movie studio sustains no losses from it being shared, so it shouldn’t count as copyright infringement.

jcranmer2• 4 years ago

Is the number itself copyright?

For a work to be copyrightable (including derivative works), there needs to be a quantum of creativity. Since there's no creativity in the process of constructing a cryptographic hash, there's nothing copyrightable there.

seydor• 4 years ago

why does society and technology have to constantly adapt to ancillary legal requirements, while the laws themselves rarely adapt (e.g. never expire)?

sharperguy3• 4 years ago

Creating a new technology and using it doesn't require the consensus of an entire society. Changing the law does, and so takes forever by comparison.

subroutine• 4 years ago

Does copyright law protect “in the style of”? I don’t think anyone knows.

I think the is the crux of the issue. I think it would not. IMO the spirit of copyright law protect certain essential elements of a work being reproduced identically or nearly-identically (1), but not a composition that is reminiscent of a general style of some other work, and certainly not some collective body of work. To me the source of the training data should not matter insofar that it does not matter if an artist studies a body of work to create pieces in the style of X. Does it concern copyright law that an artificial rather than biological neural network has been trained on a body of work? Would a copyright judgment ever be made on the basis that an artist studied too much of another person's work, and not simply the similarities between the piece they created and another specific work?

It is against copyright law to release either a computer generated lowfi cover or a human generated violin cover of stairway to heaven. On the other hand you could go and buy an amp, guitar, pedals, etc. to perfectly recreate a classic 'jimmy page sound', and whatever you compose would be fine as long as the song does not borrow specific melodies found in led zeplin songs. See any Greta Van Fleet song. Moreover, that GVF reminds everyone of led zeplin is no coincidence. From wikipedia: "Greta Van Fleet is often compared to Led Zeppelin. Jake said he went through a year of really intensely studying what Jimmy Page did to the point I knew how he thought."

At the end of the day I think there is no basis to claim a copyright over a general "style", and that's a good thing.

(1) https://www.briffa.com/blog/classic-copyright-cases-ice-ice-...

_0ffh• 4 years ago

> Does copyright law protect “in the style of”

Well, pastiches are a thing, and I've read and watched a few. Never heard about anyone getting into trouble for them.

pornel• 4 years ago

Existing laws were written before this technology existed.

Copyright isn't a law of nature, but a tool to balance needs of creators and everyone else. With this tech the power has definitely shifted, so the existing law may not be adequate any more.

Animats• 4 years ago

I suspect we'll see "trade dress" lawsuits, because copyright doesn't apply. Probably from Disney.

fergal_reid• 4 years ago

I feel sympathy for the artists, and who knows what jobs AI will leave!

But I don't think the artists I'm seeing on Twitter understand that, even in their best case, where training on a copyright image is considered total infringement, this genie still isn't going back in the bottle.

Models are getting better and training more efficient all the time.

Soon you'll be able to train a model entirely on public domain content and have great generative art output.

It might be harder to express the style you want than just sharing a name of an artist in that style, but that won't hold things back.

With this worldview the discussion about whether training on something is infringing its copyright fast becomes irrelevant, for better or worse.

thot_experiment• 4 years ago

Absolutely true. It won't be long before models trained on purely public domain works will be able to oneshot style transfer from a piece of art fed to it.

We need IP law reform and UBI/strong social safety net, the problem isn't automation, it's that we allow people to own the automation and charge rent on it.

heavyset_go6• 4 years ago

Copilot itself is a commercial product that is built a body of training data, even though it is completely different from that data. It’s clearly “transformative.”

I don't buy this, given that NNs can and do encode data from their training sets in the network itself, similar to compression, and they can regurgitate that code when given the right incantations. They can also encode "the heart of the work" when it comes to algorithms from copyrighted work being recycled in the NNs output.

And I don't buy that it's transformative, either, given that it's taking someone's code as input and just outputting more code. The purpose is literally the same, building software for profit. It's not like the code is being used in an art project, it's being used for the same purposes that the copyrighted works were created. This is not like making a collage out of magazine cut outs, it's like taking "People" magazine's content and rearranging it and selling it as "Persons" magazine for profit.

When it comes to fair use, transformative derivative work is not the only factor considered. Another factor is the purpose and character of use, especially whether or not the derivative work is used for commercial purposes. Code regurgitated by these models are used for the same thing as the code they were trained on: building software commercially.

Another factor is "the effect of the use upon the potential market for or value of the copyrighted work". It's beyond obvious that laundering copyrighted code for commercial use affects the potential market for and value of the copyrighted work.

OctopusLupid• 4 years ago

This seems to be rather US-centric. Here's the EU version: https://copyrightblog.kluweriplaw.com/2021/02/17/tdm-excepti...

Singapore version: https://www.twobirds.com/en/insights/2021/singapore/coming-u...

Other countries have already made copyright exceptions for AI training.

poulpy123• 4 years ago

A good piece that express a balanced point of view on the question, even if I'm not convinced by the conclusion and solution proposed

What does copyright say about generative models?

Related Stories