Hacker News Reader

> If you are going to use someone else's work in order to make something that you are going to profit off of, I believe that original author should be compensated. And should also be able to decide they don't want their work used in that way.

> Note that I'm not talking about what existing copyright law says; I'm talking about how I believe we should be regulating this new facet of the industry.

Is it really new? Humans have always learnt by studying what's out there already. Our whole culture is built on what's been done and published before (and how could it be otherwise?). Without Bach there would be no Mozart, and then down the line their influence permeates everything you hear today.

If anything I'd like to make it easier to reuse parts of our shared culture, and limit the ability of organisations to control how things that they've published are reused. You can make private works if you want to keep control of them, but at some point the public deserves to share and rework the things that have been pushed into the public consciousness.

> Humans have always learnt by studying what's out there already. Our whole culture is built on what's been done and published before

Are you implying that educators should not be compensated or credited? Because that is not how it works in the real world.

If I read a lot of fantasy books as a kid, then start writing my own fantasy book, should I have to pay royalties to the authors of the books I read?

Does your ability to write fantasy books absolutely depend on having read those fantasy books as a kid? Was gaining the ability to write your own fantasy books and profit from them your only motivation to read those fantasy books? After gaining the ability to write fantasy books thanks to having read them, can you now produce fantasy books at a qualitatively different speed, scale, and conditions than any of the authors of the books that you read?

If the answer to those three questions is "yes", then I would argue that yes, you absolutely should have to pay royalties to the authors.

> If the answer to those three questions is "yes", then I would argue that yes, you absolutely should have to pay royalties to the authors.

Copyright lobbies can't have their own cake and eat it too, if you want to enforce such a different way of thinking about copyright compared to the current one, that would destroy the current industry and rightly so.

I would appreciate if you addressed the questions as literally stated.

I could just as well argue that businesses developing or making use of LLMs can't have their cake and eat it too. If they want their computer programs to enjoy the same rights and prerogatives as human creators do, they should be ready to demonstrate that those models are truly moral agents, with their own lived experiences, and thus deserve the status of legal persons as of themselves subject to the same laws and obligations as human beings.

> If they want their computer programs to enjoy the same rights and prerogatives as human creators do

I don't think they care about that, they seems fine with output images not being copyrightable as per the current legislation.

They already paid when they bought the books. Why would they need to pay more?

Say I want to write a screenplay and produce the resulting film, for profit, but I am literally unable to have any ideas whatsoever unless I base them on book that I read. With this in mind, and with this sole motivation, I buy and read the whole collection of Brandon Sanderson's novels and create a screenplay based exclusively on their content, for I have no ideas nor experiences of my own. I already paid when I bought the books. Why would I need to pay Brandon Sanderson any more?

Anything you create comes from what you've seen, from what you've experienced.

That's a dead-end to start having to pay each and every one of the original sources of the components of your mind each time they are used to create!

So I understand you are arguing that derived works should not be subject to royalties, i.e. I should be able to produce a film based entirely on the work of a living author without restriction or the need to pay any royalties.

No, you're over-generalizing what I'm saying (a.k.a. straw man fallacy).

If you're commercializing e.g. goodies that use the exact same character designed by some artist, so that anybody can tell you it's the same, AND if the traits of this character are actually original (not just so that anybody would come up with it independently), AND if the people buying your goodies are all thinking about the original character in the first place, AND if the original author is alive, then I'd absolutely agree that royalties need to be paid.

That's just an example. To tell you that in specific cases it's obvious that royalties are required.

But in the general case, no. Because, else, anything just is a derived work. Just think about it.

It's even impossible to retrace all of the woven threads, ramified tendrils, that link your mind to the billions of other minds all over the world and over the ages.

The world of ideas is liquid. All is mixing, all dissolves and disappears in everything else, and is reborn new and different, again and again.

How much does the world owe to Gilgamesh and Homer?

I don't know. Whether the answer were to be either 'our entire human culture', or 'absolutely nothing whatsoever', or any point in between, how would that be relevant for the discussion at hand?

Yes, you do. The royalty is usually called a "college degree".

Most jobs require a degree or certification of some sort.

> Because that is not how it works in the real world.

Unless you're satirizing it or making fun of it in some other way. Then it's fair use..

Educators get paid a flat fee, not an indenture on your future work (nor do you owe anything back to the people who made the learning materials they use). And if you teach yourself from books or websites you don't pay anything (except maybe the cost of the books). All of which seems right and proper?

>> Is it really new? Humans have always learnt by studying what's out there already.

"Humans" being the important word here. I don't understand why people keep trying to compare training a model to humans learning through reading etc. They are very different things. Learning done by machines at enormous scale and done to benefit private companies financially is not the same as humans learning.

> They are very different things

Yes, but also very similar. We learn very well by spaced repetition, and by practicing things. Our whole nervous system stores information in a similar way. Is the brain and the individual neurons more complex? Yes, sure, but that doesn't negate the core similarities.

> Learning done by machines at enormous scale and done to benefit private companies financially is not the same as humans learning.

Yes, that's the important difference. That in the end if you train a robot you get a program that's easy to copy/scale, the marginal cost of using it is orders of magnitude lower than what you'd get if you would do it with humans in the loop.

We already have fair use in copyright, because there are important differences between the various forms and modalities of human imitation.

And of course maybe it's time to rename copyright to usageright. After current copyright doesn't even apply in most cases. (The results are not derivatives, there's sufficient substantial transformative difference, etc. That said, the phrasing in the US constitution still makes sense: "... the exclusive Right to their respective Writings and ..." ... if we interpret Right to mean all rights, including even the right of who can read/see it.)

How is it meaningfully different with respect to this question?

If I go to a museum and look at a bunch of modern paintings, then go home and paint something new but “in the style of”, this is well-established as within my rights, regardless of how any of the painters whose work I studied and was inspired by might feel.

If I take a notebook and write down some notes about the themes and stylistic attributes of what I see, then go home and paint something in the same style, that too is fine - right? Or would you argue the notes I took are a copyright violation? Or the works I made using those notes?

Now let’s say I automate the process of recording those notes. Does that change the fundamentals of what is happening, with respect to copyright?

Personally, I don’t think so.

The law most definitely distinguishes between the rights of a human and the rights of a software program running on a computer.

AI does not read, look at or listen to anything. It runs algorithms on binary data. An AI developer who uses millions of files to program their AI system also does not read, look at or listen to all of that stuff. They copy it. That is the part explicitly covered by international copyright law. It is not possible to use some file to "train" a ML model except by copying that file. That's just a fact. It wasn't the computer that went out and read or looked at the work. It was a human who took a binary copy of it, ran some algorithms on it without even looking at it, and published/sold/gave access to the software.

AI software is a work by an author; not an author.

> How is it meaningfully different with respect to this question?

Humans can't be owned by corporations for one.

The differences don't seem salient though. Doing a legal thing faster doesn't generally make it any less legal; doing it for profit changes the legal regime somewhat but not in ways that seem relevant to what's being claimed.

> Doing a legal thing faster doesn't generally make it any less legal

“But officer, it is legal to drive just slightly slower than I was going!”

Simply put: You are wrong. The law makes arbitrary distinctions all the time, for practical reasons.

There is a specific law against driving above a certain speed. There's no law like that against being an AI.

> You can make private works if you want to keep control of them, but at some point the public deserves to share and rework the things that have been pushed into the public consciousness.

There's already a licensing framework for artists doing this - should they wish to. It's called Creative Commons, and allows a pretty fine distinction of rights from public domain to free for personal use not commercial, and everything in between. https://creativecommons.org/

I agree our shared culture in some sense ultimately owns the productions of the culture. But that's wildly different to (and in some senses the opposite of) letting private companies enclose, privatise and sell those products back to us. As for example Disney has done over and over again, taking myths transcribed by the Brothers Grimm, or classic novels now in the public domain, 'reinterpreting' them and viciously enforcing copyright on these new interpretations.

The entire point of copyright law is to allow the Bach's of this world to profit from their work - without having to die in poverty and obscurity, as so many artists and musicians have historically, even while others have profited from their work at scale.

> Humans have always learnt by studying what's out there already.

Finally - as other commentators have noted, there's really no similarity at all between a human at human pace studying and integrating understanding of a piece or genre of art, and an AI training to replicate that work at scale as perfectly as possible. A much better comparator would be the Chinese factory 'villages' that reproduce paintings at scale for the commercial market, without creativity or 'art' being a part of the process. But even that is a poor analogy, since individual humans mediate the process. A really good analogy would be a giant food corporation like Nestle somehow scanning the product of a restaurant and then offering that chefs unique dishes that had taken years to invent for nearly free - using the same name and benefiting from the association.

What if the AI is open source and run by individual creators? That changes the slant of the argument a lot. I worry that excessive regulation will mostly come down on individual creators using AI tools.

I'm not sure what you mean? The image and text models currently becoming popular are being trained on work that is not owned or created by those creating the models. The consensus amongst the artists whose work is being used to train them (without their permission or compensation) is very high that this is a bad thing. Financial impacts are already being felt by artists across industries. There's a huge level of denial of the impact of this on professional artists, already today, here on hacker news. 'Open source' is a separate issue to training on other peoples work, replicating their style and stealing their livelihood.

"Humans have always learnt by studying what's out there already."

Neither an AI model nor an AI developer who programs that model are actually studying "what's out there already". One is copying files, and the other is running algorithms on copied files. And then the first one is raking in $$$ while bankrupting the authors of those files. That's illegal in the US, UK, EU and under the terms of the Berne Convention.

> Lossy or not, the training data provides value.

If we ignore the issue of machine learning for now; It's not the job of copyright to prevent people extracting value from a copyrighted work.

If it was, then it would be possible for copyright holders to launch lawsuits that block entities from using the knowledge that was published in copyrighted reference material. Or the rights holder of a cookbook would be able to block people from making the recipe.

We have other sections of law that provide protection along these lines. Patents give their holders a monopoly to extract value from the given invention, a much stronger protection that copyright law. But in exchange Patents are limited to 21 years, much be publicly documented and only certain types of things are covered.

Trade secret laws can be used to protect recipes, formulas and other processes, but only as long as the holder makes reasonable efforts to keep it secret. The owner can't have it both ways, have the IP publicly known and protected by trade secret law.

The only purpose of copyright is to give the holder a monopoly over the reproduction of a work. The definition of reproduction might be quite wide these days: A performance of a work is a reproduction, a cover of a song is a reproduction, distribution is reproduction in the modern age of computers... etc, but the limit of copyright law is reproduction.

Coming back to machine learning, there is a somewhat open question if training counts as a form of reproduction (well, outside of Japan). But we can't use your proposed "extraction of value" metric as a way to decide that.

Personally, I would argue that training a machine learning model is roughly equivalent to a human brain consuming copyrighted works, and should be treated the same in law.

The fact that a machine learning model is (sometimes) capable of recreating a copyrighted work later shouldn't be held against them, as a human brain is also fully capable of recreating copyrighted works from memory. From a legal perspective, recreating a copyrighted work from memory will not save a human from a copyright infringement lawsuit, it's simply not a defence. The copyright infringement happens when the work is recreated.

The purpose of copyright is to create moral and economical incentives to authors to create new copyrightable works, but giving those authors a time limited state granted monopoly. Reproduction is only one aspect to this. As an example, applying a song to a video require addition permissions even if the party has permissions for reproduction. The owners to a record can also disallow the use of a song in a political event based on the context of morality, even if the politician has bought permissions to play the song in public.

I would personally also argue that machine learning model is roughly equivalent to a compression algorithm. Converting a 4k video to a 420p video is just as lossy as feeding a learning model from a 4k video and ask it to reproduce the 420p video. It has nothing in common with how a human brain is consuming content or learns information. No person can produce a 420p video just by consuming a 4k video, nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning.

Your examples are not inherent rights that copyright law explictly grants to rights holders. They are clever side effect of how the holder licenses out their monopoly on reproduction.

Holders rarely grant unrestricted reproduction rights to anyone. Reproduction rights licenses always come with a bunch of explicit restrictions, for example: "You may reproduce this novel, in print, unmodified, only for retail sale, in North America, on this quality of paper, for the next 5 years" and so on.

The party has the license to reproduce the song as a standalone audio recording, but attaching it to a video and reproducing the combined work isn't covered and the party must enter into negotiations with the rights holder for a new license. Such licenses often only grant the rights to reproduce it with that exact video and not a different one later, which allows the rights holder to gain control over which videos their song is attached to.

Same thing with holding morality over political events. The rights holder was careful to add a bunch of restrictions to that public performance license they sell. Sure, the politician might have bought a licence, but they forgot to check the small print that blocks their type of event from actually using it.

-----

Machine learning is kind of like compression, yes... It can be a useful analogy at times.

But it is absolutely nothing like lossy video compression. It's not compressing a single file or object. The only way you could train it on a 4k video and get a 420p video out is if that model was extremely over-fitted. The resulting model would likely be bigger than a 420p h264 video file and useless for anything else.

The way that machine learning is like a compression is that it find common patterns across it's entire training set and merges them in very lossy ways.

And it's actually very much like how a human brain works. Your brain doesn't start from scratch for every single human face you recognise. Instead, your brain has built up a generic understanding of the average human face, grouping by clusters of features. Then to remember a given human face it just remembers which cluster of features it's close to and then how it differs... Which is a form of lossy compression.

> No person can produce a 420p video just by consuming a 4k video

But many people do remember entire songs, complete with lyrics and music. And people with musical skills can (and often do) reproduce that song from memory as a cover... Which is copyright infringement if preformed publicly or otherwise distributed.

> nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning.

There are many things which large LLMs like chatgpt are absolutely incapable of doing. People do over hype their capabilities.

But in my experiments, chatgpt is actually quite good at tasks that require interpreting emotions and social contexts. Does it actually understand these emotions and social contexts? shrug. But if it doesn't actually understand that just proves that true understanding isn't actually needed to preform useful tasks in those areas.

The thing with sync rights is that they are a construct made by US courts, who made the interpretation that syncing music to a video invokes the derivative works part of US copyright law. As such a party need to obtain both a license to reproduce and a license to create a derivative work. US copyright law are split into six categories, one which is reproduction. The other 5 are: preparation of derivative works, distribution (which is not the same as reproduction), public performance, public display, and public performance by means of a digital audio transmission. Copyright owner is given in the law the right to convey each of these exclusive rights separately.

Other countries like say France has moral copyright law and property copyright law as cleanly separate laws. Different rules but with similar practical implications.

In both cases they are not just fine prints in a license. A license that intend to give recipients full permissions need to include everything. This is why international written copyright licenses are a huge legal problem with no obvious solutions.

----

The human brain is vastly more complex. We don't learn by building up an understanding of the average human face, grouping by clusters of features. At best it would be a extreme oversimplification that ignore 99.9% of what the brain does. The eye neurons (still an oversimplification) sends the visual inputs to multiple parts of the brain, each interpreting and signaling each other, and the outcome of that both influences and changes growth and behavior of those paths and parts. You see a face and it talks to the Amygdala, but also to the frontal cortex, but also the limbic and pre-limbic cortex. It runs a simultaneously a simulation in the hypothalamus to test emotional reaction. We are still not at the hippocampus where long term memories is generally considered to form.

A person would be long dead if they had to go through the generic understanding of what a tiger look like, then go through memories that distinguish a lion with a cat, and last go through memories of the significance of a fenced tiger in a zoo in contrast to one in the jungle.

Trying to remember a given human face actually invokes a lot of those parts of the brains that activated when we saw them, and a face can also be remember by putting oneself into the emotional state that we were when we met them. One theory about dreaming is that we also rerun those neural pathways, but not all of them at the same time, which then allows for more context.

LLMs doesn't even come close to operate like this. It can be good at emulating what seems like learning, but comparing them is like saying that the complex system we call the immune system is like a gun. Both can kill people.

Copyright law (EU, US, UK, international under the Berne Convention) covers reproduction, distribution and exhibition. That's exactly what those FBI warnings on VHS movies used to say. Distribution and exhibition are prohibited along with reproduction.

In all the cases I mentioned, the only legal way to make any exception to that is if the copying does not harm the interests of the author or reduce the market value of the work. These are the actual laws.

"training a machine learning model is roughly equivalent to a human brain consuming copyrighted works"

A few clear differences: 1. The person "training" a machine learning model doesn't even need to view the work. They copy a file. They do not study or learn from it. 2. A human brain doesn't rely on a 100% verbatim digital copy of the work. To the extent that a brain "makes a copy" of what it observes, it is impossible for it not to. 3. Copyright law (almost everywhere) explicitly applies to making digital copies of binary files of a work (without which it is not possible to "train" a model using the work). Nowhere does it ever apply to a human brain when a person looks at the work.

Not all the things you mentioned are considered "reproduction". A cover of a song is a derivative work, and requires compensation. Showing a movie is exhibition, and is explicitly addressed in copyright laws. These things are not just considered "some form of reproduction".

The laws actually exist and are easy to find and read.

> the training data provides value

I wish it was easier to build on someone else's copyrighted value. Geforce Now shouldn't have to get permission from game makers to rent out server time for users to play games they already own. Aereo shouldn't have been legally obliterated for the way they rented out DVRs with antennas. Using a sample in a song shouldn't be automatic infringement.

> distributing that model is, in fact, copyright infringement

is it? If i distributed digits of pi (to the umpteenth billion decimals), it theoretically contains copyright information in their digits.

The distribution of the copyrighted material is the infringement, but not if the data is _meant_ to produce other effects, and it is reasonable that the data is used for some other purpose _other than_ to replicate the copyrighted works.

> the training data provides value.

and so does a textbook. A student reading the book (regardless of how that book was obtained - paid or not) does not pay royalties from the knowledge obtained.

Value derived from a work was never a component of protected IP.

If you write a song which is so inspirational that it influences the way a listener thinks or even inspires them to make similar sounding songs, you don't have any claim to that value.

> the training data provides value.

Copyright law doesn't protect "any provided value". Fair use specifically allows content creators to use copyright.

This can be for the purposes of parody, it can also be for the purposes of "reaction videos" or "commentary". The original content creators are NOT compensated for the value they helped create in a "reaction video".

https://arstechnica.com/tech-policy/2017/08/youtuber-court-b...

> If you are going to use someone else's work in order to make something that you are going to profit off of, I believe that original author should be compensated. And should also be able to decide they don't want their work used in that way.

You're posting on HN. Are you expecting a check from YCombinator?

> Regardless, copyright is about distribution. If the a model trained on copyrighted material is considered a copy or derived work of the original work, then distributing that model is, in fact, copyright infringement (absent a successful fair use defense). I'm not saying that's the case, or how a court would look at it, but that's something to consider.

The government is what ultimately decides what copyright means. And if the court wouldn't look at it that way, then it's not "in fact copyright infringement".

> Lossy or not, the training data provides value. If all the various someones had not spent time making all the stuff that ends up as training data, then the model it trains would not exist.

I don’t think deriving value is the measure for infringement. The artist derived value from society and from countless inspirations, should they be compensated?

Copyright is the right to distribute your content, not to control it in every way and derive maximum value possible. The purpose of copyright is to incentivize creators and they already get, in the US, 70+ years of monopoly preventing anyone else from selling copies.

This seems like more than enough incentive for creators to create, as evidenced by massive copyrighted material created at a rate far above population growth (more content is created and more money is made via copyright than ever before).

> If you are going to use someone else's work in order to make something that you are going to profit off of, I believe that original author should be compensated

This is insane. I learned mathematics from Houghton-Mifflin and Addison-Wesley copyrighted textbooks. Are you telling me that if I use this knowledge to, say, calculate matrices, I am violating copyright?

If I read Brandon Sanderson, absorb his writing style, setting, characters, and plot devices into my subconscious, and then a year later pen a short story that happens to be Sanderson-ish, do I have to write a check to Brandon?

No. LLMs are doing roughly the same thing our own brains do, at least at a high level: absorb information, and utilize that information in different ways. That's it.

Society and the economy will do just fine without copyright law. Some things would probably go away, like blockbuster movies. But a lot of other things would flower.