I had a startup that tried to do so. Our acquisition was torpedoed after a claim by the owners of the training data that sufficient derived data rights were not granted to us -- that our derived data rights allowed us to operate the model but not to {sell the model, sell the company} without granting a [substantial] portion of the sale back to the owners of the original training set.

In retrospect, and my advice to others, would be to try to get derived data rights way early on before there are big sums of money involved that make everyone greedy.

Tell me more about "derived data rights". So you tell me the cost of materials and I factor in labor and come up with a price for my product. I derived my price from the data you supplied. And I can't use this price because it's derived data?

No, I give you a dataset. You make transformations on that dataset and you try to sell it. I say — "hey, I own your underlying data, so I should have ownership of your transformed data, and thus the proceeds of your sale."

Correct. This is the issue we faced.

Derived has so many meanings.

In this case, the transformations are the 0s and 1s (diagnoses labels) affecting a convolutional neural network to fit to these 0s and 1s and a set of network weights that result.

I suppose it would depend on the terms of the sale of the training data (if it was purchased) or how it was acquired (if it wasn't).

We did not purchase the training data, we licensed it. In retrospect it was clear that we had obtained the license to use the data to train a model, but not entirely clear whether we had licensed the rights to commercialize the model (e.g., sell the weights on our fairly generic neural network.)

That said, lots of things start getting confusing, when confusion means owners of the data might make millions off an acquisition.

In this case, it was medical data. The owners were technically the sovereign that licensed the medical data to us. But technically, the medical images had been read by independent doctors who only granted an upstream license to the sovereign to use their reads for medical diagnoses for patient welfare -- not to downstream license the reads to help someone else train a model to theoretically put doctors out of business.

I would liken the AI with a photo camera. I can't paint very well, but I can take this miraculous machine and create a picture on the press of a button. Yet there is still artistic expression in deciding when an where to press the button, and at which angle you hold the machine when pressing the button. And we universally agree that whoever pressed the button owns the copyright of the created picture, even if most of the work is done by the machine.

With AI in general and neutral networks specifically the same applies. There's some nuance to decide if the copyright belongs to the person who trained the network or the person who prompted this specific output. The easy solution is to say "both", and let people handle it with copyright assignments and license agreements.

When we have AGI that thinks for itself we might want to rethink this, but that's in the far future.

Well, there has to be creative input, right? Suppose you take all my written works, train a model, then used that to generate text it doesn't sound meaningful to me that you now own that text. After all, someone else could easily and naturally do the same and they too could end up with similar enough weights that they generate similar text.

That's not a creative act you did. We all do that.

How about if I read and study all your books and then generate a new one perfectly in your style (maybe using one of the common plot template you use often). Would the new book be mine or yours? No AI involved, but the black box looks similar from the outside.

How do they handle this now with writers and painters who imitate other artists?

Perhaps the best example of humans doing this sort of thing is in music. Artists take samples from one another’s works and use them to create new music.

I think the problem the GP highlights is roughly analogous to the case where an artist takes all of their samples from the album of one person and uses it to create a new song that sounds like the original artist made it. That’s a different thing entirely from taking a sample from one artist and a sample from another and then combining them with a new original beat and layering on some other original tracks.

What if someone used no samples at all but used all of the stylistic choices of the artist and made something that sounds like the original artist made it despite no part being directly copied?

"How do they handle this now with writers and painters who imitate other artists?" -> not a derivative work as long as no specific protected elements were copied.

Repainting or rewriting your work in different style would be a derivative work, but painting or writing a different work in your style would not be.

Oh, that's a good one. I suppose you're right in that it would be undetectable that you used an ML model to make the text unless you said so and in that case we'd probably treat it as a creative work.

We all do what? We don't all take your specific work, train a model on it, and then select an output to present as a new work. Those are the creative acts. And certainly, the output may be indistinguishable from something you yourself created, and given the structure and nature of neural nets it might even cause you substantial existential angst to see such work, but I don't see how any of those things would weigh in to a copyright claim. Your best bets for a claim would be either a claim of a 'derivative work' being created or perhaps that the use of your original work does not fall under the Fair Use case. Fair Use would be very weak here because generally you would need to show that the 'heart' of the work (your originals) was duplicated. Similarly a claim of 'derivative work' would have to be novel and would be very difficult when the actions of the AI author are considered to be part of the authoring process while your authorship process is very different. I could see it possible, but unlikely, than an ML system would produce something which, had it been produced otherwise by a human, could be recognized as a derivative work. That would be the criteria, likely the main if not sole criteria, used in a court case.

Suppose someone takes all the words from your comment above and reorder them. Whether this new prose has the same, another, or no meaning at all, combining your words to form something else is a creative act. Sometimes, creativity doesn't take much effort.

Ok, but in the monkey-selfie case, the monkey was not considered to be a tool of the artist (the camera was, but as the artist wasn't operating the tool, that didn't matter).

I think the more important question is about ownership and society. If an AI can produce value, does that value accrue to the owner of the AI?

If yes, then we're in for another round of "let's sack all the workers and replace them with machines". Historically these haven't been fun, and have produced massive inequality, but eventually they produce richer, better-educated societies.

If no, and we say (for instance) that "all value produced by an AI is public property", then we'd head off the beaten track, and possibly to a more equitable society. Or somewhere different, at least.

Hm. In the monkey case, I think a case could be made that a monkey is "a conscious being" and "capable of true thought" and stuff along those lines, and hence something (someone?) who can own rights to their work.

Can the same be said for AI/ML as they exist today? I'm not sure.

The fact remains, AI needs to be developed and trained by someone, and that completely affects the resultant "artwork". That's why people use the "photographer and camera" analogy. Does the Dev not deserve credit for their work?

(A thought experiment that I just came up with, not sure if it makes sense: Say you and I are locked in a room for a prolonged period of time. I tell you a bunch of stuff and you make some art based on it. Following the photographer/camera analogy, can it be said that you are my "tool"? Can I claim rights to what you come up with?)

> think a case could be made that a monkey [...] can own rights to their work

This was actually decided already. PETA sued on behalf of the monkey, and lost. The courts ruled that only humans can own copyrights. However, the issue of whether the photograph is copyrighted by the photographer or in the public domain was not decided. Personally I think that the photographer could likely win the copyright based on the effort he put into the setup, if he got a good enough lawyer.

The Dev, yes. As an independent act of artistic creativity, I don't think anyone has any problem with the dev owning the output of the AI.

To answer your thought experiment: Should an Art Teacher get copyright rights on all their students' work? (what you propose is effectively art college)

Agiain, to my mind the problem comes when someone (let's call him The Baron) employs a hundred Devs to make an AI that provides cheap legal advice, based on public-domain court records. People get cheap legal advice (society wins). The Devs get paid for their professional labour. A million lawyers go out of business. The Baron becomes astonishingly rich.

Should The Baron get all that value? Should the Devs? Should the out-of-work lawyers?

>Should an Art Teacher get copyright rights on all their students' work?

In quite a lot of cases they do (or at minimum the right to reproduce and distribute such at will without compensating the originator of the work).

https://www.trademarkandcopyrightlawblog.com/2012/05/bear-an...

https://designobserver.com/feature/who-owns-student-work/126...

https://www.huffpost.com/entry/who-owns-student-artwork_b_64...

partly this is because education establishments insist on copyright assignment.

My mother got hit with this when she tried to patent the seamless sweater invention she had - suddenly the university stepped in and said it was their invention. Sure enough, she'd signed over all copyright when she started the course.

but just because it happens now doesn't mean it's morally correct. In an ideal world, do we want this?

In the art school sense, it actually impacts one of the most critical questions an art student has of themselves,

"Do I try to make my career defining art now or do I use this time to explore a variety of subjects and mediums?"

The fact the artist doesn't own the copyright to this early work tilts the question to the latter statement.

I think this is the crux of it. We need to decide at what point an AI becomes "a person" capable of owning things.

Unless you recorded the oral stuff or wrote it down, I don’t see how you can claim a right to it.

So, if I setup an HD CCTV camera in a fixed position and livestream sunsets from the beach, I cannot claim copyright of the stream.

I don’t see why this would be different from an AI.

Choosing the location and direction to point the camera is the creative part.

Still not able to copyright that.

I'm still confused by Model Release forms. If I take a photo of someone, I own the rights to the photo. So I can publish it and make money. But they own the rights to their likeness, so they can sue me for money unless I have a signed form. Is that how it works?

If you streamed your sunset shot and I walked into shot, can I sue you for using my likeness without my permission?

And how does this work with DeepFakes? If I DeepFake Sting's head denying climate change, can he sue me for likeness rights?

Model release forms are only necessary to use their likeness for advertising. If it is just art, no form needed. At least that’s my understanding.

There's going to be an argument about "derivative works". If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - That's clearly a derivative work. If I feed all of the New York Times Bestsellers for the past 50 years into an AI and it spits out something, there's much less of an argument that it's derivative of any one work.

I've seen stuff like https://www.thiswaifudoesnotexist.net/ - It's both impressive and not. Sure, it's "new" and "Unique" every time, but if you were to take a trained eye to it, you'd see that a bunch of what it's done are very simple derivatives of an already derivative artform. Do I know which image it's stolen the basic shape from and which it's stolen the color palette? No, but if I spent a few hours looking through the training dataset I'd probably find another image that looked almost exactly like the one I'm looking at. These things are impressive because they can fool a cursory glance, and quite frankly, that's often enough.

"If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - That's clearly a derivative work." does not match the current legal interpretation in the industry.

The general assumption is that machine learning models trained on some data are (usually - details matter) not considered derivative work from that data; so unless you have specific contractual restrictions (e.g. you got the date bacause you have a contract with the data owner saying what you'll do/not do with the models derived from it) then copyright does not prohibit you from using and distributing these models. That's the case even if it's trained on a particular dataset with a single, clear copyright owner.

It probably starts with old precedent on copyrightability of statistics derived from creative work - things like word frequency, most common words, etc are not considered derivative works.

Then we have statistical language models, of the kind that were used in statistical machine translation before the neural approaches overtook everything - and again, the established interpretation there is that statical models trained on a corpus are not considered derivative work from that data, because they essentially are a trivial extension of word sequence frequency counts... but they already are in the category of "feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something", e.g. a simple hidden Markov model or trigram language model from Harry Potter books will easily hallucinate (very lousy quality) Potter-themed text, but the model that can do that is not considered derivative work.

Though this all may vary in different jurisdictions.

"does not match the current legal interpretation in the industry." is the complete proof that the industry is fraudulent, vile, and sociopathic.

> No, but if I spent a few hours looking through the training dataset I'd probably find another image that looked almost exactly like the one I'm looking at.

If you look at nearest-neighbor lookups in papers like BigGAN, you'll see that for all the apparent 'simple derivatives', you will see that it is actually quite difficult to find "another image that looked almost exactly like the one I'm looking at". Similar, certainly, recognizably the same character, absolutely (particularly well-represented ones like Touhou characters), but exactly? And if they were purely copying specific points, the interpolations would not work. (It would also be a lot easier to reverse a given image into the latent vector to edit it.)

Years ago I remember there was an article of someone using brute force to generate every possible phrase of a certain length so that from that point on no one could ever claim ownership of one since they had all already been created.

Also I wonder what defines created. If I create a program that can generate every possible variation of x, do I get copyright when the program is run or do the results have to be saved. Or do I get copyright over the output of my program before it has even run since the result of the program is fixed we can know it would have resulted in the output even if it was never run. What if I just make a webserver with unlimited pages and given the correct url is entered it can output any result at all. That data was available on my website since it was created so do I have copyright over it?

If we allow the output of programs which were not extensions of the users input (photoshop, a text editor) than in theory my program that outputs random numbers now has copyright over every possible work that could ever be created.

That gave me an idea for an app which would permute all phrases of any length randomly and have forward and back pointers to the next and previous word in the series. At the end of an _n_ word phrase, the next link would give you an _n_+1 phrase.

If a crawler came along (like archive.org) it might index all permutations, giving you some evidence of when your site “coined” a phrase.

The app may want to consult a recent USPTO dump you avoid known trademarks and copyrights. Sounds like a fun exercise.

There is something similar to what you are describing.

This reminds of infinite monkey theorem.

[1]https://libraryofbabel.info/ [2]https://en.wikipedia.org/wiki/Infinite_monkey_theorem

Creating an intelligible work through a random process takes a really long time. It took Earth hundreds of millions of years.

A lot of things that seem to fall under copyright are not particularly difficult to brute force. Looking at music for example. Most music follows a set of fairly basic patterns where each part of the music (drum beat, chord progression) is just a short arrangement of only a few choices. And yet the copyright system seems to take even a tiny piece of a song as copyrightable. It would not be difficult or time consuming for a script to generate every possible 4-6 chord progression and every single kind of melody which is close enough for copyright (Often these cases use very general terms like "A descending melody starting at C")

I think the intellectual component to randomly generated content is distinguishing between the good and the bad. And there's a lot of bad.

The problem for me is by AI capability and volume. Stephen King is said to write somewhere in the ballpark of 2,000 words per day with about 100 books written. Scholastic Press publishes about 600 books per year. Now lets say in theory an AI were able to write 2,000 words per second and be able to produce thousands of books per day. This AI could have live access marketing data to tailor write books to specific demographics or even specific individuals. So how could a human author compete? I would think access to copyright would be the only edge a human could have.

Now as for using AI and AI-based writing tools like the GPT2 program that autocompletes sentences for writers is no different then an artist tool like you said. https://transformer.huggingface.co/

Even though it's not related to IP, this may open precedence for making the creators of a given AI system be responsible for any outcomes of said AI (maybe that's the way to go). That said, car companies may prosecuted if their AI driven cars cause accidents (again, maybe that's the way to go).

This discussion is on it's first steps and has a long way to go yet.

I personally am definitely of the opinion that those who develop and train AI should be held responsible for the actions of said AI. That's the only way to ensure that such systems are kept in check. Case in point: the various YouTube "algorithm" fiascos that have taken place over the past several years. If AI systems you've put in place discriminate against certain demographics, you should 100% be held accountable for the same.

Same thing when it comes to improperly trained/developed AI causing accidents or loss of life.

Sounds similar to how parents are responsible for actions of their children up to a certain age. So the common intuition is on your side.

It might also get hairy if/when we make a sentient AI and have to decide if they can have property rights. Which is a thorny issue because even if we assign them rights, they themselves will be the intellectual property of someone else. An intellectual slave?

Probably a good thing general purpose AI is still a few decades away at this point. Society is not ready for the rights implications. Not that I really expect people to have better answers to this in the future.

>Which is a thorny issue because even if we assign them rights, they themselves will be the intellectual property of someone else. An intellectual slave?

Its not that novel. Take a Corporation it may hire an employee, or contractor - or even unpaid intern - and have them develop copy all day long. That copy will be owned by the Corporation in the future, even if the employee/contractor/intern ceases to work for said Corporation (even if said employee/contractor/intern ceases to exist their work will be owned by said Corporation).

It is novel because in the case of IP of this GAI, it would be like owning the DNA of someone. Imagine it being a copyright violation to have babies.

Well if it is like owning DNA, then it can't be patented, the Supreme Court has already taken and ruled on such cases (holding: genes are from nature and can't be patented)...but its not like owning DNA of someone and imagining having babies violating copyright laws won't make that anymore true.

Algorithms are already creating media of all kind (videos, articles, music) and that media is all subject to the same copyright protections as man-made media, replacing algorithm with AI (which seems to be the popular thing to do with SV startups these days, doesn't create Earth shattering facts the law hasn't seen, its just like when the blockchain believers thought DAO code was the law outside the jurisdiction of courts and regulators)

DNA is already established as something owned. I don't think DNA is a good metaphorical proxy in this case, as the DNA of the GAI would sensibly be its set of weights assuming it is a neural net. But if it were, perhaps we would regulate it the way we currently regulate DNA. Which means we ignore the fact that DNA can self-replicate and permit corporations to own DNA patterns and prosecute people who possess or make use of organisms expressing that DNA. While many of the accusations against Monsanto are overblown or inaccurate, Monsanto does indeed own patents on a great deal of genetic material, including entirely naturally produced DNA of organisms in developing nations.

I'm sure we will have this exact discussion when human genetic modifications start to become commercialized...

Of all the things we have to worry about when we have general AI, whether or not they can own ip seems like a small one.

Billions of dollars of lawyer time is wasted annually on this small issue with just humans. It's not as trivial as it seems at first.

It turns out the AI's conquer humanity not through violence, but by highly optimized IP trolling. Hell, we already have a flavor of this on Youtube, where algorithms flag content all day long with little recourse for innocent victims.

I agree. Until the AI itself can complain that someone is getting credit for their work, this isn't a huge issue. Once an AI gets to the point where it feels cheated and wants a cut we've got a very big problem.

famous last words before an ip ownership loophole allows an AI to enslave humanity

So if the AI is the author, then who gets paid? If a copyright term is set in terms of the author's life, when does the AI stop being alive?

I think until an AI expresses a wish for something else, we should assume all AI generated work is in the public domain.

I'd argue that AI isn't the "author". It is simply a tool (like a camera). The person who fed it input and tuned its parameters ("pointed it at an object and pressed the shutter") is the "artist" and owns the right to the resulting artwork.

I'm not sure I agree. I think this is different than, say, a laser printer that just spits out my creative work.

An AI isn't necessarily under control of the person that owns it.

Who do you think owns the copyright of the famous monkey photo? It was the monkey that composed and took the shot. The person who owned the camera had no control over the creative aspects of the shot.

I've commented about this elsewhere in the thread, but basically, it can be argued that a monkey is a conscious being, capable of original thought, whereas the same can't be said about AI/ML.

an AI is nothing but a model. It only creates the illusion of creating because we are incapable of calculating as much and as fast as they can, but ultimately the product is just the result of the model, and the model is a tool made by the human.

Friendly reminder not to treat AI like humans. Ever.

Is there any evidence that our consciousness isn't just the result of computation as well? Is there anything to suggest that we actually have free will and make choices?