As with all these generative AI tools, it should be legal to train models on public-facing data, but the raw results should be public domain. Maximize the creative potential for the public,...
As with all these generative AI tools, it should be legal to train models on public-facing data, but the raw results should be public domain. Maximize the creative potential for the public, minimize the economic downsides. At best these companies should be able to sell access (both to their proprietary model and to the processing power needed to run it), but large companies should have no incentive to replace artists with them.
The output of generative AI has already been determined to be un-copyright-able in US court, which means they are indeed in the public domain at least in terms of US law. This is because copyright...
The output of generative AI has already been determined to be un-copyright-able in US court, which means they are indeed in the public domain at least in terms of US law. This is because copyright requires human authorship. Since the copyright of a creative work belongs to the creator, and non-humans cannot hold copyright, works of a non-human creator are thus in the public domain. The famous monkey selfie is also in the public domain for the same reason.
I love the monkey selfie example, but an important detail there is that the courts ruled the author to be whoever pressed the button on the camera, which is how the monkey was in contention for...
I love the monkey selfie example, but an important detail there is that the courts ruled the author to be whoever pressed the button on the camera, which is how the monkey was in contention for the copyright in the first place. Similarly, the plaintiff in the case from the first link explicitly chose to list the algorithm as the author of the works, not himself, which leads to the same outcome as the monkey case.
In practice, neural nets are a tool just as much as a word processor, or a 3D modelling package, or camera is: the human using the tool remains the author for copyright purposes.
That’s not to say there aren’t a million problems with copyright law itself, with its interactions with this technology, and with the ways that large companies are using the works of others. Those questions absolutely need to be figured out, but it’s going to be a lot more complex than just saying model output can’t be copyrighted.
I'm not under the impression that there's any case law to back this up as an actual legal claim. If there is definitely link it, as I'm happy to be corrected, but I don't think this is something...
In practice, neural nets are a tool just as much as a word processor, or a 3D modelling package, or camera is: the human using the tool remains the author for copyright purposes.
I'm not under the impression that there's any case law to back this up as an actual legal claim. If there is definitely link it, as I'm happy to be corrected, but I don't think this is something that you can claim without pointing to actual decisions by the Copyright Office.
Copyright law is indeed a complex beast and has numerous problems even without AI copyright issues in the mix... especially in the music industry, which makes this example stand out.
It's an interesting question! A lot of the answers are't definitive yet - the field as a whole is new enough that things are still being litigated and figured out, not to mention the potential...
It's an interesting question! A lot of the answers are't definitive yet - the field as a whole is new enough that things are still being litigated and figured out, not to mention the potential divergence between US/EU/China - but there are nods in that direction even in the ruling that denied copyright in the Thaler case:
Undoubtedly, we are approaching new frontiers in copyright as artists put AI in their toolbox to be used in the generation of new visual and other artistic works. The increased attenuation of human creativity from the actual generation of the final work will prompt challenging questions regarding how much human input is necessary to qualify the user of an AI system as an “author” of a generated work, the scope of the protection obtained over the resultant image, how to assess the originality of AI-generated works where the systems may have been trained on unknown pre-existing works, how copyright might best be used to incentivize creative works involving AI, and more.
[...]
This case, however, is not nearly so complex. [...] Judicial review of a final agency action under the APA is limited to the administrative record [...]. Here, plaintiff informed the Register that the work was “[c]reated autonomously by machine,” and that his claim to the copyright was only based on the fact of his “[o]wnership of the machine.” Application at 2.
This policy does not mean that technological tools cannot be part of the creative process. Authors have long used such tools to create their works or to recast, transform, or adapt their expressive authorship. For example, a visual artist who uses Adobe Photoshop to edit an image remains the author of the modified image, and a musical artist may use effects such as guitar pedals when creating a sound recording. In each case, what matters is the extent to which the human had creative control over the work's expression and “actually formed” the traditional elements of authorship.
The official line right now is pretty solidly that AI is a tool, but the open question is where exactly the bar lies for creative control, and that'll affect the real-world outcomes fairly significantly.
The input needed to make something copyrightable has a lot of non-AI precedent, and I'd argue that the bar there is already absurdly low (certainly lower than two minutes it might take to compose a decent prompt and cherry-pick a good output), so I'd actually be fine with that requirement becoming stricter as part of broader reforms.
I probably will end up looking over it in the near future - it's highly relevant to what I do day-to-day - but right now I don't have time to start diving into a 70 page document, so I can't give you a good summary of what's in there!
AI to the rescue? ;) Summary (no statement made on accuracy or quality): There is a need for new federal legislation to address the growing challenges posed by AI-generated digital replicas, which...
Summary (no statement made on accuracy or quality):
This report by the U.S. Copyright Office examines the legal and policy issues related to artificial intelligence (AI) and copyright, with a focus on the challenges posed by the rise of sophisticated generative AI models that can produce digital replicas of individuals' voices, images, and likenesses. The report analyzes existing laws, including rights of publicity and privacy, as well as proposed federal legislation like the NO FAKES Act, to assess the need for new federal protections against the unauthorized use of digital replicas. Key findings include the recommendation for a federal law that covers both commercial and non-commercial uses, provides a notice-and-takedown framework with safe harbors for online service providers, and balances First Amendment concerns through carefully crafted exceptions. The report also suggests that any federal law should not preempt stronger state-level protections, and should clarify that it does not conflict with copyright law's limitations on the protection of style or ideas.
The key implications from the report are:
There is a need for new federal legislation to address the growing challenges posed by AI-generated digital replicas, which can be used to cause significant harm to individuals.
Any new law should strike a careful balance between protecting individuals' rights and preserving free speech and expression.
Coordination between federal and state laws, as well as collaboration with online platforms, will be critical to effectively addressing the misuse of digital replicas.
The report's recommendations provide a framework for policymakers to consider in developing comprehensive legislation to address this emerging technological challenge.
This case gets trot out often for finding that AI art is non-copyrightable, but that's not actually what the case held. There are very, very important caveats which get skipped over that...
This case gets trot out often for finding that AI art is non-copyrightable, but that's not actually what the case held. There are very, very important caveats which get skipped over that effectively make this case irrelevant with respect to whether or not you can copyright an AI generated image.
The most important fact is that the plaintiff never claimed to be the artist of the work. His position going into the case was that the AI engine (that he created) was the artist and that he owned the copyright by means of work for hire. The defense (the copyright office) did not challenge his assertion of authorship. As such, the only legal question under consideration was whether an AI could be considered the artist of a work for the purpose of copyright. As you mentioned, this was already a settled legal question, thus his lawsuit was denied.
Importantly, upon appeal he changed his position and claimed that he was the original artist and that the AI was only a tool, however appeals courts do not consider questions of fact, only law, and his attempt to change his position was denied and the original verdict upheld.
Ergo, the question as to whether you can copyright an AI image if a human claims to be the author is still legally unresolved.
It’s not black and white. Most AI works are a collaborative effort between humans and AI. Is using “content aware fill” in Photoshop going to invalidate your copyright? No way. So if you mash...
It’s not black and white. Most AI works are a collaborative effort between humans and AI. Is using “content aware fill” in Photoshop going to invalidate your copyright? No way. So if you mash together a few song sections, each from an AI, that could have a copyright on it as well.
The case over the graphic novel specifically said he could still copyright the arrangement (since even though the images were from Midjourney, a human chose how to arrange them in combination with...
The case over the graphic novel specifically said he could still copyright the arrangement (since even though the images were from Midjourney, a human chose how to arrange them in combination with the writing), so yeah. The person I replied to referred to the "raw output" though, so I assume they were thinking of the AI generated work itself, not any later works derived from them.
Yeah, a lot of people don't realize that in a future where AI can replace artists, the big conglomerates already have the rights to enough high quality photographs (stock image companies), text...
Yeah, a lot of people don't realize that in a future where AI can replace artists, the big conglomerates already have the rights to enough high quality photographs (stock image companies), text (publishers), music (music labels), drawings (Disney and friends), etc saved up in their vaults to create competent models. The other day I read a paper of how a team recreated Google's BERT using one GPU trained for a week by using a higher quality text dataset, as it turns out you don't need web scale data but rather good quality data. This means that if AI is able to replace people then only big companies will be able to afford doing so and the only question left is how the split is gonna go between the people that execute the tech and the people that owns the rights.
Why? What’s the difference between me distilling the things I have seen into something new and me using a machine that does the same thing? Put another way, why shouldn’t my work product also be...
Why? What’s the difference between me distilling the things I have seen into something new and me using a machine that does the same thing? Put another way, why shouldn’t my work product also be public domain then?
I mean, philosophically I think all things should be public domain. Copyright is an aberration that is actively harmful to artistic expression, trading the freedom of everyone for the profit of a...
I mean, philosophically I think all things should be public domain. Copyright is an aberration that is actively harmful to artistic expression, trading the freedom of everyone for the profit of a select handful of commercial artists and the corporate fiefs that employ them.
This is especially true of music, which thrives on remixing existing ideas, setting new lyrics to existing music, setting existing words to new music, sampling existing things and building new things with them, etc.. That's how culture is formed: ideas become ubiquitous and are used by many people freely, not treated as property.
I also generally believe that any output from the current generation of generative tools is isn't art, as art is a process of human expression, not a productized end result. It lacks sufficient human involvement and control in the process. (e.g. science is a process of ascertaining fact, not an end result of that process.)
I would also be very offended if someone sent me an email at work that was written by an LLM. It's a statement that someone isn't worth the effort of communicating with properly (worse: they want to waste your time), or that they don't understand what they want to say enough to articulate it themselves...in which case a non-psychic word vomit generator isn't going to know any better than they do and is just useless noise.
It may have become that, but the original spirit of copyright was in fact to focus on protecting the smaller time creators. When you make a convuluted process to obtain copyright, the corporations...
Copyright is an aberration that is actively harmful to artistic expression, trading the freedom of everyone for the profit of a select handful of commercial artists and the corporate fiefs that employ them.
It may have become that, but the original spirit of copyright was in fact to focus on protecting the smaller time creators. When you make a convuluted process to obtain copyright, the corporations win. Because most people won't know about it (and then corporations copyright their work), some of those thst do get bogged down in beauracray, and some of those may still have edge cases to argue. For a corporation, this would simply be someone's full time job.
But yes, corporations find loopholes and lobby behind closed doors. So it's been perverted to something that basically hoards IP's well past the creator's time to benefit. I see that more as a means to return to older time limits rather than abolish the whole thing.
I also generally believe that any output from the current generation of generative tools is isn't art, as art is a process of human expression, not a productized end result
I philosophically agree, but I'm worried in reality because where's the line? Once we establish that, people will simply skirt that line and guess who benefits from that most? Corporations will generate the art and hire some cheap "artist" who simply edits It just enough in some assembly line to be copyright. Small time creators would be legislated out of using it for any livelihoods.
On a technical levels: scale. You can only produce so many drawings in a day. A computer can generate thousands, and is more scalable and prectiable than a human. It's the same difference between...
On a technical levels: scale. You can only produce so many drawings in a day. A computer can generate thousands, and is more scalable and prectiable than a human. It's the same difference between you choosing to save a web page and a web crawler devouring terabytes of data that starts to significantly burden the server owner. Most people don't care about saving or copying a few pages here and there.
It's also an obvious logistical issue. You share a favorite web page and it's a symbiotic relationship. You get the communal value of sharing something you like, the website gets more traffic to show interest or deliver ads to. You can't advertise to bots (and I'm sure there's enough trackers to identify a human for a bot with some degree of accuracy). You run into the tracker problem once again when all your traffic is bots. Lots of server load with no human curiosity to engage woth with nor advertise to.
Put another way, why shouldn’t my work product also be public domain then?
The objective reason is simply because the government determed it not being public domain by default. Reasons being due to historical events and the fact that these laws were designed originally from to protect the smaller, less law saavy inventors from being poached by corporate.
But you are more than free to release it if you feel differently
Another difference is that humans and LLMs have large differences in how they function. LLMs are not capable of producing things they have not been trained on; they can only output giga-mashups...
Another difference is that humans and LLMs have large differences in how they function.
LLMs are not capable of producing things they have not been trained on; they can only output giga-mashups comprised of tiny chopped-up bits of many hundreds or thousands of pieces of training material. Even the most novel looking LLM output is pieces of other existing work mashed up and reshaped.
Humans on the other hand, with exception to those with photographic memories, don’t typically make direct use of anything they experience. Instead, we filter input through numerous mind-lenses (which are unique to each individual) which is then boiled down to sets of highly abstracted concepts and principles. Furthermore, we can do this on very little input; we only need to see an object once to grasp its most essential traits. This means that even attempted copies, on some level, are some percentage original; we’re fighting our own nature to produce 1:1 copies and while our creations may have strong influences (blatant plagiarism aside), they’re original in a way that LLM output as it stands simply cannot be.
Now I have no doubt that eventually machine learning will progress to the point that it closely resembles human minds and gains similar capabilities. When that happens, I think there will be much more of an argument that their creations are original work, but we’ll be running into ethical dilemmas too because chances are that these constructs will have something approaching sentience, which makes using them as tools morally questionable.
I think separating the concepts of “novelty” and “creativity” is probably a good idea here. Creativity I’m pretty much with you on, it’s a much more complex philosophical concept that would...
I think separating the concepts of “novelty” and “creativity” is probably a good idea here. Creativity I’m pretty much with you on, it’s a much more complex philosophical concept that would necessitate intent and quite plausibly self-awareness under most definitions. But novelty - stuff that isn’t just mashed up pieces of other works - is a different beast.
The chopped-up fragments that language models work in are tokens, and at a good-enough approximation for now those are just individual words. Everything else is probability: what’s the likelihood of token X existing in this proximity to token Y.
It’s easy enough to show that statistical models can be novel in that context with a toy example: constrain the parameter count enough and all you’ve got room for is a basic grammar. Tokens with this mathematical property go here, tokens with that property go there, and suddenly the model has derived subject-verb-object word order in English. Now it can generate practically unlimited plausible, simple sentences with no need to mash up existing work.
It obviously gets vastly more complex with larger models. Those do have room to store entire token sequences from training work, or parameters that might recreate those sequences with high enough probability that they’ll come up eventually in normal use. It demonstrably happens, given they can sometimes quote existing work on demand. But it’s not all they do; that distillation of “essential traits” (a phrase that I really like in this context, btw!) is exactly how ML models compress hundreds of terabytes of training input down to a few gigabytes of parameters - and in turn can produce hundreds more terabytes of new, different content from those few GB of probabilities.
So yeah, I’m with you on the big questions. LLMs do function differently to humans, and creativity is a deep concept. But reducing them to engines that just remix bits of input is too far in the other direction - ML models absolutely do build conceptual abstractions of their training set, and they can create entirely new content guided by those abstractions.
The thing I’d say that pushes me “too far in the other direction” on this is how particularly in visual content generated by LLMs, there’s a very visible “statistical averageness” look that feels...
The thing I’d say that pushes me “too far in the other direction” on this is how particularly in visual content generated by LLMs, there’s a very visible “statistical averageness” look that feels exactly what one would expect to see if you had a machine that perfectly blended a large pile of pre-existing images. Human works don’t give this impression.
This, to me, makes it difficult to shake the notion that they’re “creating” by mashing up data instead of creating from scratch based on highly abstracted concepts and principles. Perhaps they aren’t abstracting or isolating enough, which leads to irrelevant/wrong details from training data getting encoded into the concepts/principles and then lending to that bizarre effect of “averageness”? I don’t know.
I believe this is largely because it is difficult to move far from the average in current models and frameworks. Significant portions of heavily weighted data likely are very similar in the same...
... there’s a very visible “statistical averageness”
I believe this is largely because it is difficult to move far from the average in current models and frameworks.
Significant portions of heavily weighted data likely are very similar in the same way that stock photos of office workers have the exact same "averageness". Image generators were initially trained on controlled data because it gave the best results the earliest, especially for GANs.
There are some renderings of textures and materials which are simply not consistent or of even remotely decent quality from base models, and this should be doubly true for elements of composition and lighting which are beyond laypersons to articulate and thus underrepresented in datasets.
There are actually many elements which are - generously speaking - subpar, but I don't think this will be the case when generators have sufficiently improved.
Got you, that makes a lot of sense! There actually are a couple of extra layers of (more or less) averaging going on in image models via the diffusion process and VAE, but even for text I know...
Got you, that makes a lot of sense! There actually are a couple of extra layers of (more or less) averaging going on in image models via the diffusion process and VAE, but even for text I know what you mean, the output can feels like the edges have all been rounded off a bit.
The way I’d visualise it is a bit like a line of best fit on a scatter plot. The cloud of dots is your training input in all its nuanced, granular glory - but the model only has space for the line. You ask it for a new output, it creates one by taking the equation of the line and giving you a point nearby: you can give it some leeway, introduce a little noise, but the chances are your generated output points from just the equation will be more clustered, more generic, more average than the input points that the line was drawn through in the first place.
If you repeat that equation fitting process a couple of billion times, you’ve pretty much got yourself a neural net. That’s a whole lot of degrees of freedom, and we’re figuring out more effective ways to make use of each one by the day, but ultimately the output is still the product of an enormous stack of common denominators. Genuinely novel, but often still a bit generic - for now, at least.
Imho its all murky and there no definitive technically calculated answer. If I personally distilled hundreds of images of Micky and start to draw and sell these images (few years ago) I will...
What’s the difference between me distilling the things I have seen into something new and me using a machine that does the same thing?
Imho its all murky and there no definitive technically calculated answer. If I personally distilled hundreds of images of Micky and start to draw and sell these images (few years ago) I will quickly get quite demanding and costly letters from Disney Co.
And honestly Im still not sure what the difference in this case between me and OpenAI? Its ok to steal (sorry distill) from general people but not from big corporations?
As with all these generative AI tools, it should be legal to train models on public-facing data, but the raw results should be public domain. Maximize the creative potential for the public, minimize the economic downsides. At best these companies should be able to sell access (both to their proprietary model and to the processing power needed to run it), but large companies should have no incentive to replace artists with them.
The output of generative AI has already been determined to be un-copyright-able in US court, which means they are indeed in the public domain at least in terms of US law. This is because copyright requires human authorship. Since the copyright of a creative work belongs to the creator, and non-humans cannot hold copyright, works of a non-human creator are thus in the public domain. The famous monkey selfie is also in the public domain for the same reason.
I love the monkey selfie example, but an important detail there is that the courts ruled the author to be whoever pressed the button on the camera, which is how the monkey was in contention for the copyright in the first place. Similarly, the plaintiff in the case from the first link explicitly chose to list the algorithm as the author of the works, not himself, which leads to the same outcome as the monkey case.
In practice, neural nets are a tool just as much as a word processor, or a 3D modelling package, or camera is: the human using the tool remains the author for copyright purposes.
That’s not to say there aren’t a million problems with copyright law itself, with its interactions with this technology, and with the ways that large companies are using the works of others. Those questions absolutely need to be figured out, but it’s going to be a lot more complex than just saying model output can’t be copyrighted.
I'm not under the impression that there's any case law to back this up as an actual legal claim. If there is definitely link it, as I'm happy to be corrected, but I don't think this is something that you can claim without pointing to actual decisions by the Copyright Office.
Copyright law is indeed a complex beast and has numerous problems even without AI copyright issues in the mix... especially in the music industry, which makes this example stand out.
It's an interesting question! A lot of the answers are't definitive yet - the field as a whole is new enough that things are still being litigated and figured out, not to mention the potential divergence between US/EU/China - but there are nods in that direction even in the ruling that denied copyright in the Thaler case:
And from the US Copyright Office guidance on AI works:
The official line right now is pretty solidly that AI is a tool, but the open question is where exactly the bar lies for creative control, and that'll affect the real-world outcomes fairly significantly.
The input needed to make something copyrightable has a lot of non-AI precedent, and I'd argue that the bar there is already absurdly low (certainly lower than two minutes it might take to compose a decent prompt and cherry-pick a good output), so I'd actually be fine with that requirement becoming stricter as part of broader reforms.
[Edit] I've also just realised that the first part of the US Copyright Office's ongoing research and reporting on the topic was just published on Wednesday: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-1-Digital-Replicas-Report.pdf
I probably will end up looking over it in the near future - it's highly relevant to what I do day-to-day - but right now I don't have time to start diving into a 70 page document, so I can't give you a good summary of what's in there!
AI to the rescue? ;)
Summary (no statement made on accuracy or quality):
This case gets trot out often for finding that AI art is non-copyrightable, but that's not actually what the case held. There are very, very important caveats which get skipped over that effectively make this case irrelevant with respect to whether or not you can copyright an AI generated image.
The most important fact is that the plaintiff never claimed to be the artist of the work. His position going into the case was that the AI engine (that he created) was the artist and that he owned the copyright by means of work for hire. The defense (the copyright office) did not challenge his assertion of authorship. As such, the only legal question under consideration was whether an AI could be considered the artist of a work for the purpose of copyright. As you mentioned, this was already a settled legal question, thus his lawsuit was denied.
Importantly, upon appeal he changed his position and claimed that he was the original artist and that the AI was only a tool, however appeals courts do not consider questions of fact, only law, and his attempt to change his position was denied and the original verdict upheld.
Ergo, the question as to whether you can copyright an AI image if a human claims to be the author is still legally unresolved.
It’s not black and white. Most AI works are a collaborative effort between humans and AI. Is using “content aware fill” in Photoshop going to invalidate your copyright? No way. So if you mash together a few song sections, each from an AI, that could have a copyright on it as well.
The case over the graphic novel specifically said he could still copyright the arrangement (since even though the images were from Midjourney, a human chose how to arrange them in combination with the writing), so yeah. The person I replied to referred to the "raw output" though, so I assume they were thinking of the AI generated work itself, not any later works derived from them.
Yeah, a lot of people don't realize that in a future where AI can replace artists, the big conglomerates already have the rights to enough high quality photographs (stock image companies), text (publishers), music (music labels), drawings (Disney and friends), etc saved up in their vaults to create competent models. The other day I read a paper of how a team recreated Google's BERT using one GPU trained for a week by using a higher quality text dataset, as it turns out you don't need web scale data but rather good quality data. This means that if AI is able to replace people then only big companies will be able to afford doing so and the only question left is how the split is gonna go between the people that execute the tech and the people that owns the rights.
Why? What’s the difference between me distilling the things I have seen into something new and me using a machine that does the same thing? Put another way, why shouldn’t my work product also be public domain then?
I mean, philosophically I think all things should be public domain. Copyright is an aberration that is actively harmful to artistic expression, trading the freedom of everyone for the profit of a select handful of commercial artists and the corporate fiefs that employ them.
This is especially true of music, which thrives on remixing existing ideas, setting new lyrics to existing music, setting existing words to new music, sampling existing things and building new things with them, etc.. That's how culture is formed: ideas become ubiquitous and are used by many people freely, not treated as property.
I also generally believe that any output from the current generation of generative tools is isn't art, as art is a process of human expression, not a productized end result. It lacks sufficient human involvement and control in the process. (e.g. science is a process of ascertaining fact, not an end result of that process.)
I would also be very offended if someone sent me an email at work that was written by an LLM. It's a statement that someone isn't worth the effort of communicating with properly (worse: they want to waste your time), or that they don't understand what they want to say enough to articulate it themselves...in which case a non-psychic word vomit generator isn't going to know any better than they do and is just useless noise.
It may have become that, but the original spirit of copyright was in fact to focus on protecting the smaller time creators. When you make a convuluted process to obtain copyright, the corporations win. Because most people won't know about it (and then corporations copyright their work), some of those thst do get bogged down in beauracray, and some of those may still have edge cases to argue. For a corporation, this would simply be someone's full time job.
But yes, corporations find loopholes and lobby behind closed doors. So it's been perverted to something that basically hoards IP's well past the creator's time to benefit. I see that more as a means to return to older time limits rather than abolish the whole thing.
I philosophically agree, but I'm worried in reality because where's the line? Once we establish that, people will simply skirt that line and guess who benefits from that most? Corporations will generate the art and hire some cheap "artist" who simply edits It just enough in some assembly line to be copyright. Small time creators would be legislated out of using it for any livelihoods.
On a technical levels: scale. You can only produce so many drawings in a day. A computer can generate thousands, and is more scalable and prectiable than a human. It's the same difference between you choosing to save a web page and a web crawler devouring terabytes of data that starts to significantly burden the server owner. Most people don't care about saving or copying a few pages here and there.
It's also an obvious logistical issue. You share a favorite web page and it's a symbiotic relationship. You get the communal value of sharing something you like, the website gets more traffic to show interest or deliver ads to. You can't advertise to bots (and I'm sure there's enough trackers to identify a human for a bot with some degree of accuracy). You run into the tracker problem once again when all your traffic is bots. Lots of server load with no human curiosity to engage woth with nor advertise to.
The objective reason is simply because the government determed it not being public domain by default. Reasons being due to historical events and the fact that these laws were designed originally from to protect the smaller, less law saavy inventors from being poached by corporate.
But you are more than free to release it if you feel differently
Another difference is that humans and LLMs have large differences in how they function.
LLMs are not capable of producing things they have not been trained on; they can only output giga-mashups comprised of tiny chopped-up bits of many hundreds or thousands of pieces of training material. Even the most novel looking LLM output is pieces of other existing work mashed up and reshaped.
Humans on the other hand, with exception to those with photographic memories, don’t typically make direct use of anything they experience. Instead, we filter input through numerous mind-lenses (which are unique to each individual) which is then boiled down to sets of highly abstracted concepts and principles. Furthermore, we can do this on very little input; we only need to see an object once to grasp its most essential traits. This means that even attempted copies, on some level, are some percentage original; we’re fighting our own nature to produce 1:1 copies and while our creations may have strong influences (blatant plagiarism aside), they’re original in a way that LLM output as it stands simply cannot be.
Now I have no doubt that eventually machine learning will progress to the point that it closely resembles human minds and gains similar capabilities. When that happens, I think there will be much more of an argument that their creations are original work, but we’ll be running into ethical dilemmas too because chances are that these constructs will have something approaching sentience, which makes using them as tools morally questionable.
I think separating the concepts of “novelty” and “creativity” is probably a good idea here. Creativity I’m pretty much with you on, it’s a much more complex philosophical concept that would necessitate intent and quite plausibly self-awareness under most definitions. But novelty - stuff that isn’t just mashed up pieces of other works - is a different beast.
The chopped-up fragments that language models work in are tokens, and at a good-enough approximation for now those are just individual words. Everything else is probability: what’s the likelihood of token X existing in this proximity to token Y.
It’s easy enough to show that statistical models can be novel in that context with a toy example: constrain the parameter count enough and all you’ve got room for is a basic grammar. Tokens with this mathematical property go here, tokens with that property go there, and suddenly the model has derived subject-verb-object word order in English. Now it can generate practically unlimited plausible, simple sentences with no need to mash up existing work.
It obviously gets vastly more complex with larger models. Those do have room to store entire token sequences from training work, or parameters that might recreate those sequences with high enough probability that they’ll come up eventually in normal use. It demonstrably happens, given they can sometimes quote existing work on demand. But it’s not all they do; that distillation of “essential traits” (a phrase that I really like in this context, btw!) is exactly how ML models compress hundreds of terabytes of training input down to a few gigabytes of parameters - and in turn can produce hundreds more terabytes of new, different content from those few GB of probabilities.
So yeah, I’m with you on the big questions. LLMs do function differently to humans, and creativity is a deep concept. But reducing them to engines that just remix bits of input is too far in the other direction - ML models absolutely do build conceptual abstractions of their training set, and they can create entirely new content guided by those abstractions.
The thing I’d say that pushes me “too far in the other direction” on this is how particularly in visual content generated by LLMs, there’s a very visible “statistical averageness” look that feels exactly what one would expect to see if you had a machine that perfectly blended a large pile of pre-existing images. Human works don’t give this impression.
This, to me, makes it difficult to shake the notion that they’re “creating” by mashing up data instead of creating from scratch based on highly abstracted concepts and principles. Perhaps they aren’t abstracting or isolating enough, which leads to irrelevant/wrong details from training data getting encoded into the concepts/principles and then lending to that bizarre effect of “averageness”? I don’t know.
I believe this is largely because it is difficult to move far from the average in current models and frameworks.
Significant portions of heavily weighted data likely are very similar in the same way that stock photos of office workers have the exact same "averageness". Image generators were initially trained on controlled data because it gave the best results the earliest, especially for GANs.
There are some renderings of textures and materials which are simply not consistent or of even remotely decent quality from base models, and this should be doubly true for elements of composition and lighting which are beyond laypersons to articulate and thus underrepresented in datasets.
There are actually many elements which are - generously speaking - subpar, but I don't think this will be the case when generators have sufficiently improved.
Got you, that makes a lot of sense! There actually are a couple of extra layers of (more or less) averaging going on in image models via the diffusion process and VAE, but even for text I know what you mean, the output can feels like the edges have all been rounded off a bit.
The way I’d visualise it is a bit like a line of best fit on a scatter plot. The cloud of dots is your training input in all its nuanced, granular glory - but the model only has space for the line. You ask it for a new output, it creates one by taking the equation of the line and giving you a point nearby: you can give it some leeway, introduce a little noise, but the chances are your generated output points from just the equation will be more clustered, more generic, more average than the input points that the line was drawn through in the first place.
If you repeat that equation fitting process a couple of billion times, you’ve pretty much got yourself a neural net. That’s a whole lot of degrees of freedom, and we’re figuring out more effective ways to make use of each one by the day, but ultimately the output is still the product of an enormous stack of common denominators. Genuinely novel, but often still a bit generic - for now, at least.
Imho its all murky and there no definitive technically calculated answer. If I personally distilled hundreds of images of Micky and start to draw and sell these images (few years ago) I will quickly get quite demanding and costly letters from Disney Co.
And honestly Im still not sure what the difference in this case between me and OpenAI? Its ok to steal (sorry distill) from general people but not from big corporations?