This again? Or still, I should more accurately say? The US Copyright office has already started ruling that non-human creators (i.e. computers) do not qualify for copyright. This is correct, and...
Exemplary
This again? Or still, I should more accurately say?
The US Copyright office has already started ruling that non-human creators (i.e. computers) do not qualify for copyright. This is correct, and should remain so, because we all know what automation of copyrightable works will do to the "economics" (both fiscal as well as consumption, thus directly affecting creation) of copyright. Of creation. Namely, flooding everything because computers never stop.
ALL ART is built upon that which comes before. Artists, of any flavor, any media, consume art of others in their area of interest(s). They then build upon what they've consumed. Are influenced by it. Inspired by it. They combine styles, they harness the ideas generated by consuming that prior art. Using them to craft works of their own. Which then go on to inspire and stimulate other artists, and so the creative wheels of the world turn.
Under US law, "look and feel" is not copyrightable. It's not anyway, but there's at least one case this non-lawyer is aware of:
A lot of this has been going on in the indie scenes for a while now. Certainly in books. Romance is an easy example because it's super high volume in the world of readable fiction. And indie authors have turned being nimble and voracious in their production of content into a specialty. There have been various authors who hit upon some "look and feel" that connected with the audience (i.e., consumers). They then stick with that, since it's selling, and start rolling out books with that "look and feel."
And some of these authors have had the inept audacity to be shocked when other authors, equally as eager to get sales because, you know money, started writing similar books. These upset authors are then doubly shocked when they find out, surprise, "look and feel" or "sounds like" or "that's like mine" is not a legal argument. Just because you write books about billionaires sweeping lonely or unfulfilled small town heroines off their feet to Big Cities for a Life of Luxury doesn't mean you own that concept, that "look and feel."
Similarly, having a "writing style" isn't a copyrightable concept. If you've certain turns of phrase, or a "method", or a "way words go down on the page" ... that's not copyrightable. It just isn't. Specific stories are copyrightable, how you write them isn't. The idea of them isn't. The "general framework of it" isn't.
Ask any copyright lawyer, and certainly any copyright judge, and they'll tell you specific works of art can enjoy copyright protection. If I write a story about one of those small town heroines, that story is mine. I don't own yours, or Shelia's, or Hari's.
Further, I can't stop Shelia or Hari from reading my book and writing their own version that "copies" mine. It doesn't matter that I set it in a tropical paradise where the Big Problem that arises midway through is a natural disaster, and every restaurant the couple dines in includes waitstaff doing performance art as they ferry the food out, and at the end they take a slow boat back to civilization to live happily ever after. Shelia and Hari can do that too, note for note, and as long as it's their words ... it's their story.
Them suing me, or me suing them makes lawyers happy, because hourly billing and up-front retainers, but it doesn't change the outcome; ideas are not copyrightable. Neither is "style". Specific works.
A whole lot of people who have never really studied copyright beyond some blog post by a pissed off "creator" upset at their lack of riches seem to think generative AI (AGI) is "stealing" when it analyzes a body of works.
How is that stealing? It's studying. If you go to college for a type of art, you study the art. How? By looking at, reading, listening, whatever, thousands and thousands of examples of that art. You look at the history of the creators of that art. You look at the origin of styles, movements, and influences. You're made aware of, encouraged to tabulate and consider, similarities and differences within that field. You'll be introduced, in academic detail, to how someone originates some "new thing" that begins to resonate with the audience, and how other artists join that thing.
It's how art works. It's how humans work. You don't even have to go to college to do it. You can sit down and study the same way as a student; by honestly and purposefully finding art, consuming it, thinking about it, and then further contemplating how you might incorporate what you've now concluded or even learned into either your analysis of further examples or in the creation of your own works inspired by your education.
How do you write a law that says "consumption of art for purposes of emulation is a violation of copyright?" Go on, I'll wait. Write the text of that law. But you have to do it without making human consumption of art illegal. It's not as easy as saying "computers bad, humans good." The computer processes, "reads" every book I bring up on my ereader app. Some ereaders are capable of, or could be if any programmer spends a few hours on the code, giving statistics like page count, reading time, search, and more; are those infringements?
My grandfather loved western movies. He kept a journal about the ones he watched. With notes. Who was in it, wrote it, directed it. What he thought about it. Themes. This guided his tastes; he knew he liked revenge westerns, but not brutal ones. And so on. These things played a role in which unwatched westerns he'd view next.
What if he'd done that on a computer? With a spreadsheet, or a notes app? How do you write that law so grandpa isn't an infringer now? What about formal students of a body of art; are they infringers now too? What if someone takes their observed statistics and notes and sits down to use it as their exacting guide as they try their hand as creation? Are they an infringer? Is the professor?
How do you prove art has been stolen? Can't copyright an idea, or a style. If you could, then every everything in all the art is theft.
How do you write a law that says "this is like that, so it's a violation" when you can't copyright style or idea? Do you just put in a new law that says now you can copyright style and ideas? Sit and think about that. Don't try to "win the debate", but just think about how that might work out.
Who owns, say, mafia movies? Does a different person own mafia books? What about mafia paintings? And do we draw a line between mafia paintings vs mafia digital wallpaper? What about mafia poetry; is that infringing on the mafia book author's copyright? What about superhero stories? Revenge stories? Does it matter if we have a romantic revenge story versus a tragic one? Or a happy ending one? Or a bittersweet one? Does the tone, "flavor", or "purpose" of the story make it a new idea that doesn't infringe on the one we already know someone owns?
Do we start looking at what "kind" of mafia story it is? "Well, Your Honor, plaintiff owns the concept of revenge mafia stories. But defendant's mafia story is one of love; the mafia is simply used as a setting."
And we'd have to repeat that basically infinitely, for all the genres, for all the flavors, and sub flavors, and sub-sub-sub-sub-sub flavors, and so on. It's a ridiculous as the concept of that company owned cloud computer endlessly churning out copyrightable works.
Laws, like copyright, deal with specifics. A specific thing is copyrightable. Not the feeling of it, or the tone of it, or the genre of it. Laws have to outline specific acts. Not "well, see, they meant to take what they'd learned and use it to create other stuff." That's not illegal. Copying is, learning isn't. And copying means literally copying, not "well it's sort of similar" copying. That's why it's called copyright; it's the right to issue exact copies.
You'd have to meticulously detail what "taking" means, what "learning" means, "using" means, "creating" means. How do you do that without making the mere act of someone reading a book, looking at a painting, watching a movie illegal? Are we to say artists would only exist in a legally provable vacuum in the wake of that law? And how do you prove that in the first place?
AI can't generate a copyright. This is the correct legal avenue, because it nips most of what laypeople want to get outraged about in the bud. Draws a line right under it. If I can't set up a computer system (or rent cloud time) and run an AI to "generate ideas", or even generate "whole art" that I can then take, sell, and (most importantly) sue others for infringing ... copyright is fine. The issue with AI isn't theft, it's swamping the system. And copyright isn't about ideas or creativity, but about finances.
You can't steal my book (legally) and sell it as your own. You can read my books, and write your own versions. I might be upset that you seem to have followed chapter by chapter to mine, but unless you literally copied it, you're just another author. You wrote your story. It has no legal weight that you read mine first. The same as all the books I read don't mean those authors can sue me.
Take this AI "influencer" out of Spain. Say it happens in New York next. Some Madison Avenue company generates Alison, aka Ali the hip influencer. She goes viral. She's not a real person, so she doesn't own her likeness (and, technically, it's not exactly settled law that even people own their own likenesses; only some states have those laws, and I don't believe there's a federal level version of it yet). The computer that generated her isn't a person either, so there's no copyright on her creation that transfers.
If a marketing person writes her copy, that's copyrightable. If a human created (and trademarked) her name, then they own Alison aka Ali so the others would have to use their own names.
But Ali herself isn't copyrightable; she's computer generated (not human artist using a computer generated, but AGI created). So four, a dozen, all the other marketing companies can immediately feed Ali's videos, pictures, whatever, into their own generative AIs and in a matter of days (hours?) have Ali plugging stuff for them. Which is what "a computer can't own a copyright" means; Ali was created by a computer, so no one owns her, which means the original Madison Avenue firm can't own her either so they can't sue the other companies for copying her. Copyright wouldn't attach to "other versions", or "copies" if you will, of Ali.
Which is the protection people pissed at computers flooding the copyright system want, and should be happy with. Sure many different versions of Ali could be created, but the inability to legally control them will act to limit how eager entities are to create these unownable "things". If my AGI creates a hit TV show, yours and four dozens others' could copy it immediately. I don't own the copy right (spaced to indicate what copyright means, the right to copy) because the computer created it, and can't transfer it to me, so I can't sue (and, at least for now, neither can the computer; check back for the fresh chaos when a computer achieves sentience).
This is a non-issue. Silverman is another indie author pissed she can't own an idea, or a style. What she owns is her text. Which she released to the public under copyright. Her story is hers. But only the story. That one, that specific story. Anyone else, even an AGI, isn't infringing if they read it, even if they analyze it. They're only infringing if they copy it and then release that copy. Even if they released that copy for free, it's still infringement; they'd be harming her sales market.
But simply reading her book, even reading it repeatedly and taking detailed and voluminous notes which you then studiously refer to for guidance as you write a story of your own ... that's not infringement.
Get up in arms, light the torches and find your pitchforks, if the Copyright Office is strong-armed into issuing copyrights for computer generated works. Keeping a human in the loop for issuance of copyright deals with just about every problematic issue in the whole thing. Companies won't be able to crank out endless "content", and human creativity still has financial value.
Copyright is about finances, not creativity. It offers a legal framework for allowing a creator to put a price on creativity that has been shaped into a specific form. Name, the form of a book, or a movie, a script, a painting, a song, whatever. A specific piece of art.
They have workarounds to copyright contemporary Mickey, but it also says this:
They have workarounds to copyright contemporary Mickey, but it also says this:
Additionally, Disney owns trademarks for Mickey Mouse. So, anyone attempting to copy and use the character on goods or services in commerce may be liable to Disney for trademark infringement (instead of for copyright liability) once the copyright expires. A trademark is any word, phrase, symbol, design, or combination of these things that identifies where goods or services come from. Trademarks help consumers distinguish sources of goods and services.
That being said, trademark usually only applies precisely to 'Trademarks help consumers distinguish sources of goods and services.' I'm betting using Mickey Mouse as a bit character, not a title...
That being said, trademark usually only applies precisely to 'Trademarks help consumers distinguish sources of goods and services.'
I'm betting using Mickey Mouse as a bit character, not a title one, could win a trademark battle so long as they didn't appear in any marketting images.
I wonder if anyone's made the argument that using copyrighted material in training data in a way that does not actually reproduce the copyrighted material as output is sort of like me reading a...
I wonder if anyone's made the argument that using copyrighted material in training data in a way that does not actually reproduce the copyrighted material as output is sort of like me reading a bunch of books by a single author and then writing a book that follows that other author's style.
I think there must be some layer of wiggle room for how complete the style mimicry is. I would say it's reasonable for humans to be able to imitate the style of someone else, because humans are...
I think there must be some layer of wiggle room for how complete the style mimicry is.
I would say it's reasonable for humans to be able to imitate the style of someone else, because humans are much more of a 'lossy' copy. And that most people copying style aren't, to my knowledge, prefixing their works with "written in the style of Sarah Silverman."
But a computer excels at copying things verbatim. And, to the best of most expert's knowledge, are not yet capable of original thought. It's more like running a copywritten work though a thesaurus than copying style from what you've learned at this juncture.
Saying that having a computer copying an author's style isn't infringment feels a lot like saying that transcoding a movie with a slight blur filter isn't infringement.
I would say you're right, but then again, people have been accused of writing the Marvel movies, and plot-wise they can be very hard to tell apart. Nevertheless, I haven't heard of anyone being...
I would say you're right, but then again, people have been accused of writing the Marvel movies, and plot-wise they can be very hard to tell apart. Nevertheless, I haven't heard of anyone being accused of copyright infringement for making another film that reuses a plot with different comic book characters.
I just had to come in and say that your joke did not go unnoticed, nor properly appreciated. But more to the point, this argument would literally make every version of Journey to the West illegal....
people have been accused of writing the Marvel movies
I just had to come in and say that your joke did not go unnoticed, nor properly appreciated.
I haven't heard of anyone being accused of copyright infringement for making another film that reuses a plot with different comic book characters.
But more to the point, this argument would literally make every version of Journey to the West illegal. Everything from Dragon Ball to The Miraculous Ladybug would be sued to oblivion for copying the concepts.
That was kind of my point. You can make two very similar movies without violating copyright, as long as they are different movies somehow. Here's an even better example: Ian Fleming's novel...
That was kind of my point. You can make two very similar movies without violating copyright, as long as they are different movies somehow.
Here's an even better example: Ian Fleming's novel "Thunderball" was adapted into a movie in 1965 starring Sean Connery. In 1983 a different movie studio adapted the same novel into "Never Say Never Again" also starring Sean Connery as James Bond. They are literally the exact same story with minor cosmetic differences.
I'm making fun of the Marvel movies, which despite different aesthetics (different characters, setting, music, etc.) the actual plot of each movie has gotten pretty repetitive.
I'm making fun of the Marvel movies, which despite different aesthetics (different characters, setting, music, etc.) the actual plot of each movie has gotten pretty repetitive.
Yeah, to me the legal question is that is algorithmically consuming copyrighted material and then manifesting it in a new form actually breaking copyright law.
Yeah, to me the legal question is that is algorithmically consuming copyrighted material and then manifesting it in a new form actually breaking copyright law.
Maybe, I don't know. But it seems to me like an easy distinction to make between LLM blackbox input-output and a person. I'm not an expert so I welcome an explanation why this isn't as simple as...
Maybe, I don't know. But it seems to me like an easy distinction to make between LLM blackbox input-output and a person. I'm not an expert so I welcome an explanation why this isn't as simple as it seems.
Most art today is created in digital creation tools like Adobe Photoshop. These tools are gaining increasingly advanced features like object removal, background auto-fill, and most recently, full...
Most art today is created in digital creation tools like Adobe Photoshop. These tools are gaining increasingly advanced features like object removal, background auto-fill, and most recently, full object generation. These are all "AI" features, though utilize different technologies.
Similarly, modern cameras take extreme liberties when processing sensor data. They use advanced models to identify features, and fill them in how they think they're supposed to look, to enhance the picture quality beyond what was captured.
LLMs, GANs, and diffusion models are the latest technologies and rightfully raise a lot of interesting questions of copyright and authorship. However, the line between human and computer-generated content has been blurred for a long time already, and it's not likely the genie will go back into its bottle.
As far as I know, this is more a feature of the cellphone camera segment of the market. My 5DIV, C100, and R50 have none of those 'features' and I'd be annoyed if they did. Including Sony's...
modern cameras take extreme liberties when processing sensor data.
As far as I know, this is more a feature of the cellphone camera segment of the market. My 5DIV, C100, and R50 have none of those 'features' and I'd be annoyed if they did.
Including Sony's in-camera authorization, which I'm sure every other manufacturer is following suit, the trend for professional cameras is proving that you really shot it and the AI stuff poses more ethical questions for professionals than the 'possibilities are endless' optimism crowd.
Hell, using the wrong LUT will adjust enough values that my organization will no longer call it a photo because they say it's technically been modified too much that it no longer represents reality.
Cameras still aren't a perfect replica of the human eye, so it's easy to "lie" about what things really looked like. This is true even without techniques like long-exposure shots. My favourite...
Cameras still aren't a perfect replica of the human eye, so it's easy to "lie" about what things really looked like. This is true even without techniques like long-exposure shots. My favourite example of this is the fact that auroras to the naked human eye usually just look like glowy clouds. They're still cool, but they rarely have the vibrant hues that a camera can capture.
Well, metadata included on the picture tells you everything about it. "Taking in 40 seconds worth of light" isn't a lie to me compared to "I know what the Aurora looks like so I replaced this one...
Well, metadata included on the picture tells you everything about it. "Taking in 40 seconds worth of light" isn't a lie to me compared to "I know what the Aurora looks like so I replaced this one with a better one."
Thanks for taking the time to write this. Taking Adobe from your example, shouldn't Adobe know how and what data each of their individual Photoshop AI tools use?
Thanks for taking the time to write this. Taking Adobe from your example, shouldn't Adobe know how and what data each of their individual Photoshop AI tools use?
Adobe Firefly encompasses all the creative generative AI models that will show up in Adobe products. Firefly’s first model trains on a database of Adobe Stock images, openly licensed content, and public domain content where copyright expired to turn text descriptions into AI creations.
I'm not an expert, but it helps if you start from the basic point that these models do not retain full text, they analyze and retain their analysis. They are trained on data similarly to how we...
I'm not an expert, but it helps if you start from the basic point that these models do not retain full text, they analyze and retain their analysis. They are trained on data similarly to how we observe it. The model, to my knowledge, does not have the full text of Silverman's book in its memory. This claim by Silverman's team was disputed by the creators of the models passing hese legal tests via presented arguments.
And now we're back to what I said: the issue comes down to the creative process and human-style memory. If all human memory is simply a series of fleshy analyses of a work, how do we differentiate between that and computational analysis that leads to a similar result?
At risk of undercutting my previous comment, we can draw the line at computational "inspiration" and the claim that these materials were not created with the intent of being LLM training materials.
I'm not a philosophy major either (there seems to be a lot of things I'm not the expert of) and this probably sounds a bit ignorant - but why focus on the ambiguous end of creative process instead...
If all human memory is simply a series of fleshy analyses of a work, how do we differentiate between that and computational analysis that leads to a similar result?
I'm not a philosophy major either (there seems to be a lot of things I'm not the expert of) and this probably sounds a bit ignorant - but why focus on the ambiguous end of creative process instead of the easily observable? Is the thing that's spitting the end result out a person or a machine?
The point I tried to get to is that we don't need to differentiate as much as take issue with the training itself. I'm legally allowed to read any work Silverman puts out as a human being, but to...
The point I tried to get to is that we don't need to differentiate as much as take issue with the training itself. I'm legally allowed to read any work Silverman puts out as a human being, but to train a model is a different endeavor. You can't do much once the model has analyzed it, even purge the data (you'd need to generate a new model). You can go after the people training the model for an unauthorized use of your work, since this isn't reading, it's using it for a commercial endeavor.
You wouldn't necessarily have to generate an entirely new model to remove a bit of data. It is possible to create negative weights that cancel out the bias introduced by the target material during...
You can't do much once the model has analyzed it, even purge the data (you'd need to generate a new model).
You wouldn't necessarily have to generate an entirely new model to remove a bit of data. It is possible to create negative weights that cancel out the bias introduced by the target material during the initial training. Now, it might not be easy, and might end up being as computationally expensive as generating a new model. Or, it might be as simple as training a new Anti-Silverman lora. I'm not really aware of any research into negating the weights of specific training data. But, in theory at least, it should be possible to teach a model to ignore what it learned about a specific piece of training data, assuming that the data was actually a part of its training. If the data that we negate wasn't actually a part of the initial training data, then we'd just be introducing an entirely new bias into the output, sort of like asking GPT to "write me a book that sounds like the exact opposite of Sarah Silverman."
Edit. I think I've misunderstood you, too drunk. My apologies.
I suppose we disagree there because I see a clear difference between a person reading a lot of books and a machine with a perfect memory taking all that information in in an instant.
But despite that, you helped me understand your point of view, at least a bit. Thank you.
Edit. I think I've misunderstood you, too drunk. My apologies.
How involved does a human need to be for the end result to be considered their result? If I push a button and get some random art, am I responsible for the end result since the art would not have...
How involved does a human need to be for the end result to be considered their result? If I push a button and get some random art, am I responsible for the end result since the art would not have existed had I not pushed the button? If I paint all but one pixel of a piece, and ask a machine to randomly fill in the remaining pixel, am I responsible for the end result even though a machine decided the final steps for me? Where is the sweet spot where we can demarcate the point at which an end result no longer has protection? If we cannot provide one, then I suggest that we err on the side of caution, since otherwise people will invest real time and effort (however minimal or maximal that be) into creating works that are not protected and indeed considered infringing, without knowing that they are doing so.
I think if you use AI models that are infringing copyrights, your end result shouldn't be copyright protected either. At least not without an agreement/settlement with the original copyright...
I think if you use AI models that are infringing copyrights, your end result shouldn't be copyright protected either. At least not without an agreement/settlement with the original copyright holder(s). But that's really not my area, I'm sure people much smarter than me have some kind of precedence to cite.
I don't mind AI, it's the stealing of others' work that's the problem. If a program can generate beautiful art with you inputting one pixel there, I don't see a problem. That's probably more of a question about the license agreement with the program's developer.
Fortunately, there's already strong precedent that non-human works can't be copyright protected in the US. There's a famous case of a photo taken by a monkey not being protected by copyright. If...
Fortunately, there's already strong precedent that non-human works can't be copyright protected in the US. There's a famous case of a photo taken by a monkey not being protected by copyright. If machine learning output can't be effectively monetized, I feel this solves a lot of issues.
But you can't ensure that if you use copyrighted material as input, in fact you're increasing the probability that it will reproduce it. Plenty of examples where image generators reproduced an...
in a way that does not actually reproduce the copyrighted material as output
But you can't ensure that if you use copyrighted material as input, in fact you're increasing the probability that it will reproduce it. Plenty of examples where image generators reproduced an entire signature of the original author. Such large fragments cannot be considered "style".
If you don't use such material as input, any output that contains a fragment of copyrighted material is purely coincidental and highly improbable.
Well it's not, because you're a person and own your own creative output. AI is a tool that 'reads' the book and someone else prompts it to produce the output. If the two situations were equivalent...
Well it's not, because you're a person and own your own creative output. AI is a tool that 'reads' the book and someone else prompts it to produce the output. If the two situations were equivalent the AI would own the rights to the art it generates, which it obviously isn't capable of.
U.S. District Judge Vince Chhabria on Monday offered a full-throated denial of one of the authors’ core theories that Meta’s AI system is itself an infringing derivative work made possible only by information extracted from copyrighted material. “This is nonsensical,” he wrote in the order. “There is no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs’ books.”
Another of Silverman’s arguments that every result produced by Meta’s AI tools constitutes copyright infringement was dismissed because she didn’t offer evidence that any of the outputs “could be understood as recasting, transforming, or adapting the plaintiffs’ books.” Chhabria gave her lawyers a chance to replead the claim, along with five others that weren’t allowed to advance.
…
The ruling builds upon findings from another federal judge overseeing a lawsuit from artists suing AI art generators over the use of billions of images downloaded from the Internet as training data. In that case, U.S. District Judge William Orrick similarly delivered a blow to fundamental contentions in the lawsuit by questioning whether artists can substantiate copyright infringement in the absence of identical material created by the AI tools. He called the allegations “defective in numerous respects.”
…
This means that plaintiffs across most cases will have to present evidence of infringing works produced by AI tools that are identical to their copyrighted material. This potentially presents a major issue because they have conceded in some instances that none of the outputs are likely to be a close match to material used in the training data. Under copyright law, a test of substantial similarity is used to assess the degree of similarity to determine whether infringement has occurred.
I feel like chasing copyright is too hard of a battle to win.. it's too fungible because it's almost impossible for it to be logically consistent across different works...I recall reading an...
I feel like chasing copyright is too hard of a battle to win.. it's too fungible because it's almost impossible for it to be logically consistent across different works...I recall reading an article about that a ways back. And the end result of that will be "largest army of lawyers wins".
But I feel like pushing for a law on the basis of control over one's identity might stand a better chance. Sarah Silverman wouldn't be able to be impersonated if her works were not consumed by the LLM. So a law that mandates that LLM's must exclude the material of anyone whom requests it from the training data, regardless if they own the direct copyright. Perhaps with an exception for data that is legally required to be public. Maybe bake it into an extension of GDPR.
The tech companies would argue this is impossible, but I would question whether it is actually impossible to comply with or merely expensive to comply with. Because those are very different things. We really shouldn't allow "but it would really be expensive to follow the law" to be a valid defense for violating it.
Otherwise copyright law would instantly lose any and all teeth, at least for consumers.
Feels like in a lot of cases, where nations rule against ai's use under this or that circumstance (or perhaps unions get agreements trying to protect their jobs) they're handing a huge competitive...
Feels like in a lot of cases, where nations rule against ai's use under this or that circumstance (or perhaps unions get agreements trying to protect their jobs) they're handing a huge competitive advantage to other nations.
EG, if animation can be made 80% more cost effective using ai, there are 3 possible outcomes:
the US can be part of a booming animation scene that drastically increases output volume at lower per-unit cost retaining its market share and total jobs.
no increase in volume follows and the number of jobs shrinks, but the US retains its current market share.
market share moves to places that permit use of AI and the sector more or less entirely vanishes in the US.
That's just a single example, using the US (who I assume are the current world leaders in animation), but the same applies to many industries I'd imagine.
I mean, slavery is a pretty big competitive advantage too. There's a reason that few companies can honestly say "we know with reasonable certainty there's no slavery in our supply chain." Game...
I mean, slavery is a pretty big competitive advantage too. There's a reason that few companies can honestly say "we know with reasonable certainty there's no slavery in our supply chain."
A crucial difference there is that slavery is immoral, whereas AI is amoral. I don't see that the AI revolution is meaningfully different from any other technological revolution ethically speaking.
A crucial difference there is that slavery is immoral, whereas AI is amoral. I don't see that the AI revolution is meaningfully different from any other technological revolution ethically speaking.
Is it amoral if it's undermining the labor of artists? There's always a capacity for audiences to support human-made art and punish studios that rely heavily on AI art. Literally nobody cares...
Is it amoral if it's undermining the labor of artists? There's always a capacity for audiences to support human-made art and punish studios that rely heavily on AI art. Literally nobody cares about the price of a movie ticket, so they won't be passing the savings onto consumers.
Was the loom not amoral when it undermined the labor of seamsters (?) and seamstresses? Was CGI not amoral when it undermined the labor of traditional practical effects teams? I'm not trying to...
Was the loom not amoral when it undermined the labor of seamsters (?) and seamstresses? Was CGI not amoral when it undermined the labor of traditional practical effects teams?
I'm not trying to say what should happen, I'm saying that I think the Luddites failure was inevitable, and I think the situations are almost perfectly analogous.
People did make the case it was wrong to replace seamstresses, but it benefited consumers enough for society to move on. Consumers don't benefit from AI generated art. If not immoral, it can be...
People did make the case it was wrong to replace seamstresses, but it benefited consumers enough for society to move on. Consumers don't benefit from AI generated art. If not immoral, it can be certainly be understood to be ethically wrong. You don't need to be a Luddite to support audiences and the artistic community enforcing a norm.
In my personal use case, I'm a bit of a hoarder when it comes to wallpaper art, and I've noticed that 90% (ish) of the wallpapers I'm downloading and using these days are AI generated, whether...
In my personal use case, I'm a bit of a hoarder when it comes to wallpaper art, and I've noticed that 90% (ish) of the wallpapers I'm downloading and using these days are AI generated, whether generated by myself or by others.
Now in this case no money was changing hands in the first place, so nobody is missing out, but I think it demonstrates how consumers can benefit from AI generated art. They can have a more tailored experience, and those without innate ability can become - to some extent - creators.
Actually, did it benefit society now, hundreds if years on? Clothes are now substantially less durable and repairable. The neverending march of mechanization has clothed us all in cheap,...
Actually, did it benefit society now, hundreds if years on?
Clothes are now substantially less durable and repairable. The neverending march of mechanization has clothed us all in cheap, disposable garbage. How much energy is wasting growing, harvesting, and manufacturing cheap cotton/poly blends that fall apart inside of 2 years, and can't be recycled or reused?
Back in the days of seamstresses, you might have only had a handful of shits as an adult. But between the durability and repair, you'd have those shirts almost your entire life.
I could see a reasonable case that sewing machines that enabled seamstresses to work faster/better on their own terms were a good thing. I think factory-style clothing manufacturing is less of a win.
Now, to take the case of my career, software engineering: there are many tools that have sped up development along the way, and potentially reduced the number of developers required, or have...
Now, to take the case of my career, software engineering: there are many tools that have sped up development along the way, and potentially reduced the number of developers required, or have enabled greater productivity from those same developers. One such tool is now AI.
I wonder if you consider it ethically wrong to use AI in this industry as well? If not, I'm curious as to the distinction in your eyes.
(Two separate points, so in separate posts as I think they're easier to respond to this way, should further discussion be desired anyway.)
I don't disagree that AI itself is amoral. Like a hammer, or a gun, or a nuke, it's only as moral as the weilder. I think I'd be a pretty sweet weilder of justice if I could shoot...
I don't disagree that AI itself is amoral. Like a hammer, or a gun, or a nuke, it's only as moral as the weilder. I think I'd be a pretty sweet weilder of justice if I could shoot whatever/whomever I wanted consequence-free, but I doubt most people would trust my word on that one....with good reason.
Luddites weren't smashing machines because they were anti-technology. They were anti-worker-exploitation. And as far as I can see, most non-trivial uses of it are primarily aiming at exploiting humans, especially if they have labor protections. If it wasn't, I don't think Hollywood have been so stubborn about the writer's strike.
Interesting. I went and read the top link on wikipedia about luddites and you're not wrong about that. I'm certainly all for labor rights, and not opposed to strikes etc.. That said, restriction...
I'm certainly all for labor rights, and not opposed to strikes etc.. That said, restriction of a useful tool seems like an unworkable approach to protecting worker rights.
I think it'd be strange if the work of 10 people could be achieved by 1 person, and we chose to continue having 10 people do it anyway. That would be a disservice to everyone involved. And I really don't think it's feasible. The cost savings are potentially so extreme that someone will fill that gap.
The problem is that said 1 person is going to be paid roughly the same now doing 10x the work. And that 9 other people now need to find a new job. In a world that provides no assistance to retrain...
The problem is that said 1 person is going to be paid roughly the same now doing 10x the work. And that 9 other people now need to find a new job. In a world that provides no assistance to retrain and punishes you for transitioning out of your expertise.
Maybe we should be focusing on making sure labor-saving tools result in hours-reductions rather than hiring reductions.
Couldn't agree more on that! That's where I'd want to focus efforts: societal shift. As opposed to keeping people in - what becomes in essence - pointless labor.
Couldn't agree more on that!
That's where I'd want to focus efforts: societal shift. As opposed to keeping people in - what becomes in essence - pointless labor.
I don't think the economics checks out here (audiences are limited in what they want to consume and what they time-wise can consume) The product of each country's animation industry can't be...
increases output volume at lower per-unit cost retaining its market share and total jobs
I don't think the economics checks out here (audiences are limited in what they want to consume and what they time-wise can consume)
market share moves to places that permit use of AI and the sector more or less entirely vanishes in the US.
The product of each country's animation industry can't be directly compared, since there's an interest and value placed on the country it's coming from by audiences and the community. The US is also only competing with a handful of Anglophone countries that it could relatively easily coordinate with in terms of regulation.
As to point 1: Studios/nations are in competition with one-another. If Netflix can produce 10 times the content at the same cost and quality as Amazon Prime, they're inevitably going to capture...
As to point 1:
Studios/nations are in competition with one-another. If Netflix can produce 10 times the content at the same cost and quality as Amazon Prime, they're inevitably going to capture more market-share I'd imagine.
Point 2:
You might be right that internally regulation within the Anglosphere could be possible. Maybe.
Aside:
I remain unconvinced that legislating specifically to reduce worker productivity is intrinsically a good idea. It appears to me preferable to find an alternative approach that embraces AI whilst considering existing industry professionals during the transitional phase (should the change prove so dramatic).
In point 1, my point is that there cannot be sufficiently greater overall output to maintain the number of jobs, since consumption cannot keep up with the efficiency (you would need a roughly 2x...
In point 1, my point is that there cannot be sufficiently greater overall output to maintain the number of jobs, since consumption cannot keep up with the efficiency (you would need a 5x increase roughly 2x increase in supply/consumption to compensate for an 80% increase in labor-efficiency). Also, the market for media works dramatically differently from, say, commodities. Brand loyalty/perception, goodwill, and a bunch of other woo-wooey stuff matter a lot (these factors can also be made to matter more under cultural enforcement).
My view, generally, is that the law and regulations should be set up to assist in cultural enforcement of norms around intellectual property, value of artistic labor, etc. (ex. disclosures of to what extent a product is AI-generated).
In other words, downloading a number of copyright books from books3 is still clearly illegal. OpenAI & Meta are still accused of knowingly infringing copyright. If only they had legally purchased...
Notably, Meta didn’t move to dismiss the allegation that the copying of books for purposes of training its AI model rises to the level of copyright infringement.
In other words, downloading a number of copyright books from books3 is still clearly illegal.
OpenAI & Meta are still accused of knowingly infringing copyright.
If only they had legally purchased digital copies, then they apparently wouldn't have broken any laws.
Napkin math time, because I don't think it will be hard to prove intentional infringement in this case. Call books3 150,000 illicit books (to give some leeway for public domain works like...
In a case where the copyright owner sustains the burden of proving, and the court finds, that infringement was committed willfully, the court in its discretion may increase the award of statutory damages to a sum of not more than $150,000.
Napkin math time, because I don't think it will be hard to prove intentional infringement in this case.
Call books3 150,000 illicit books (to give some leeway for public domain works like shakespear). Since max punitive damages are per-work, square that: $22.5 Billion dollars.
Now, that's a per-offense, per-work maximum. They've narrowed out that using the model is not infringement, but how about training it?
I think it would be reasonable, based on the logic that the various copyright holders use to punish users, to count each full read of books3 as an offense. And that number is likely in the multiple billions, given parallel processing, but we'll just call it 1 billion. We now have a total payout of 225,000 trillion dollars.
A $15,000 annual UBI for the world (10 billion people) sits around 150 trillion dollars annually.
Turns out AI creators have worked themselves into a legal hole, where a jury would be legaly in the right to award authors in books3 enough to fund a worldwide UBI of $15,000 for...150,000 years.
You might think this reasonable, but the US courts likely would not. Copyright protects against copying. It doesn't protect against the one copy being read and reread a thousand times. Think of...
I think it would be reasonable, based on the logic that the various copyright holders use to punish users, to count each full read of books3 as an offense.
You might think this reasonable, but the US courts likely would not.
Copyright protects against copying. It doesn't protect against the one copy being read and reread a thousand times. Think of libraries, which offer both physical and digital books.
The US law system calculates copyright harm based off lost profits/ damages. The way they do this is kind of bullshit, but not the way you articulated it. They assume that every download would have been a legitimate purchase, and so calculate the list price times the number of downloads.
In this case, either courts would calculate how much FB would have to spend to purchase a digital copy at list for every copy they made, or if there were no digital copies for purchase, it would have the trickier job of figuring out how much it would cost FB to purchase distribution rights.
Also, the statutory damages capped at $150k mean plaintiffs have to prove intent, which seems reasonable, but is still pending.
You're missing a thing from the previous subsection of statuatory damages. Although I did as well. I quoted directly from the code. The previous subsection: So to recap: Default is damage-incurred...
You're missing a thing from the previous subsection of statuatory damages. Although I did as well. I quoted directly from the code. The previous subsection:
the copyright owner may elect, at any time before final judgment is rendered, to recover, instead of actual damages and profit, an award of statutory damages for all infringements involved in the action, with respect to any one work, for which any one infringer is liable individually, or for which any two or more infringers are liable jointly and severally, in a sum of not less than $750 or more than $30,000 as the court considers just. For the purposes of this subsection, all the parts of a compilation or derivative work constitute one work.
So to recap:
Default is damage-incurred profits/loss, as you describe.
However, copyright owner may at any time before judgement, opt for statutory damages instead. Normally capped at $30k per work, which can be upped to $150k if infringement is willful.
And the bit I missed: It explicitly is for all violations (per work) in the action. So my potential $22.5B judgement stands, but not the subsequent copying math.
If they opted for their copying math along with profit/loss calcs, they could probably surpass that $150k for many works. They could probably reasonably argue, at a bare minimum, they would need to legally purchase 1 license for every copy online simultaneously. So multiply retail cost by number of computing threads...
I would say proving intent would be pretty easy in this case, but IANAL. Point at books3, question company: did you or did you not use this for the sole purpose of making this product and then profiting off of its use?
Without this illegal source, would the LLM be half as functional as it is? The multi-billion-dollar question.
We decided that results of LLM is not copyrightable. So, if LLM produces image of that looks like Mickey Mouse I`m free to use it? Honestly I think we should not focus on question "is results of...
We decided that results of LLM is not copyrightable. So, if LLM produces image of that looks like Mickey Mouse I`m free to use it?
Honestly I think we should not focus on question "is results of LLM are copyrightable, or not?". We should focus on question: does this LLM was trained on materials from authors that gives consent to such training, or not?
The reason this gets debated is that there aren't any laws about AI training material. It's a new issue that's come up. (For people, you can use whatever training material that you have access...
The reason this gets debated is that there aren't any laws about AI training material. It's a new issue that's come up. (For people, you can use whatever training material that you have access to.)
People try to come up with a legal arguments why it's not allowed, based on existing laws, none of which are a perfect fit. Also, companies use technical measures and their terms of use agreements to restrict it.
I'm rather confident that is not how it works. If you share images containing a character with the likeness of a trademarked character, in this case Mickey Mouse, you are potentially violating...
We decided that results of LLM is not copyrightable. So, if LLM produces image of that looks like Mickey Mouse I`m free to use it?
I'm rather confident that is not how it works. If you share images containing a character with the likeness of a trademarked character, in this case Mickey Mouse, you are potentially violating Disney's trademark.
This again? Or still, I should more accurately say?
The US Copyright office has already started ruling that non-human creators (i.e. computers) do not qualify for copyright. This is correct, and should remain so, because we all know what automation of copyrightable works will do to the "economics" (both fiscal as well as consumption, thus directly affecting creation) of copyright. Of creation. Namely, flooding everything because computers never stop.
ALL ART is built upon that which comes before. Artists, of any flavor, any media, consume art of others in their area of interest(s). They then build upon what they've consumed. Are influenced by it. Inspired by it. They combine styles, they harness the ideas generated by consuming that prior art. Using them to craft works of their own. Which then go on to inspire and stimulate other artists, and so the creative wheels of the world turn.
Under US law, "look and feel" is not copyrightable. It's not anyway, but there's at least one case this non-lawyer is aware of:
A lot of this has been going on in the indie scenes for a while now. Certainly in books. Romance is an easy example because it's super high volume in the world of readable fiction. And indie authors have turned being nimble and voracious in their production of content into a specialty. There have been various authors who hit upon some "look and feel" that connected with the audience (i.e., consumers). They then stick with that, since it's selling, and start rolling out books with that "look and feel."
And some of these authors have had the inept audacity to be shocked when other authors, equally as eager to get sales because, you know money, started writing similar books. These upset authors are then doubly shocked when they find out, surprise, "look and feel" or "sounds like" or "that's like mine" is not a legal argument. Just because you write books about billionaires sweeping lonely or unfulfilled small town heroines off their feet to Big Cities for a Life of Luxury doesn't mean you own that concept, that "look and feel."
Similarly, having a "writing style" isn't a copyrightable concept. If you've certain turns of phrase, or a "method", or a "way words go down on the page" ... that's not copyrightable. It just isn't. Specific stories are copyrightable, how you write them isn't. The idea of them isn't. The "general framework of it" isn't.
Ask any copyright lawyer, and certainly any copyright judge, and they'll tell you specific works of art can enjoy copyright protection. If I write a story about one of those small town heroines, that story is mine. I don't own yours, or Shelia's, or Hari's.
Further, I can't stop Shelia or Hari from reading my book and writing their own version that "copies" mine. It doesn't matter that I set it in a tropical paradise where the Big Problem that arises midway through is a natural disaster, and every restaurant the couple dines in includes waitstaff doing performance art as they ferry the food out, and at the end they take a slow boat back to civilization to live happily ever after. Shelia and Hari can do that too, note for note, and as long as it's their words ... it's their story.
Them suing me, or me suing them makes lawyers happy, because hourly billing and up-front retainers, but it doesn't change the outcome; ideas are not copyrightable. Neither is "style". Specific works.
A whole lot of people who have never really studied copyright beyond some blog post by a pissed off "creator" upset at their lack of riches seem to think generative AI (AGI) is "stealing" when it analyzes a body of works.
How is that stealing? It's studying. If you go to college for a type of art, you study the art. How? By looking at, reading, listening, whatever, thousands and thousands of examples of that art. You look at the history of the creators of that art. You look at the origin of styles, movements, and influences. You're made aware of, encouraged to tabulate and consider, similarities and differences within that field. You'll be introduced, in academic detail, to how someone originates some "new thing" that begins to resonate with the audience, and how other artists join that thing.
It's how art works. It's how humans work. You don't even have to go to college to do it. You can sit down and study the same way as a student; by honestly and purposefully finding art, consuming it, thinking about it, and then further contemplating how you might incorporate what you've now concluded or even learned into either your analysis of further examples or in the creation of your own works inspired by your education.
How do you write a law that says "consumption of art for purposes of emulation is a violation of copyright?" Go on, I'll wait. Write the text of that law. But you have to do it without making human consumption of art illegal. It's not as easy as saying "computers bad, humans good." The computer processes, "reads" every book I bring up on my ereader app. Some ereaders are capable of, or could be if any programmer spends a few hours on the code, giving statistics like page count, reading time, search, and more; are those infringements?
My grandfather loved western movies. He kept a journal about the ones he watched. With notes. Who was in it, wrote it, directed it. What he thought about it. Themes. This guided his tastes; he knew he liked revenge westerns, but not brutal ones. And so on. These things played a role in which unwatched westerns he'd view next.
What if he'd done that on a computer? With a spreadsheet, or a notes app? How do you write that law so grandpa isn't an infringer now? What about formal students of a body of art; are they infringers now too? What if someone takes their observed statistics and notes and sits down to use it as their exacting guide as they try their hand as creation? Are they an infringer? Is the professor?
How do you prove art has been stolen? Can't copyright an idea, or a style. If you could, then every everything in all the art is theft.
How do you write a law that says "this is like that, so it's a violation" when you can't copyright style or idea? Do you just put in a new law that says now you can copyright style and ideas? Sit and think about that. Don't try to "win the debate", but just think about how that might work out.
Who owns, say, mafia movies? Does a different person own mafia books? What about mafia paintings? And do we draw a line between mafia paintings vs mafia digital wallpaper? What about mafia poetry; is that infringing on the mafia book author's copyright? What about superhero stories? Revenge stories? Does it matter if we have a romantic revenge story versus a tragic one? Or a happy ending one? Or a bittersweet one? Does the tone, "flavor", or "purpose" of the story make it a new idea that doesn't infringe on the one we already know someone owns?
Do we start looking at what "kind" of mafia story it is? "Well, Your Honor, plaintiff owns the concept of revenge mafia stories. But defendant's mafia story is one of love; the mafia is simply used as a setting."
And we'd have to repeat that basically infinitely, for all the genres, for all the flavors, and sub flavors, and sub-sub-sub-sub-sub flavors, and so on. It's a ridiculous as the concept of that company owned cloud computer endlessly churning out copyrightable works.
Laws, like copyright, deal with specifics. A specific thing is copyrightable. Not the feeling of it, or the tone of it, or the genre of it. Laws have to outline specific acts. Not "well, see, they meant to take what they'd learned and use it to create other stuff." That's not illegal. Copying is, learning isn't. And copying means literally copying, not "well it's sort of similar" copying. That's why it's called copyright; it's the right to issue exact copies.
You'd have to meticulously detail what "taking" means, what "learning" means, "using" means, "creating" means. How do you do that without making the mere act of someone reading a book, looking at a painting, watching a movie illegal? Are we to say artists would only exist in a legally provable vacuum in the wake of that law? And how do you prove that in the first place?
AI can't generate a copyright. This is the correct legal avenue, because it nips most of what laypeople want to get outraged about in the bud. Draws a line right under it. If I can't set up a computer system (or rent cloud time) and run an AI to "generate ideas", or even generate "whole art" that I can then take, sell, and (most importantly) sue others for infringing ... copyright is fine. The issue with AI isn't theft, it's swamping the system. And copyright isn't about ideas or creativity, but about finances.
You can't steal my book (legally) and sell it as your own. You can read my books, and write your own versions. I might be upset that you seem to have followed chapter by chapter to mine, but unless you literally copied it, you're just another author. You wrote your story. It has no legal weight that you read mine first. The same as all the books I read don't mean those authors can sue me.
Take this AI "influencer" out of Spain. Say it happens in New York next. Some Madison Avenue company generates Alison, aka Ali the hip influencer. She goes viral. She's not a real person, so she doesn't own her likeness (and, technically, it's not exactly settled law that even people own their own likenesses; only some states have those laws, and I don't believe there's a federal level version of it yet). The computer that generated her isn't a person either, so there's no copyright on her creation that transfers.
If a marketing person writes her copy, that's copyrightable. If a human created (and trademarked) her name, then they own Alison aka Ali so the others would have to use their own names.
But Ali herself isn't copyrightable; she's computer generated (not human artist using a computer generated, but AGI created). So four, a dozen, all the other marketing companies can immediately feed Ali's videos, pictures, whatever, into their own generative AIs and in a matter of days (hours?) have Ali plugging stuff for them. Which is what "a computer can't own a copyright" means; Ali was created by a computer, so no one owns her, which means the original Madison Avenue firm can't own her either so they can't sue the other companies for copying her. Copyright wouldn't attach to "other versions", or "copies" if you will, of Ali.
Which is the protection people pissed at computers flooding the copyright system want, and should be happy with. Sure many different versions of Ali could be created, but the inability to legally control them will act to limit how eager entities are to create these unownable "things". If my AGI creates a hit TV show, yours and four dozens others' could copy it immediately. I don't own the copy right (spaced to indicate what copyright means, the right to copy) because the computer created it, and can't transfer it to me, so I can't sue (and, at least for now, neither can the computer; check back for the fresh chaos when a computer achieves sentience).
This is a non-issue. Silverman is another indie author pissed she can't own an idea, or a style. What she owns is her text. Which she released to the public under copyright. Her story is hers. But only the story. That one, that specific story. Anyone else, even an AGI, isn't infringing if they read it, even if they analyze it. They're only infringing if they copy it and then release that copy. Even if they released that copy for free, it's still infringement; they'd be harming her sales market.
But simply reading her book, even reading it repeatedly and taking detailed and voluminous notes which you then studiously refer to for guidance as you write a story of your own ... that's not infringement.
Get up in arms, light the torches and find your pitchforks, if the Copyright Office is strong-armed into issuing copyrights for computer generated works. Keeping a human in the loop for issuance of copyright deals with just about every problematic issue in the whole thing. Companies won't be able to crank out endless "content", and human creativity still has financial value.
Copyright is about finances, not creativity. It offers a legal framework for allowing a creator to put a price on creativity that has been shaped into a specific form. Name, the form of a book, or a movie, a script, a painting, a song, whatever. A specific piece of art.
But you can't copyright idea or style.
Might want to look into how Disney has handled those laws since then. If your animation "looks and feels" like mickey mouse, you will lose.
That sounds like trademark law.
It’s not in this case.
As an example https://blogs.luc.edu/ipbytes/2023/08/13/is-disney-losing-mickey-mouse-because-of-copyright-law/
They have workarounds to copyright contemporary Mickey, but it also says this:
That being said, trademark usually only applies precisely to 'Trademarks help consumers distinguish sources of goods and services.'
I'm betting using Mickey Mouse as a bit character, not a title one, could win a trademark battle so long as they didn't appear in any marketting images.
I wonder if anyone's made the argument that using copyrighted material in training data in a way that does not actually reproduce the copyrighted material as output is sort of like me reading a bunch of books by a single author and then writing a book that follows that other author's style.
I think where things get tricky is that the illegality of that becomes extremely questionable. You explicitly can't copyright style.
I think there must be some layer of wiggle room for how complete the style mimicry is.
I would say it's reasonable for humans to be able to imitate the style of someone else, because humans are much more of a 'lossy' copy. And that most people copying style aren't, to my knowledge, prefixing their works with "written in the style of Sarah Silverman."
But a computer excels at copying things verbatim. And, to the best of most expert's knowledge, are not yet capable of original thought. It's more like running a copywritten work though a thesaurus than copying style from what you've learned at this juncture.
Saying that having a computer copying an author's style isn't infringment feels a lot like saying that transcoding a movie with a slight blur filter isn't infringement.
I would say you're right, but then again, people have been accused of writing the Marvel movies, and plot-wise they can be very hard to tell apart. Nevertheless, I haven't heard of anyone being accused of copyright infringement for making another film that reuses a plot with different comic book characters.
I just had to come in and say that your joke did not go unnoticed, nor properly appreciated.
But more to the point, this argument would literally make every version of Journey to the West illegal. Everything from Dragon Ball to The Miraculous Ladybug would be sued to oblivion for copying the concepts.
Edit: autocorrect weirdness
That was kind of my point. You can make two very similar movies without violating copyright, as long as they are different movies somehow.
Here's an even better example: Ian Fleming's novel "Thunderball" was adapted into a movie in 1965 starring Sean Connery. In 1983 a different movie studio adapted the same novel into "Never Say Never Again" also starring Sean Connery as James Bond. They are literally the exact same story with minor cosmetic differences.
https://en.wikipedia.org/wiki/Thunderball_(film)
https://en.wikipedia.org/wiki/Never_Say_Never_Again
I'm making fun of the Marvel movies, which despite different aesthetics (different characters, setting, music, etc.) the actual plot of each movie has gotten pretty repetitive.
How do most experts define "original thought?"
Yeah, to me the legal question is that is algorithmically consuming copyrighted material and then manifesting it in a new form actually breaking copyright law.
That will be hard to codify into law in a way that doesn't prevent people from doing the same.
Maybe, I don't know. But it seems to me like an easy distinction to make between LLM blackbox input-output and a person. I'm not an expert so I welcome an explanation why this isn't as simple as it seems.
Most art today is created in digital creation tools like Adobe Photoshop. These tools are gaining increasingly advanced features like object removal, background auto-fill, and most recently, full object generation. These are all "AI" features, though utilize different technologies.
Similarly, modern cameras take extreme liberties when processing sensor data. They use advanced models to identify features, and fill them in how they think they're supposed to look, to enhance the picture quality beyond what was captured.
LLMs, GANs, and diffusion models are the latest technologies and rightfully raise a lot of interesting questions of copyright and authorship. However, the line between human and computer-generated content has been blurred for a long time already, and it's not likely the genie will go back into its bottle.
As far as I know, this is more a feature of the cellphone camera segment of the market. My 5DIV, C100, and R50 have none of those 'features' and I'd be annoyed if they did.
Including Sony's in-camera authorization, which I'm sure every other manufacturer is following suit, the trend for professional cameras is proving that you really shot it and the AI stuff poses more ethical questions for professionals than the 'possibilities are endless' optimism crowd.
Hell, using the wrong LUT will adjust enough values that my organization will no longer call it a photo because they say it's technically been modified too much that it no longer represents reality.
Cameras still aren't a perfect replica of the human eye, so it's easy to "lie" about what things really looked like. This is true even without techniques like long-exposure shots. My favourite example of this is the fact that auroras to the naked human eye usually just look like glowy clouds. They're still cool, but they rarely have the vibrant hues that a camera can capture.
https://phogotraphy.com/2015/10/04/aurora-lies-cameras/
Well, metadata included on the picture tells you everything about it. "Taking in 40 seconds worth of light" isn't a lie to me compared to "I know what the Aurora looks like so I replaced this one with a better one."
Thanks for taking the time to write this. Taking Adobe from your example, shouldn't Adobe know how and what data each of their individual Photoshop AI tools use?
https://www.adobe.com/sensei/generative-ai/discover/firefly-vs-dalle.html
That sounds like a fair and responsible way to use this tech.
I'm not an expert, but it helps if you start from the basic point that these models do not retain full text, they analyze and retain their analysis. They are trained on data similarly to how we observe it. The model, to my knowledge, does not have the full text of Silverman's book in its memory. This claim by Silverman's team was disputed by the creators of the models passing hese legal tests via presented arguments.
And now we're back to what I said: the issue comes down to the creative process and human-style memory. If all human memory is simply a series of fleshy analyses of a work, how do we differentiate between that and computational analysis that leads to a similar result?
At risk of undercutting my previous comment, we can draw the line at computational "inspiration" and the claim that these materials were not created with the intent of being LLM training materials.
I'm not a philosophy major either (there seems to be a lot of things I'm not the expert of) and this probably sounds a bit ignorant - but why focus on the ambiguous end of creative process instead of the easily observable? Is the thing that's spitting the end result out a person or a machine?
The point I tried to get to is that we don't need to differentiate as much as take issue with the training itself. I'm legally allowed to read any work Silverman puts out as a human being, but to train a model is a different endeavor. You can't do much once the model has analyzed it, even purge the data (you'd need to generate a new model). You can go after the people training the model for an unauthorized use of your work, since this isn't reading, it's using it for a commercial endeavor.
You wouldn't necessarily have to generate an entirely new model to remove a bit of data. It is possible to create negative weights that cancel out the bias introduced by the target material during the initial training. Now, it might not be easy, and might end up being as computationally expensive as generating a new model. Or, it might be as simple as training a new Anti-Silverman lora. I'm not really aware of any research into negating the weights of specific training data. But, in theory at least, it should be possible to teach a model to ignore what it learned about a specific piece of training data, assuming that the data was actually a part of its training. If the data that we negate wasn't actually a part of the initial training data, then we'd just be introducing an entirely new bias into the output, sort of like asking GPT to "write me a book that sounds like the exact opposite of Sarah Silverman."
I suppose we disagree there because I see a clear difference between a person reading a lot of books and a machine with a perfect memory taking all that information in in an instant.But despite that, you helped me understand your point of view, at least a bit. Thank you.Edit. I think I've misunderstood you, too drunk. My apologies.
How involved does a human need to be for the end result to be considered their result? If I push a button and get some random art, am I responsible for the end result since the art would not have existed had I not pushed the button? If I paint all but one pixel of a piece, and ask a machine to randomly fill in the remaining pixel, am I responsible for the end result even though a machine decided the final steps for me? Where is the sweet spot where we can demarcate the point at which an end result no longer has protection? If we cannot provide one, then I suggest that we err on the side of caution, since otherwise people will invest real time and effort (however minimal or maximal that be) into creating works that are not protected and indeed considered infringing, without knowing that they are doing so.
I think if you use AI models that are infringing copyrights, your end result shouldn't be copyright protected either. At least not without an agreement/settlement with the original copyright holder(s). But that's really not my area, I'm sure people much smarter than me have some kind of precedence to cite.
I don't mind AI, it's the stealing of others' work that's the problem. If a program can generate beautiful art with you inputting one pixel there, I don't see a problem. That's probably more of a question about the license agreement with the program's developer.
Why? Isn’t “use in an ai training data set” a separate right than “use for people to read/consume”? That seems pretty different to me.
Fortunately, there's already strong precedent that non-human works can't be copyright protected in the US. There's a famous case of a photo taken by a monkey not being protected by copyright. If machine learning output can't be effectively monetized, I feel this solves a lot of issues.
But you can't ensure that if you use copyrighted material as input, in fact you're increasing the probability that it will reproduce it. Plenty of examples where image generators reproduced an entire signature of the original author. Such large fragments cannot be considered "style".
If you don't use such material as input, any output that contains a fragment of copyrighted material is purely coincidental and highly improbable.
Well it's not, because you're a person and own your own creative output. AI is a tool that 'reads' the book and someone else prompts it to produce the output. If the two situations were equivalent the AI would own the rights to the art it generates, which it obviously isn't capable of.
From the article:
…
…
I feel like chasing copyright is too hard of a battle to win.. it's too fungible because it's almost impossible for it to be logically consistent across different works...I recall reading an article about that a ways back. And the end result of that will be "largest army of lawyers wins".
But I feel like pushing for a law on the basis of control over one's identity might stand a better chance. Sarah Silverman wouldn't be able to be impersonated if her works were not consumed by the LLM. So a law that mandates that LLM's must exclude the material of anyone whom requests it from the training data, regardless if they own the direct copyright. Perhaps with an exception for data that is legally required to be public. Maybe bake it into an extension of GDPR.
The tech companies would argue this is impossible, but I would question whether it is actually impossible to comply with or merely expensive to comply with. Because those are very different things. We really shouldn't allow "but it would really be expensive to follow the law" to be a valid defense for violating it.
Otherwise copyright law would instantly lose any and all teeth, at least for consumers.
Feels like in a lot of cases, where nations rule against ai's use under this or that circumstance (or perhaps unions get agreements trying to protect their jobs) they're handing a huge competitive advantage to other nations.
EG, if animation can be made 80% more cost effective using ai, there are 3 possible outcomes:
That's just a single example, using the US (who I assume are the current world leaders in animation), but the same applies to many industries I'd imagine.
Kant says this is a bullshit argument. (Moral argument, that is; as a game-theoretic argument, it seems fine.)
I mean, slavery is a pretty big competitive advantage too. There's a reason that few companies can honestly say "we know with reasonable certainty there's no slavery in our supply chain."
Game theory is a shit basis for law and justice.
A crucial difference there is that slavery is immoral, whereas AI is amoral. I don't see that the AI revolution is meaningfully different from any other technological revolution ethically speaking.
Is it amoral if it's undermining the labor of artists? There's always a capacity for audiences to support human-made art and punish studios that rely heavily on AI art. Literally nobody cares about the price of a movie ticket, so they won't be passing the savings onto consumers.
Was the loom not amoral when it undermined the labor of seamsters (?) and seamstresses? Was CGI not amoral when it undermined the labor of traditional practical effects teams?
I'm not trying to say what should happen, I'm saying that I think the Luddites failure was inevitable, and I think the situations are almost perfectly analogous.
People did make the case it was wrong to replace seamstresses, but it benefited consumers enough for society to move on. Consumers don't benefit from AI generated art. If not immoral, it can be certainly be understood to be ethically wrong. You don't need to be a Luddite to support audiences and the artistic community enforcing a norm.
In my personal use case, I'm a bit of a hoarder when it comes to wallpaper art, and I've noticed that 90% (ish) of the wallpapers I'm downloading and using these days are AI generated, whether generated by myself or by others.
Now in this case no money was changing hands in the first place, so nobody is missing out, but I think it demonstrates how consumers can benefit from AI generated art. They can have a more tailored experience, and those without innate ability can become - to some extent - creators.
Actually, did it benefit society now, hundreds if years on?
Clothes are now substantially less durable and repairable. The neverending march of mechanization has clothed us all in cheap, disposable garbage. How much energy is wasting growing, harvesting, and manufacturing cheap cotton/poly blends that fall apart inside of 2 years, and can't be recycled or reused?
Back in the days of seamstresses, you might have only had a handful of shits as an adult. But between the durability and repair, you'd have those shirts almost your entire life.
I could see a reasonable case that sewing machines that enabled seamstresses to work faster/better on their own terms were a good thing. I think factory-style clothing manufacturing is less of a win.
"... you might have only had a handful of shits as an adult", love that typo, gave me a good giggle.
I live dangerously by having neither autocorrect nor spellcheck on my phone.
Now, to take the case of my career, software engineering: there are many tools that have sped up development along the way, and potentially reduced the number of developers required, or have enabled greater productivity from those same developers. One such tool is now AI.
I wonder if you consider it ethically wrong to use AI in this industry as well? If not, I'm curious as to the distinction in your eyes.
(Two separate points, so in separate posts as I think they're easier to respond to this way, should further discussion be desired anyway.)
I don't disagree that AI itself is amoral. Like a hammer, or a gun, or a nuke, it's only as moral as the weilder. I think I'd be a pretty sweet weilder of justice if I could shoot whatever/whomever I wanted consequence-free, but I doubt most people would trust my word on that one....with good reason.
Luddites weren't smashing machines because they were anti-technology. They were anti-worker-exploitation. And as far as I can see, most non-trivial uses of it are primarily aiming at exploiting humans, especially if they have labor protections. If it wasn't, I don't think Hollywood have been so stubborn about the writer's strike.
Interesting. I went and read the top link on wikipedia about luddites and you're not wrong about that.
I'm certainly all for labor rights, and not opposed to strikes etc.. That said, restriction of a useful tool seems like an unworkable approach to protecting worker rights.
I think it'd be strange if the work of 10 people could be achieved by 1 person, and we chose to continue having 10 people do it anyway. That would be a disservice to everyone involved. And I really don't think it's feasible. The cost savings are potentially so extreme that someone will fill that gap.
The problem is that said 1 person is going to be paid roughly the same now doing 10x the work. And that 9 other people now need to find a new job. In a world that provides no assistance to retrain and punishes you for transitioning out of your expertise.
Maybe we should be focusing on making sure labor-saving tools result in hours-reductions rather than hiring reductions.
Couldn't agree more on that!
That's where I'd want to focus efforts: societal shift. As opposed to keeping people in - what becomes in essence - pointless labor.
Artifacts have politics.
I don't think the economics checks out here (audiences are limited in what they want to consume and what they time-wise can consume)
The product of each country's animation industry can't be directly compared, since there's an interest and value placed on the country it's coming from by audiences and the community. The US is also only competing with a handful of Anglophone countries that it could relatively easily coordinate with in terms of regulation.
As to point 1:
Studios/nations are in competition with one-another. If Netflix can produce 10 times the content at the same cost and quality as Amazon Prime, they're inevitably going to capture more market-share I'd imagine.
Point 2:
You might be right that internally regulation within the Anglosphere could be possible. Maybe.
Aside:
I remain unconvinced that legislating specifically to reduce worker productivity is intrinsically a good idea. It appears to me preferable to find an alternative approach that embraces AI whilst considering existing industry professionals during the transitional phase (should the change prove so dramatic).
In point 1, my point is that there cannot be sufficiently greater overall output to maintain the number of jobs, since consumption cannot keep up with the efficiency (you would need a
5x increaseroughly 2x increase in supply/consumption to compensate for an 80% increase in labor-efficiency). Also, the market for media works dramatically differently from, say, commodities. Brand loyalty/perception, goodwill, and a bunch of other woo-wooey stuff matter a lot (these factors can also be made to matter more under cultural enforcement).My view, generally, is that the law and regulations should be set up to assist in cultural enforcement of norms around intellectual property, value of artistic labor, etc. (ex. disclosures of to what extent a product is AI-generated).
Or lower-cost more-niche media can reach fewer consumers and still be considered successful. People's needs can be more specifically catered for.
The point still stands that output cannot increase sufficiently to maintain the number of jobs.
In other words, downloading a number of copyright books from books3 is still clearly illegal.
OpenAI & Meta are still accused of knowingly infringing copyright.
If only they had legally purchased digital copies, then they apparently wouldn't have broken any laws.
Napkin math time, because I don't think it will be hard to prove intentional infringement in this case.
Call
books3150,000 illicit books (to give some leeway for public domain works like shakespear). Since max punitive damages are per-work, square that: $22.5 Billion dollars.Now, that's a per-offense, per-work maximum. They've narrowed out that using the model is not infringement, but how about training it?
I think it would be reasonable, based on the logic that the various copyright holders use to punish users, to count each full read of books3 as an offense. And that number is likely in the multiple billions, given parallel processing, but we'll just call it 1 billion. We now have a total payout of 225,000 trillion dollars.
A $15,000 annual UBI for the world (10 billion people) sits around 150 trillion dollars annually.
Turns out AI creators have worked themselves into a legal hole, where a jury would be legaly in the right to award authors in books3 enough to fund a worldwide UBI of $15,000 for...150,000 years.
I guess AI really could set us free.
You might think this reasonable, but the US courts likely would not.
Copyright protects against copying. It doesn't protect against the one copy being read and reread a thousand times. Think of libraries, which offer both physical and digital books.
The US law system calculates copyright harm based off lost profits/ damages. The way they do this is kind of bullshit, but not the way you articulated it. They assume that every download would have been a legitimate purchase, and so calculate the list price times the number of downloads.
In this case, either courts would calculate how much FB would have to spend to purchase a digital copy at list for every copy they made, or if there were no digital copies for purchase, it would have the trickier job of figuring out how much it would cost FB to purchase distribution rights.
Also, the statutory damages capped at $150k mean plaintiffs have to prove intent, which seems reasonable, but is still pending.
You're missing a thing from the previous subsection of statuatory damages. Although I did as well. I quoted directly from the code. The previous subsection:
So to recap:
I would say proving intent would be pretty easy in this case, but IANAL. Point at books3, question company: did you or did you not use this for the sole purpose of making this product and then profiting off of its use?
Without this illegal source, would the LLM be half as functional as it is? The multi-billion-dollar question.
We decided that results of LLM is not copyrightable. So, if LLM produces image of that looks like Mickey Mouse I`m free to use it?
Honestly I think we should not focus on question "is results of LLM are copyrightable, or not?". We should focus on question: does this LLM was trained on materials from authors that gives consent to such training, or not?
The reason this gets debated is that there aren't any laws about AI training material. It's a new issue that's come up. (For people, you can use whatever training material that you have access to.)
People try to come up with a legal arguments why it's not allowed, based on existing laws, none of which are a perfect fit. Also, companies use technical measures and their terms of use agreements to restrict it.
I thought that was the whole issue when I first read about this case in July.
I'm rather confident that is not how it works. If you share images containing a character with the likeness of a trademarked character, in this case Mickey Mouse, you are potentially violating Disney's trademark.