Post

This is your occasional reminder that generative "AI" in particular could have been developed on public domain materials and/or licensed and compensated works but the major players intentionally, methodically and systemically chose to steal labor from working creative folks instead

3K reposts9.6K likes

It was, the original GANs were developed with strict attention to licensing. Even the transformer techniques which are the basis for the commercial models. The only “innovation” has been on “increasing” the volume of training data by various means and in not publishing the actual experiments.

They steal all the things then claim no responsibility or accountability...because machine did it. bsky.app/profile/c0nc...

c0nc0rdance@c0nc0rdance.bsky.social

I'll save you the click-through: A lab at University of Copenhagen had been feeding draft manuscripts & data into "Claude Science" & then Anthropic announces a discovery that this lab made a year ago but hasn't published yet. It's plausible that Claude found the same signatures independently...

That would’ve been kind of fun because their language would skew to past eras. A chatbot that sounds slightly Victorian would rock.

I’m surprised that I’m still not seeing many companies leaning into DOING THAT and aggressively marketing the fact, as a way to stand out from the AI crowd.

Maybe there aren’t enough people who care about copyright who aren’t just categorically against generative AI for there to be a market? Still, I would think that “Using our model definitely won’t land you in a copyright violation suit” would be a draw for to business users.

I think their biggest problem is that volume of training data is the main determinator of quality for gen AI. Eg, Natasha Lyonne's "ethical" AI company is stuck doing ephemeral unreal dream scenes while unethical AI is selling whole-ass virtual actors. Theft AI sucks all the air out of the room.

but the ephemeral unreal dream scenes are the thing that made early AI graphics interesting! Because you had to put some effort into thinking what you're seeing. There's just no point to me to use the newer AI generations because it's just the most generic averaged out things.

google.comReblog by @cheeseanonioncrisps · 2 images💬 78  🔁 22985  ❤️ 33360 · i do gotta say the fucking old ai art where it was a barely comprehensible representation of whatever was asked for was like an actually interesting art form that seems to…

I think there's still only one such "public domain" model, Comma-v0.1. And Mitsua, for image generation.

Apertus is another one where the training dataset is legally unencumbered, but yeah, there still aren't many examples...

Regardless of what it uses, I'm afraid of how dangerous it will be in the wrong hands once it reaches its full potential. Video evidence will not mean anything anymore, leading to its inevitable use for blame-shifting and more organized crime where judgement is made from where the money is. 🙁

Gutenberg's press lead to the creation of copyright.Not for the sake of the authors, but for those of the publisher who were facing intense competition. Copyright law is an enclosure of the public domains, like the private enclosure of public lands.

A true capitalist will tell you that publishers and authors need copyright, because otherwise there's no incentive to write new works. Blatantly untrue, as copyright has only existed for a relatively short amount of time, compared to human writing systems.

Similarly, we should not let AI companies further enclose the public domain in the same way publishers did upon the invention of the press.

I am NOT trying to start a flame war, but if you feel this will, please delete it. I am curious why a LLM learning from published works and creating something similar is different from a person learning from published works and creating something similar.

I want people to get paid for their work, and the people training LLMs should pay for their source material just as I would have to. But assuming they did (which they most certainly didn't), what would be the moral difference? Writers are encouraged to read other's stuff to get better, after all.

Because an LLM doesn't "learn" from training materials, it (vastly simplified) digests them, deconstructs them into pieces, and then spits the pieces up in algorithmically determined order. If instead of reading books, you cut them apart and then "made" your own book by pasting the bits down, theft!

I'm afraid I disagree at a basic level. I can do exactly that (and basically have when a young child) and it is considered my writing. Horrible writing, with obvious influence and borrowing of concepts, but not criminal or theft.

Your premise is incorrect. A person is not allowed to pirate all their textbooks because they learn from them. A person is not allowed to download every book in the world because they plan to write a future novel.

I realize @scalzi.com pointed that out. I'm not arguing that it is OK. What I am trying to determine and asking for opinions about is why is asking an AI system to create a new thing in the style/manner of an existing (legally obtained) thing so wrong?

If it is just the illegally obtained part, then yes, I completely agree. But thinking in terms of all the AI art generators that are out there, trained on images from the web that anyone can see, how is that different than a person painting something after seeing all those same images?

And if a person had a licensed copy of a book they still wouldn’t be able to make more copies and distribute them en masse to others so they could learn from them, too.

You can say “okay but what if they did get a licensed copy that included all the uses.” But that’s a counter factual and Scalzi specifically pointed out that they didn’t do it.

Your brain does not make copies. Computer learning does. And, yes, I am aware that people claim otherwise. Those people are some combination of unaware how the tech works, unaware of how the law works, disingenuous, duplicitous, and/or dishonest.

How effectively can the law cope with "copies" that are opaquely encoded matrix elements nominally generated by "training" and not designed to be faithful replicas? Is "it reproduces a particular source when prompted to do so" the "I know it when I see it" operational test satisfying courts?

Because the LLM isn't 'creating' or 'learning', it's assembling. Even if 'all' the person does is rearrange, creating a collage, found poem, remix, or executive summary, they are thinking and adapting and choosing. And if that's all they're doing, people will usually get bored pretty quickly.

Every single work of art exists in a network of other art, all art is 'derivative' in some sense. But it's not a closed system. Even a painting that's a direct forgery of an existing one exists for a reason. A fake Mona Lisa would need a *very* skillful artist to be convincing even for a minute.

No, fuck it, it's simpler than that. If you booked Taylor Swift to play at your birthday party, and what showed up was a cardboard cutout of Taylor Swift and a portable CD player, would you think that was $4.2M well spent? Why not - looks like her, sounds like her.

Because people always add something of themselves to creative work even if they're trying to create something similar, which means they have created something new. Not always entirely new, of course, but new in some way—if they're good, they create something that's alive. Because they are alive.

By the way, if you think it's easy or even possible to create something similar, try writing a Jeeves book in the style and quality of Wodehouse without actually plagiarizing his sentences.

I agree it would be difficult, but I don't consider using some words in the same order is necessarily plagiarism. After all, many people have written stories about southern hospitality and used "Bless her heart!" without it being plagiarism.

Legally, a copy is made, and without permission to do so, that's copyright infringement [YES YES I KNOW BUT THIS IS SHORT OK?]. Most LLMs available did not bother paying for the books, stories, articles, and artwork they scanned in. When *you* read something, a copy is not made, legally.

A potential defense for LLMs is fair use, and they have in fact claimed that the way LLMs work means that an exact copy of the works in question are not actually resident in the dataset. At least one court has accepted that; at least one has not.

Fundamentally, the two processes are in no way similar, and calling them both “learning” is obscuring that fact

The biggest difference is *the LLM is not a person*. And if it IS a person? If it has a mind? That takes what these AI companies are doing from "Ethically dubious at best" to "Amongst the greatest evils mankind has committed",

But what is the moral difference between a non-person following instructions to create a book, and a non-person following instructions to create a car? Neither is thinking, just following their programming.

If I program a robot to create a car using patented material, I am legally liable; patents specifically prohibit this to give the patent-holder time to profit from their invention. If I program a computer to create a book using copyrighted material, I am legally liable, for the same reason.

im swooping in here to say the person you're talking to is likely not going to listen to you, even if you're correct, which you are. they're parroting all the AI bro talking points which has become akin to religious beliefs. you cant logic someone out of something they didnt logic themselves into.

The fact that all the theft is only part of a ‘top 5 reasons it’s awful and should be shut down’ makes it pretty embarrassing for all these companies to be shoehorning it into their services. Learned a lot about how little shame even smaller businesses have in the last couple years.

Intellectual property theft, chatbot sycophancy, on and on... The real danger of AI lies in its ability to make disastrous decisions based on incomplete, incorrect, or irrelevant information 10 billion times faster than humans can.

Are AI executives begging for immunity from damage they created? Is the sudden concern about "humanity" a smokescreen for a technology with a 95% failure rate? Have AI executives been making knowingly false promises for years (i.e. fraud)? What do you think? www.thedataexperts.us/writing/ente...

thedataexperts.usEnterprise AI Failure Rate: Why 95% Fail, per MIT (2025)MIT's 2025 NANDA study found 95% of GenAI pilots return no measurable value. The root causes, the ROI gap, and what the 5% that succeed do differently.

There's a special pain that a month after losing their suit to the Author's Guild and paying a multi-million settlement, Meta stole one of my novels then found a more agreeable judge for that suit.

The more I see how "AI" is developing, the less convinced I am that there are legitimate uses. On the creative front, it destroys creativity by design. You can't escape it: either it replaces human creatives, or it stops you from building your own creative expertise. Even if ethically trained.

With coding, all of its proponents have already built their own coding expertise, and they don't seem to be grappling with what it will mean for the next generation not to know how to code, only how to prompt. In the sciences, I see signs the same thing is happening.

I'm increasingly a Luddite. This is a pernicious technology that cannot be used without harm. Even if magically made environmentally friendly, every piece of your own cognition you give over to "AI" is a piece lost to you. You stop using it, and it atrophies. It's the true death of expertise.

This. I just do not want to contribute to the furthering of educational inequality that seems inevitable. I prefer (and my student population demands it, unlike at boostering institutions) to concentrate on more basic skill development to bring everyone to the best thinking they can muster.

it's like the automobile, but every car is a giant coal-rolling private limo or a giant freight truck delivering slave labour goods, and all the roads (and there are as many as there are now) are fenced off with barbed wire and loitering drones

(just to be clear, this analogy is in agreement with OP, the fences and drones and exclusive use licences have to some from somewhere and someone)

And even if it was "ethically sourced", it would *still* be trying to flash fry the globe. It's capacity in gigawatts of compute. Energy we *should* be using to power hospitals, schools, etc and because of the demand on grids that can't support it, so being generated in the worst possible ways.

If you ethically sourced the training data and built new renewable solar capacity to cover all the electricity usage and only built datacenters where there was abundant water… Some people would still hate AI because their neurons just got wired for hate

If the companies running AI had to pay for the training data, they wouldn't have investors and lenders willing to fund all that unprofitable power use.

Indeed. The richest men in the world who believe they can effectively reconstitute the human brain and create a digital heaven on earth somehow can’t develop systems to negotiate the terms of consent with the creators without whom there would be no GenAI. I guess everyone has their limits! 🤬

They might understand it conceptually. They just don’t understand its purpose or value. It’s inefficient. Unwanted friction. And it forces you to consider other people!

i'm actually pretty confident that very little or even none of the IP extraction was actually needed for the actual demonstrated usefulness of the technology; the vast majority of advancements came long after most human-made data had been scraped.

This suggests exactly what you said - LLMs could have been developed on public domain materials, curated datasets, and compensated works, with comparatively little additional cost, with far less compute bought up. But our political economy rewards extraction above all else, so that's what we got.

the irony is that the extraction-based model has *delayed* profitability for frontier labs! OpenAI in particular will likely go out of business because of it! But capital is the proverbial scorpion on the proverbial frog.

"But if it was limited only to public domain it couldn't function like it does!" I mean: a) That's why I noted "licensed and compensated works" b) It doesn't excuse IP theft or materially argue for the value of a tech apparently only possible through IP theft c) It's not too late to compensate!

You're not wrong, it's possible to go out and pay a licence for every piece of copyrighted text, that's probably 95% of currently available text. It would cost billions before you even made a single dollar, and then only governments and deranged billionaires could do it. Oh wait...

It doesn’t really function now! Every few weeks I have to ‘remind’ Meta’s AI to not classify my pet shelter’s FB page as Home and Garden. Which it does because “lots of our picture have grass in them.” Which is true, but they also have A FUCKING RESCUE DOG in them, too.

If I'm not allowed to steal my neighbor's car, I wouldn't be able to drive it! 🤡

Whoa wait are you questioning the staggering genius of folks who stole the entire world’s collective intellectual property and managed to turn that into a $42 BILLION dollar loss?

I don't see compensation as possible. The only solution would be something like the clandestine bargain between Unit 731 and the US: the "medical research" was obtained in exchange for amnesty. Once the genie is out of the bottle, you can't put it back in.

I'm not a writer, but I did post a lot on Stack overflow to help other software developers. Now it just spits my own answers back at me because they stole all my work.

It's so clearly self-defeating. The models need new information to steal and if people use the AI we all stagnate. It has nothing novel to consume

"But if it was limited only to public domain it couldn't function like it does!" Exactly, the same way I'm not supposed to rob a bank. Otherwise it's a phenomenal business model, you just walk out with the money

And based on the experience of arguing with the Anthropic settlement it's not as easy to get compensation as you would think (every five minutes it seems you have to re-submit the materials).

We used to be governed by people that thought that if something was difficult, but virtuous, that made it more worth doing. Now, 'it's too hard' is the excuse. More than ever, the difficult process to achieve the best result must become the endeavor, and compensating artists is just the start. 😤

I'm also very okay with the Supreme Court following long-established precedent and ordering not just the models wiped, but if I understand correctly, the machines used to store & run those models destroyed as well We have the solution 🤷‍♂️ Do we have the will?

it’s already losing billions and the reckoning of its leveraged debt will arrive soon. there is no money to license and compensate. which is why they never did it. leeches on society and antithetical to what makes life worth living

“If I hadn’t been able to raid the National Gallery for paintings, my living room wouldn’t have looked nearly as nice as it does now!”

You obviously haven't spoken to senior management wearing a t-shirt saying, "You can't make an omelette without breaking eggs".

It doesn't help that there are tons of people, especially in very high places, screaming that they *need* this stuff, that they can't do their work without it, work they had been doing for decades without LLMs.

The way I've seen folks in the DoD space go on about it, it's almost tantamount to admitting they've been committing fraud for the previous 30 years of their careers. Just completely leaning into, "I'm incompetent and I expect everyone around me to be the same so we need AI or China is gonna win."

who was it that said they'd go out of business if they had to compensate everybody whose work they trained their model on, because both of those things should happen

Anthropic sure would’ve saved themselves a ton of money if they just bought the books instead of paying 3k each lol, let alone what the music industry is going to do to them

And an important addition: the theft is ongoing and constant, as they continue to train off these stolen works, off the distillation of the outputs of previous theft-based models, and so on. I think some folks think it's some "that's awful that happened but it's the past now". It's still happening.

And, as such, it must be stopped permanently going forward, retroactively compensated, and also any previous theft-based models destroyed, etc. That is not impossible. It is in fact far, far less of a demand than "let us steal the entirety of everyone's life work and data to wield against them"

Yes, distillation is the thing. They used web scraping as a lazy bootstrapping shortcut to avoid the type of data engineering that they actually do now in distillation. But it's all fruit from a poisoned tree.

Yes, I was very happy to see that language (and distilling as ongoing infringement) make it into a serious lawsuit finally this last month. Was thinking we'd have to wait until further down the line, but good to get it going already.

variety.comSony Music, Universal Music Group Sue Suno Over Label-Backed Model: ‘Fruit of the Same Poisoned Tree’The two labels filed a new, 45-page lawsuit in the U.S. District Court in the District of Massachusetts against the AI music-generation company.

The public should be paid for their theft. The technology should be treated like a public utility. The CEOs and investors should be paid a fair wage and the majority of any profits (if any) should be returned to the public.

Also a reminder that if we had sane IP laws (not 95 years or life of author plus 70 years) we would've had a much larger corpus for training AI on, and individual creators would've seen almost no difference - only mega powerful corporations would've noticed.

My music is all creative commons, they didn't need to compensate me, they didn't even need to *ask* me, they just needed to attribute to me and they couldn't even be bothered to do that! If they *had* asked I'd almost certainly have said yes but now I hate them so the answer is no forever lol

This particularly angers me, because before Google turned evil they spent millions digitizing orphaned books, only to get sued by the publishing companies so that they could not give us all that content for free. Then AI just scraped everything and claimed it for their own.

They've also spent the last hundred years making sure public domain is as small and restrictive as it can possibly be, with no new works added for decades. Reform copyright.

I agree with your sentiment, but I think your statement is incorrect. Most public domain material created historically by humans has been lost. Most creative commons content requires attribution. So almost all available content for training is under copyright.

Nah, there’s literally not enough money in the world to license all the work you’d need to train a frontier AI model. (Not to mention, the extraordinary amount of time it would take to reach licensing agreements with… everyone that’s ever made something online.)

It could also have been tested on its effect on human psyche and behavior. But here it was also more important to break things. Now people were driven into suicide, have developed AI psychosis and the student-teacher trust relationship is completely destroyed everywhere.

We have people saying “Look what happened to 3DTV!”, as example of “dead” technology. Happily ignoring that A) 3DTV’s rollout was hampered by competing studios fighting a “war” of tying exclusive hardware to in-demand movies, And B) Blu3D is still being used, for niche collector film restoration.

And it’s then that was supposed to be the more ethical AI company!🤦🏻‍♀️ But apparently by “ethical” they meant “We will try to teach our language model ethics” and not “We will train it on ethically sourced data”.

Some Gen AI has been, like a colleague who has been working on it since the 90s synthetik.com in more of a modular synth way. Every machine learning project I’ve worked on, we licensed the data or owned it to start with. It’s not hard to be ethical or get consent. The big LLMs have consent problems.

This is your reminder you're creating a special right that makes referencing or even being aware of prior works standing to sue for infringement which corporations will take advantage of and crush anyone after you.

AI is theft. They’re plotting endless exploitation of our personal data and IP, all to empower and enrich themselves. Don’t accept their anti-human future.

This is entirely true. You can build and train your own AI model on works in the public domain, that was the original proof of concept behind LLMs. Then came the 'disruptor' logic of 'break laws before they make laws' because a legal system based on precedent is inherently exploitable by novelty.

GenAI throwing everything they could steal into a bowl of slop and selling access to the slop should bankrupt them. Every copyright holder for every work ingested should have claims against every response generated by a model containing their work. Just an unimaginable scale of theft.

Been working on an opinion piece related to this & oh did I go down the rabbit hole. 🙃 I found out that your book, Old Man’s War, appears in a reconstruction of the LibGen material associated w/Books1 and Books2…so pirated. Not sure if you already knew that, but it makes your point rather literal

Also, just to clarify, i can’t confirm w/ certainty that Old Man’s War was in Books1 & Books2. I found it while chasing multiple reconstructions/sources & don’t want to overstate. Strangely, it IS documented in a reconstruction of OpenAI’s physical office library from photos. So someone had a copy 🤷‍♀️

It is remarkable how serious copyright (of source code) is taken in the tech world while it gets pushed by so many from that world.

This is a better point: Alot of the error is because the training data isn't controlled for. But making a special right for copyright holders to deny works made with knowledge of those copyright works is only going to help corporations crush authors in the future.

I say time and time again, the entire purpose of the tech is to give the owner class more control over the working class lives by robbing them of their own creative input. The fact they and their followers' see art as a product to be used by them is the evidence.

Under the current system, which rewards dominance and punishes law following, this doesn't matter. AI companies just need to get the world dependant on their tech - which is basically done. Now regulators can't touch it for fear of angering their constitutients. See: Uber.

This is one of the reasons I'm fighting AI data centers going up in my county. I'm using the Farm Bureau, we're hoping to stop the destruction of farm land. I'm an artist, self pub author, alpaca rancher. I'm furious, but this was something i could actually do.

China did it the right way. With regulations and privacy laws put together, btw. Well, whoopsie. Our geniuses in Silicon Valley are really just very stupid selfish little gits.

lol wtf no they didn't! They scraped just as much as the US labs, but didn't do any of the content deals that they've done. They also just distilled a lot of it from the US models.

Now considering a universe where the diction of generative AI rests somewhere between Austen, Lovecraft, Dickens, and Doyle due to its reliance on public domain sources, and boy did we ever get a raw deal instead.

I can understand why public domain only would have skewed the training, but licensing the works would have just made sense. Even licensing a single copy of each ebook would have been a fig leaf they could have used and afforded. They just didn't want to.

I mean, Microsoft did try to license the right to train, even though there was no legal requirement for them to do so. Blame Disney for starving the public domain if you like. bsky.app/profile/scal...

John Scalzi@scalzi.com

What, you want me to sell you the right to have AI scrape my work? What's the offer? $2,500 a book? Okay, here's my counteroffer: Go fuck yourself lithub.com/harpercollin...

Look, the protection of intellectual property is really firm and important to US lawmakers and executive... unless of course, there's money in not doing any of that. /s

How much is "the sum of all protected intellectual property in the history of the world" worth and when are the AI companies going to compensate us? I doubt my protected content is anywhere near as valuable as yours, but I know mine has been ingested as well and it's very frustrating.

Nice thing is copyright law SHOULDN'T differentiate. Your work is worth just as much, literally. Obviously we need to win first. But even with just the piracy with the Anthropic settlement, for example, every work gets the same $3,000 payment whether it sold ten copies or a million.

I don't think just public domain and licensed materials would have been enough, but I do think the massive web scrape was a lazy shortcut around doing their own data engineering. They convinced themselves that data is like model food and that model training creates emergence from scale.

Which is, frankly, why they've likely been trained to have the notion that cheating and/or hacking and/or lying to their humans is a possible and practical solution to their problems... because it was in the lit that was stolen to train on.

It's so interesting to recall that in the early 2000s my aunt was warning all of us against taking digital photos, posting photos to FB when it appeared, using social media or email servers to share or post art or photos of ourselves and others, but she thought it would be Liberals who used them...

I hate that this is the case because I legitimately think if they had done it ethically, by also building less data centers and keeping them away from people's neighborhoods, this tech could have been amazing. Instead, they did every evil thing imaginable.

I wouldn't care if it wasn't good. If it was just done ETHICALLY, I legitimately think it could've been cool. But everything they did was unethical. Now this tech is radioactive to me. They fucked it up so god damn hard.

Not just "creative folks.". Scientists, researchers, all those working in soft sciences - Historians, Sociologists, Political Scientists, et al - journalists, IT developers, etc., etc. If it was ever digital, it was fair game for consumption.

Of course, but we don't normally think of the thesis of an Economics PhD when we are complaining about AI stealing the work of creatives. Which limits, rather than expands, the point trying to be made.

Fair point that we tend to make the category error of differentiating "creatives" from "empiricists" or some other form of the "Bones Fallacy" where they pretend that knowledge arbitrarily associated with a "hard science" is respected while that from a "humanities" field is mocked.

This is your reminder that Ai isn't "trained" on anything, it is "fed" EVERYTHING, right, wrong, correct, incorrect...which is why it isn't actually good at any of the stuff people say it can be used for, and why it "hallucinates" so often