The data that powers AI is disappearing fast
nytimes.com
nytimes.com
Currently, I think most of the training use cases can be covered by the existing "you can't copyright a fact" carve out in the law. That's probably better for society and creators than my licensing regime scenario.
Anyway, I'm rooting for "no regulation" for now. The whole industry is still being screwed over by market distortions created by the DMCA, and this could easily be 10x worse.
Sad that you didn't consider the content creators, people who's faces, voices, writings, art, personal information, etc. are being used without consent and without any compensation as the ones "being screwed" here.
If the pattern holds, the "cost" of training non-paywalled data will be attempts to hack/influence the model, as opposed to hacking humans.
People in the AI (using this term very loosely here) business are worried that only the biggest players effectively monopolize AI with the help of badly thought out (or plain corrupt) regulations that may be set in place in the near future.
These two issues are not mutually exclusive problems, these two groups are not fighting against eachother. It feels like a typical divide and conquer, where the regular people are being pitted against eachother because of your view, while they should work together towards a solution for both of their problems and fight bad regulation (and monopolies in AI).
Generative "AI" enthusiasts fear that bigger fish will squeeze them out of the niches they created for themselves through theft.
Suggesting that artists and genAI enthusiasts should "work together" to defend the regulatory environment of the latter is ludicrous; it would be like a burglar breaking into my house, stealing my appliances, and asking for my support in his campaign for mayor!
gatekeeping the information that is being learnt off these data is not a right they have been granted.
I do not see an ai weight/model as a derivative works of the individual works used to train it, provided that you can obtain new works that would not have violated the copyright of the original had it been manually made.
Why is this the worst possible outcome? Companies using AI would be training with properties they own or have licensed appropriately, rather than the existing scheme of ignoring copyright law to extract $$$ from the creative works of ordinary people.
Most who earn their income from IP are society’s elites. Perhaps the lowest-paid least elite IP profession I can think of is journalists, and I’m pretty sure they are doing works for hire and not owning the copyright.
Arguably when it comes to the IP questions, ordinary people are far more likely to be benefiting from AI (for instance getting it to make art they couldn’t express because of lack of talent) than they are to be be robbed of their IP in a way that matters.
Obviously the question of AI harming ordinary people in terms of automating mundane jobs is a very real one, but interestingly it’s totally irrelevant to the IP issues.
That is one thing I can't get. Why is everyone not seeing the elephant in the room? The user and the user prompt - they add new intention and purpose to the material being referenced. Not only that, but a LLM response is usually shorter than a full length article or book. On top of that, LLMs retrieve from multiple sources when they use search, and respond from even more sources when they generate answers in closed-book mode. They are clearly doing synthesis.
LLMs are not even good tools for infringement. You have to work really hard to put the LLM into an infringing mood, and it requires snippets from the target content, and only works like 1% of the time. If you have snippets, you probably have the whole thing already. Copying is so much easier and faster. LLMs would just reproduce approximatively a desired content. What a LLM gets from a source text in closed-book mode is on the same level with a image thumbnail from a full resolution image.
What LLMs do well is to recombine ideas in new ways, as demanded by the prompt. Should that be infringing? If the answer is yes, then humans should also be barred from recombining ideas. After all, any human could secretly be using LLMs. But that would just tank creativity for fear of litigation. Isn't it strange how copyright promotes creativity by restricting it?
Even the name "copyright" refers to copying, as it was initially envisioned. Now authors want to expand it to idea ownership. I think it's a power grab.
That’s the big question I was alluding to. Does the prompt provide sufficient novelty to avoid claims of copyright? Or is it too derivative? I don’t have a strong opinion but can see both sides and suspect eventually it will be decided by the courts.
The distinction you may be missing is that copyright refers to licensed reproduction. IP is, after all, about property/ownership rights.
Under your moral logic, everyone should just be free to pirate/steal any creative work, as why shouldn't the un-talented (or un-trained, or un-dedicated) have equal rights to the works.
All this leads to a case where there is no longer an incentive to create and popularize creative work, and suddenly all that is available are AI rehashes of AI summaries. I, for one, don't look forward to such a marketplace of non-ideas.
I do think I should be allowed to read your book & do math with it.
This is such a painfully obtuse construction of what is occurring that it escapes any reasonable discussion.
What is strange in ignoring the idea of copyright? If you write a song and I can play it in my computer, I use it the way I want in my computer. If you don't want that, don't make your work public. Copyright is a men's creation, nobody is forced to respect it.
The code I develop can be accessed for free in github and even in the browser "view source", I won't be fighting for other people to have the right to force others to pay for using their creation while they don't pay for mine and all other open source and open science creations.
I was afraid the discussion would go this route.
Yes, copyright is a convention. It’s subject to change. However, IP protections are written into the US Constitution and the bar to change it is relatively high.
Murder is also a crime by convention, there’s no natural law against it. But we generally recognize that to live in a stable society, we must live by certain conventions.
You can play a song on your computer because that is considered appropriate use. You selling tickets to play it, or to copy it and sell it is not because they conceivably limit the authors ability to make money from their creation.
You may not realize it but software is also covered by copyright. For most intents and purposes, it’s considered the worlds worst book; you cannot legally copy and sell it if the license doesn’t allow it.
Edit: added the word "legally" to be more precise
Well, good luck trying to force me to not sell your work in my country.
In classical copyright, fair use allows certain.. fair uses. Going beyond that is considered stealing. You can chop things up into small parts, and there are rules governing how you can put the pieces back together and claim them wholly or partly as your own work.
LLMs behave more like a solvent. Everything goes in the pot, gets melted down, and by the time it's recast you can't say for sure where anything came from or who it once belonged to. Even if sometimes you might get a strong whiff.
Clean things up just enough that it's difficult to prove where their semantic content came from.
How useful this is or isn't, and how much of a threat to rightful IP owners, depends very much on the type of IP.
Copyright laws didn't need to be invented until the printing press came along, because the act of copying was slow and difficult.
Not a fan of the patent model for software, for example, but perhaps this is an argument for it. Or, we just get used to the fact that idea-reuse is cost-free, accept a couple of decades of uncomfortable economic dislocation, and get on with it.
It's reasonable to be worried about this scenario. If there was no incentive to produce creative work, our society would be much worse.
But the notion that there's only one way to prevent this scenario, and it requires a drastic expansion of the already sweeping "intellectual property" regime...well it just lacks creativity.
It's not that we want to eliminate the incentive to be creative, it's that we believe there are better ways to prevent that scenario than to further entrench a broken system.
I'm not sure why this is a relevant point. Is IP only important if it's earning money?
I don't think that copyright restrictions really apply to the use of data to train LLMs, and that fact is why I, and a number of others, have removed our works from the public web entirely. There is no other real way to protect ourselves from people using the works in a way we object to.
As you say, removing your content from the public sphere is the only way to protect yourself from people using it in ways you object to. This would seem to be the case regardless of the state of copyright protection since people gonna do what people gonna do.
Copyright allows the copyright holder to decide what uses their work can be put to. The non-monetary risks are that the work will be used for a purpose the creator strongly objects to.
> This would seem to be the case regardless of the state of copyright protection since people gonna do what people gonna do.
Not true. Copyright provides for the possibility of legal repercussions when "people gonna do what people gonna do". Even for ordinary people.
No. It allows the copyright holder to monopolize the copying and public performance of the work. That’s it.
It’s not a copyright violation for me to take your copyrighted painting and draw on it with crayons or hang it on a wall labeled “World’s Worst Painters.” As long as the copy was sanctioned by you I can put it to any use I want as far as copyright law cares. (One popular use now is having a computer compute facts about its word patterns)
Yes, we aren't disagreeing here.
> It’s not a copyright violation for me to take your copyrighted painting and draw on it with crayons or hang it on a wall labeled “World’s Worst Painters.
Exactly correct. I wasn't talking about purely private use. Copyright is about distribution of the results. If you do something with a work and never distribute the results, copyright doesn't enter into it (ignoring the complications of the absurd anti-circumvention clause of the DMCA, anyway).
But these companies want to have their cake and eat it.
(Disney: takes fairy tales from the public domain. Elsevier: You pay them to get published)
The problem with Disney is their extremely aggressive and repeated push to safeguards their own golden goose. (And IMHO they have a serious quality problem. "Somehow Palpatine returned.")
In general the quasi-forever current copyright regime is simply too blunt, and arguably waay over the optimum point in terms of length of granted protection with regards to "incentivizing the creation of arts", and of course it's incentivizing the collection of money-making properties and flooding the market to extinguish most of anything else. (Naturally attention is a limited resource, so "arts is a zero sum game".)
Cool - you want to train on medical data? That will be $1M per paper per day to Elsevier for a licence. (Apply similar for movies, news, books, etc.) Ensuring no knowledge remains "common".
The moment they can, the big data holders will charge whatever they can get away with. For an average person that may be even worse than the current content stealing.
And, importantly if papers become so valuable ... authors will suddenly start to want their cut and will go to the publisher that pays the more. And so on.
It would be amazing if papers would worth that much (as it would mean they can help creating at least that much value downstream), and it would mean there would be a lot of money in writing new papers. (Oh imagine that flood of low quality shit! Oh no, it's almost the same as now! Ehhh.)
The logic seems to be that "some of the training input was copyright, therefore the inference output is copyright." The same logic says "some of the training input was copyleft, therefore the inference output is copyleft."
Maybe future training sets will have to worry about the license compatibility between all the data in the training set?
The problem in this case is the AI companies aren’t ignoring copyright and then releasing their output with no copyright, which would happen in a copyleft or copyrightless world.
They instead want the benefits of closing their output but they want the benefits of taking other peoples work to make that output.
The entire effort to find a problem for them to solve has meant ecological harms at a scale that makes crypto mining seem trivial.
The data required has copyright and other commercial use restrictions on it. Much of the data being overtly restricted was never available for harvest to start with.
Given these facts, bringing and end to the LLMs an snaping techbros out of their latest delusion is a good thing.
Right now all we're seeing is gentlemens' agreements and self-enforcement. A robots.txt is just asking you nicely. I don't know of any country that has any judicial ruling on the topic of AI scraping yet?
https://www.deeplearning.ai/the-batch/japan-ai-data-laws-exp...
This is a complicated topic and I don't know the right answer. I hope people with far greater skills in IP can figure out a meaningful solution that everyone doesn't hate. It might be that there is no good solution.
Elsevier on the other hand. Same deal as ever, if you want knowledge, pay for it. Writers deserve food on the table, and have the right to have control over who sees their work for economic or strategic reasons.
No one owes you anything
Public money paid the researchers, results should be free.
As for the rest, I'm in favor of being able to train AI freely on works of culture. Current copyright regime is suffocating culture already pretty badly, let's not make it worse.
Not me. I want appropriate regulation. Without any regulation, these companies will just continue their abusive practices without restraint.
Will regulation encourage a concentration in a couple of large companies? Maybe. It depends on the nature of the regulation. Even if it does, though, I'd take that over the current status quo.
If anyone would like to join in, there's an actively maintained robots.txt here:
https://github.com/ai-robots-txt/ai.robots.txt
Yes, I know this isn't legally binding and scrapers can ignore it if they want to.
Crawls each and every date link on my synology dsm hosted calendar without throttling.
I'm thinking synthetic datasets are how it's going to go. Of course you can't get information from nothing, but they might not need nearly as much seed data to generate lots of examples that are specifically designed to train reasoning skills.
Maybe it will take a while to get over the hump, but they'll be motivated to make it work.
There was also a paper on persistent computer programs after running random data iteratively over millions of cycles. This doesn't even have a 'goal'. which by itself is very interesting because of how similar it feels to how organisms develop by the act of persisting after a few billion years.
That is to say I largely agree with the premise that you can train and improve alot by absense of data. But if your original dataset is garbage, then training over and over might get limited by a hidden nash equilibrium that the algoritm can't escape due to lack of new information.
Obligatory Anton petrov: https://www.youtube.com/watch?v=L_IWVZPmc-E Paper: https://arxiv.org/pdf/2406.19108
Wouldn't it make more sense to say trying to understand general relativity from the perspective of a flat-earther? At the quantum level, the world is flat--I mean, ontologically. "Dimension" is not used to describe space and time in the same way as in Einsteinian physics which retains a very Kantian model of the cosmos. But then you could certainly re-interpret Kant to make it cohere with quantum physics, I'm thinking of people like Deleuze and Latour here, which is why I alluded to a "flat" ontology.
But in truth, it sort of negates your point, since when it comes to gathering things from the totality of the productive process (which is what an LLM does), indexing them and creating a surrealist wet-dream of output, there is not going to be a "right" answer, in just the same way that a quantum physicist and a cosmologist are not going to be able to agree on the correct definition of time. Truth is at the level of feeling, not sense. No scientific instrument can detect the feeling that one thing caused another, even though we are absolutely sure of its continuity, and it is such continuity that stands as the basis of all scientific investigation. In the end, empirical laws, just like moral laws, simply "feel right," and that's it. The LLM tells us things that feel right. It is a good-feeling machine, yet it is precisely when we notice something wrong, when it seems totally off, in those sublime gaps, that it becomes the most enigmatic and, in fact, the most useful. The real beauty of the AI is in those gaps, in its imagination, because it is actually our own, projected outward and reflected back like a society-wide mirror. There is something wonderful in the dreams of a machine.
For example, AlphaZero uses synthetic data (playing against other versions of itself) but it does gain information because the data is generated against the rules of the game.
Similarly, word2vec and other vector representations use synthetic training data, but that data is derived from observations of text (no, it's not using raw text, each training instance is generated from the text, with varying levels of simplicity of course).
Tesseract is trained using synthetic data generated from font files and simulated "rendering errors."
Synthetic data should not be underestimated.
[1] https://en.m.wikipedia.org/wiki/Data_processing_inequality
It's probably true for people too. Instead of asking an expert for an opinion right away, ask them to discuss the options out loud first.
Also, "natural" data taken from the Internet is probably quite inefficient as training material. It's going to have a lot of duplication. You only need each fact once to be able to synthesize more examples of it.
https://www.nytimes.com/2024/07/19/technology/generative-ai-...
In contrast, training an AI model you download a piece of info once and then don’t resell that (exact) piece of info. Nobody is claiming they own the image when they don’t.
An AI generated work may still be derivative enough to be judged as infringing by a court. AI is not a black box which magically performs copyright laundering.
does the CC used by wikipedia allow training?
> and no contemporary art in sight
will note per above, they may be able to scrape other CC if they -can- scrape wikipedia.
> It's mostly deleterious to the culture.
Yep and based on how I already sometimes catch people treating LLM hallucinations as fact, it's likely gonna get worse anyway.
This is backwards. Absent something that prohibits it, training is allowed. The question should be this: what do you believe prohibits it in this case?
In the specific case of Wikipedia I would guess it's allowed by the license, but that's not generally true.
https://web.archive.org/web/20240720013944if_/https://archiv...
This is how I read NYT now:
tnftp -4o"|yy093" https://www.nytimes.com/2024/07/19/technology/ai-data-restrictions.html > 1.htm
links 1.htm
yy093 outputs a single HTML page with all the fulltext articles, the way I like it. 100% Javascript-free.
The web keeps getting better for text-only. More JSON, less HTML.
edit: okay, it seems those are scripts written by you, right? https://news.ycombinator.com/item?id=41035155
okay, WTF is this https://news.ycombinator.com/item?id=40894851 ? :D
That is such a weird and misleading way to put it. There was no consent in the first place. Take YouTube for example. Google did not consent to the videos it hosts being used by OpenAI. The uploader certainly did not ever consent to their face, voice and content being used to train models either.
That being said - the general crux of the third-party doctrine (at least in the United States) is that information told to another party has no expectation of privacy. That seems to apply to a vast amount of 'user generated content' on the internet. Someone decided to utilize someone else's megaphone to post into the commons and now expects to retain some ownership. It's hard to have it both ways "just because someone else noticed."
Unfortunately this also extends to general artistic styles. If you, as an artist, continues to create in a consistent style and someone else notices you don't have a lot of recourse when some entity is able to recreate a similar style.
On a technical side this seems similar to your server responding to any incoming request with a 200 status and some content (LinkedIn v HiQ). In this case a user-agent asked (was it a person? scraper? AI training process?) and you gave it to "them".
I guess we end up in the position where if you don't want to train AI then don't post your content publicly.
I truly detest this phrase, but: “Citation Needed”
Alice (content creator) told Bob (YouTube). Bob told Carol (OpenAI et al.) when Carol asked Bob for a video.
This is only true in the absence of an agreement between Alice and Bob that Bob won't tattle.
Google was training on all of YouTube back in 2012 - https://slate.com/technology/2012/06/google-computers-learn-...
I'm also not saying I _support_ ToS
Consent requires the person knowing what they are consenting to. How many people do you think know that posting your face on Facebook consents to using your face to generate photos? Your voice to create voice generators?
What about when someone posts a group photo or video and despite you not having an account that the same rules apply? What about all the photos created by others and uploaded by those with no ownership or rights to the photos in the first place? They can't legally give away the rights.
Then we also have to consider that the environment has changed and the terms changed under people. You may have signed up for Facebook and knew they were going to use your data to sell ads and do some analysis. But you didn't know AI was coming (let's be real, no one knew even at the time of AlexNet. It didn't become clear in tech groups until maybe 2016/2017). Sure, the agreements might "be the same" but people always contextualize things. In 2012 you might not care that Facebook "owns" your voice because you know they have to host the file, and contextually it is "just over legalize". But in 2024, now the context is that this trains a model that can replicate your voice and this means something VERY different.
I think there's this common attitude of "Oh, well you agreed to the terms" as if this makes the terms fair, reasonable, or okay in the first place. It doesn't account for the reality of the situation. It just further legitimizes this world where a person is legally culpable for not having domain expertise. News flash, we live in a specialized world and you can't be a domain expert in everything. It doesn't account for the fact that there is often no alternative and there is serious cohesion at this point. Unfortunately, not all the choices in your life are up to you and many are social (and many of those are idiotic and can force you into decisions nearly no one in the group wants[0]) and not all the choices you make ultimately come down to you. If it was up to me everyone would be using Signal to communicate with me and no one would have bought Apple products when they made their devices unupgradable and charged more for _increasing_ your disk space than it would cost to buy 1.5 drives in the first place. But also, thank god that everything isn't up to me because I'm also a fucking idiot.
No, I don't think there was consent this whole time and as time progressed we amount and capacity you have had to consent to has decreased. And dismissing it (the fact that consent is not binary) just makes the problem exponentially worse. Not to mention all the dark patterns which I can do a whole other rant about.
[0] Prime example: the American presidential race. The vast majority of people would rather have neither Trump nor Biden running. The vast majority of people do not want a gerontocracy but will vote for one of the two candidates because what other choice do they have? People __could__ agree to to vote for a third party, but doing so gives such a strategic disadvantage that people are rational in attempting to choose from the options provided to them (see primaries).
I think we've been able to see this coming for many, many years.
You may think this now, but I'm willing to bet a lot of money that you didn't believe this in 2012 and even 2015. Be careful to not rewrite history. It's an easy thing to do.
Arguably this is just not consent, plain and simple. There could be no reasonable expectation for a person to believe they are agreeing to this. Whereas uninformed consent, I'd argue, is more like how people were agreeing to giving away their data but do not actually understand how the data is taken and used (let's be real, this includes even a lot, if not most, tech people).
I can’t see that argument holding up in court. You’re using a definition of informed and consent that are rather novel.
Funnily enough, you can actually use LLMs to make sense of these things. We're more empowered than ever to process legal text. It used to be that your only option was to consult somebody with legal training. Which of course isn't something a lot of users would do.
In any case, you are using a moralistic argument in a legal context. That's simply not how the law gets interpreted in court rooms. It might motivate politicians to write new laws though. Which then get interepreted in court rooms. But given the complexity of the matter at hand, the outcome might not be something that is a net positive. In fact that is extremely unlikely.
#1 It's not all that certain AI training is fair use. There could be significant damages if it is found to not be fair use.
#2 "We're going to steal your shit even if you won't give us consent to do that" is a bad idea. We're already seeing the practical results: People stop/reduce posting their works to the clearnet. What's your scraper bot going to do? Create an account on every 'private' forum and immediately get sued? Pay a subscription fee to every single Patreon page?
And the big one, #3: It destroys public support for your technology, which is essential if you want to survive the oncoming government regulation.
Look at the public response to various tech companies stopping their AI rollout in europe over EU regulations. Varying from "Yes, this is exactly what we asked for" to "Good fucking riddance".
And it will get only worse yet. Europe's privacy authorities are already issuing quiet statements that the scraping of social media posts, such as is done for AI data collection, is not legal under the GDPR. (The law's pretty clear on this, it doesn't matter that it was "posted publicly", you're not allowed to use personal data like that) There's already whispers of going after the LLMs themselves, as they contain and continue to process personal data as well.
AI needs public support, and the lack of consent is slowly bleeding it dry.
> What's your scraper bot going to do? Create an account on every 'private' forum and immediately get sued?
* In the case of big IP holders (e.g. media companies, news organizations) this is just obtaining consent. The only fun quirck is that OpenAI's purchasing of this drastically lowers the "AI training is fair use" claim by proving the existance of a market.
* In the case of platforms like Reddit, it just kicks the problem one layer down. The platform does obtain "consent" through it's ToS. (Beware that this consent is legally weak, and won't protect your ass from anything outside copyright) But users will still see it as "stealing" and may flee the platform.
There's still a notable shift away from the clearnet.
I want to add nuance. I don't think it is the "move fast and break things" mentality that creates all the shit, but that that there is no "time to clean up, everybody do your share" mentality to complement it. Doing things often creates a mess, and doing hard things often creates a bigger mess. You can't make a fancy meal without dirtying a bunch of dishes. But are we seriously not hiring "dishwashers"? Creating a mess is unavoidable, and certainly it shouldn't be too large of a mess, but the dishes are piling up and we can't hide it anymore. It isn't a linear problem because the mess compounds and the mess itself generates more mess. We won't refactor. We won't rewrite. So we just have patchwork on top of patchwork. That's enshitification.
We also have a status quo that we sprint into a sprint and try to move as fast as we can but only measure how fast we're going by looking at how far we moved in a quarter. There's no long term measurement because "that's too hard." But this is like trying to circumnavigate the world and choosing to walk. You'll make progress every day and more importantly, __measurable__ (and easily measurable) progress. But you could spend 11 months building a fucking Cesna and still beat the person that could walk on water. You need to move fast (like the Cesna), but to move fast requires also slowing down. Who is willing to slow down?
I think most people don't have a problem with people using public data to do research or similar activities. That people wouldn't be up in arms if OpenAI scraped all of YouTube, trained on it, proved internally that they could do cool stuff with it, AND THEN either started to generate their own data for training or started to purchase data. Even though this would still be costly to YouTube and be a weird ethical ground (like getting a "free trial" (or theft) of the data and pay only if it works).
As someone with anxiety, I can assure you, it is not a good idea to constantly be rushing around chasing everything that needs to be solved. You just make more messes because you sloppily "fix" the issues, trying to move onto the next. The trick is to fight your own mind, slow down, triage, and solve anything that isn't a literal fucking fire with calm and care, no matter how much your own mind wants to convince you it is an emergency. But when everything is an emergency, nothing is. And that's the problem. We created an economy based on a business strategy that is functionally equivalent to an untreated and severe anxiety disorder.
It's really interesting. Ever since I started publicly stating that I've removed all my works from the public web because there is no realistic method of defending myself against AI scrapers, I've been encountering a surprising number of people who say they've done the same thing.
I don't know how far this will go, but at least I know I'm not alone.
That will have a finite lifetime.
We've already seen issues with just the couple of years' difference that the first LLMs have, and it will definitely get worse, as time goes on.
They will, however, be able to improve what they do with the data, and some of them will probably get enough funding to buy the data.
Don't be shocked to find that the reason so many organizations are restricting access to their data, has nothing, whatsoever, to do with things like privacy and copyright.
They just see a lot of money sloshing around, and want to hold out for their share.
The pipes will reopen, but the free lunch is probably done.
One big issue will probably be, that the free data will be extremely low-quality (possibly poisonous) data, so that might be very bad for folks on the lower tiers.
Eager to create "safety" the wikipedia-ish discussion where no two truths can exist simultaneously will happen in private and the official narrative of everything however preposterous will be the only one.
We should also create LLM driven content moderation so that wrong think can be silenced immediately.
The masses of brainwashed slaves will no doubt learn to like it.
It's easier to ask for forgiveness than consent,
the cost of doing business,
etcetera, etcetera.
Well so it's not really disappearing at all.
having worked in this topic, creating synthetic data is not always easy, and you still need real data to get the best results. if you look beyond basic internet media, there are tons of fields where "real data" does not exist in quantities that could help create effective models. these might be good case studies on how to proceed further in small endeavours.
"Changing the license on the data doesn’t retroactively revoke that permission, and the primary impact is on later-arriving actors, who are typically either smaller start-ups or researchers."
And also this:
"Mr. Longpre said that one of the big takeaways from the study is that we need new tools to give website owners more precise ways to control the use of their data."
This will happen once everything is looted/not needed anymore.
The truth is that no one knows where the dust will settle. All covered wagons are venturing west of the Rockies at the same time. Content creators are not really facing a worse outlook than everyone else. They are, however, in a better position to hinder the advance.
In principle, an AI learning from a scientific textbook is no different than a human student doing the same. Neither will violate copyright law when they're done learning, except perhaps accidentally - paraphrasing and facts are not violations. Unfortunately, legal and ethical principles can differ from legal reality. We're left hoping that some altruistic legal minds will open up a Northwest Passage for us, like Thomas Penfield Jackson in US v. Microsoft.
The worst possible outcome is that we end up with an all-powerful AI cartel which negotiates massive deals with IP conglomerates, locking out competition and open and free alternatives.
Once AI is used as the buffer between all forms of communication (making sure the emails I send put me in the best light, filtering incoming email to make me more productive, generating web content, reading web content in a way that makes me more efficient, even voice and video calls, the nation state can not only monitor, but also subtly alter all communications, in real time, and in context to guide/enforce how people interact/communicate and think.
This is already happening in areas such as search, getting "answers", and content generation. Your information is constrained/mapped into what the the AI deems appropriate.
We already know that the NSA taps into the communication backbone of the internet. Why not add a real-time AI monitor/filter to best manage how we all communicate ?
If you think artistic people of various kinds are doing particularly well and need to have their privilege checked to stop the discrimination against AI tech bros...then I think you are out of touch. The starving artist meme is a thing for a reason, and most content creators are not making a lot of money. They enjoy it, and maybe make a little on the side. But most work a shitty day job as well to fund their passion, the thing that makes them enjoy life and helps them find meaning. They just want to contribute to culture and enjoy making art but wouldn't be able to justify it if all the existing low wage starving artist work dries up in favor of essentially free AI generated crap.
I'd hazard a guess that the majority of the country would take their side and call Silicon Valley tech bros and VCs privileged and anti-social.
I think a culture in which it is non-viable for an ordinary person to participate in art and culture on the side is kind of sad. I value these things because there is a person on the other end and it helps me empathize and try and understand how they are feeling and thinking. An AI kills that. You'll get great stock photos generic advertising, but you'll destroy something fundamental to human culture.
I think pointing out that it's crappy is a much better critique of crappy content than looking at how it was generated.
As for tech bros vs. artists, that's a reality show I haven't watched and one I'm not interested in. I'm interested in progress and human advancement, and I'm definitely not a Libertarian or tech bro.
If you were writing highly derivative textbooks, your college textbook's author would be peeved, too. That is the AI analogy, not that you read a chemistry textbook and are then breaking some law when you go do chemistry.
Image generator models are a few GB in size, despite copyright claimants saying there are petabytes in there.
In principle human brains don't count as a copy of a copyrighted work, whereas the AI model that is capable of regurgitating the original-- I don't understand how that's not a copy[1] or derivative work which is a good indication the judges presented with these questions won't understand either.
So, no, a brain with knowledge is a brain, and an AI model with knowledge is a potentially infringing copy or derivative work.
[1] https://www.law.cornell.edu/definitions/uscode.php?width=840...
If you limit your critique to the AI's ability to recite a poem or lyrics verbatim, then you have a point. However, I used the example of a scientific textbook for that exact reason. To the extent that the AI only paraphrases or cites facts from, it's not a copy.
The derivative work issue is more interesting. The AI can generate a derivative work, but won't do so without a prompt asking for one. If you ask for an Oz-like bedtime story, you should get an Anxious Lynx and not a Cowardly Lion.
If you ask for a story with Oz characters, the result will be a derivative work, but it will include your prompt. The work was not in the AI, though the ability to make it was. In this sense, the AI is more like an instrument that can be used to violate copyright law, like an employee who has read Oz would be. It's hard to find a close inanimate analogue, but it's probably on the same shelf as search engines and typewriters.
Many AIs have in fact reproduced copyrighted work in large portions, it’s one of the reasons there are a bunch of lawsuits.
> AI can generate a derivative work, but won’t do so without a prompt asking for one
This is an odd and vague claim. Which AI are you talking about? And why do you claim there’s any generative AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works.
> AI is more like an instrument that can be used to violate copyright law
If the training data wasn’t legally acquired, using it to train in the first place may violate the law. People are trying to change copyright’s interpretation, but pending the outcome of that it’s pretty clear that today’s AI may have violated copyright law before generating any output.
Let's exclude maliciously crafted prompts. Given an "innocent" prompt that is not meant to elicit a copyright violation, what is the probability than an LLM will generate something that will both reproduce a copyrighted work and fail the fair use tests? It's obviously non-zero, but it's exceedingly small.
I'd say it's around the same probability as Google Street View accidentally reproducing a copyrighted map that was displayed in a storefront window. A claim that Google has to pre-compensate all map publishers because of this possibility is just an attack on Street View per se.
> And why do you claim there’s any generate AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works.
Just because something sounds like a derivative work doesn't mean it will infringe copyright or fail the fair use tests (see sibling threads.)
> Just because something sounds like a derivative work doesn’t mean it will infringe copyright or fail the fair use tests
That’s true! Remixes where the source comes from multiple songs have a history of escaping copyright claims more often than derivative works of a single song.
But that doesn’t change the fact that using copyrighted data without permission and without Fair Use exemption to train AI is currently illegal. And it doesn’t change the fact that when AI reproduces substantial portions of any given work, regardless of what the probability was, it is illegal. Claims that the probability of it are “exceedingly small” don’t stand up to the (now many!) actual cases filed in court demonstrating AI reproducing copyrighted works, and not because the chances were small, but because it simply breaks the law.
I think the more sensible take is that the reproduction by the LLM violates copyright, but not the act of training nor the LLM weights themselves.
The part you’re missing here is that acquiring and using copyrighted data without permission also violates copyright. Just like copying and watching a movie you didn’t pay for or legally acquire violates copyright. It’s a worse offense to distribute a movie you don’t have the rights to, but copying one and consuming it without permission is an offense nonetheless. The act of training on data not legally acquired, therefore, certainly can be a copyright violation.
This is simply not true. There are many things that copyright law prevents, but no copyright law goes as far as to say that all use without permission is forbidden. And no copyright law expressly forbids training AI either.
The copyright office says downloading movies violates copyright.
https://www.copyright.gov/help/faq/faq-fairuse.html#:~:text=....
See also (same page) “If you use a copyrighted work without authorization, the owner may be entitled to bring an infringement action against you.”
> no copyright law expressly forbids training AI either.
That’s irrelevant. The law covers using copyrighted material without permission. Fair Use is a type of permission, but otherwise they don’t have to specify what it gets used for.
If you had memorized the textbook from a stash of torrented epubs, as a commercial activity with intent to monetize, that likely wouldn't go over well for you either.
The "big AI" companies are trying to create a world of rules for thee not for me by claiming that when copyright infringement is large scale enough, it somehow becomes fair use.
Because creation of a derivative work isn't a problem itself, it only infringes at the point / time of distribution, which is the whole point of open models. But big tech could probably argue that their training of a private model doesn't infringe per se, and it is then the users who then create prompts that cause the regurgitation to occur that hit the tripwire of triggering copyright. Presumably, big tech can solve that, either with terms of use (telling the user it's their responsibility / liability) or YouTube style where a Content ID like system skewers anything that it can't prove isn't infringing before it gets to the user.
An AI is nothing more than a large network of regression models, a very large set of equations. It is not a copy of anything.
Human brains are often treated specially in law, but I don't think they should be.
It's not about the process of building a model. It's about the copyright infringement to acquire the training set and it's about the copyright infringement when producing output. Humans, just like these AI companies, can infringe copyrights by downloading copyrighted material. Humans, just like AI programs, can produce material that is derivative of material protected by copyright. Whatever the model represents and whatever is in your brain is not high on the list of concerns.
Unless you're downloading from something like Z-Library (which incidentally is a deliberate attack on Western IP on Putin's behalf) you generally have the fair use rights to "digest" the information you receive. You can process the contents of a webpage or momentarily OCR a printed page, provided that it was legally distributed and that you don't store a copy proper. There can be specialized restrictions if it's something that you have a license for instead of a copy of, see e.g. https://en.wikipedia.org/wiki/Shrinkwrap_(contract_law)
As for derivative works, not all the works that are made "in proximity of" previous works are derivative works. You can make a simulated reality movie without infringing on the Matrix, for instance. Just write a somewhat different plot and don't name your lead character Neo.
The AI companies are definitely storing proper copies of their "corpus".
> You can make a simulated reality movie without infringing on the Matrix, for instance.
You can also make a simulated reality movie, even without the name Neo, and a court could find it substantially similar to the Matrix and thus copyright infringing. It doesn't need to be the same.
Their own lawyers would be on their case if they were infringing halfway through the pipeline. They don't have to, and it would weaken their case where it counts.
> a court could find it substantially similar to the Matrix and thus copyright infringing
Substantially similar is not the standard, otherwise they would have had to stop making time travel movies about 70 years ago.
https://crsreports.congress.gov/product/pdf/LSB/LSB10922
> Under U.S. case law, copyright owners may be able to show that such outputs infringe their copyrights if the AI program both (1) had access to their works and (2) created “substantially similar” outputs
From Shaw v Lindheim.
> Their own lawyers would be on their case if they were infringing halfway through the pipeline.
This is why many of these AI companies are their own entities with sponsorships or shares held by other, larger companies. If they're on the hook for obscene copyright infringement, they just close down.
This quote is about AI, not movies. If you read the entire report, there are many other caveats and provisions:
Whether or not copying constitutes fair use depends on four statutory factors under 17 U.S.C. § 107:
1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
2. the nature of the copyrighted work;
3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
4. the effect of the use upon the potential market for or value of the copyrighted work.
The substantial similarity test is difficult to define and varies across U.S. courts.I just read your comment. I also read news articles I paid for today. None of that is copyright infringement, and scraping that same content to train an AI is no different.
Yes, when an AI produces a substantially close reproduction that would be copyright infringement, but some systems have guards to prevent that, like Github Copilot.
Scraping is copyright infringement. Try telling the RIAA that scraping songs from YouTube or Spotify is not copyright infringement. Try telling the MPAA that scraping movies from Netflix is not copyright infringement. Same for book publishers.
How so? Generative AI has developed in a way that's very amenable to the bearers of accumulated privilege, it requires so much capital expenditure that it's in the hands of a few companies that can be variously negotiated with or strong armed. This is not at all like the computer and internet revolutions of the 90s.
> The worst possible outcome is that we end up with an all-powerful AI cartel which negotiates massive deals with IP conglomerates, locking out competition and open and free alternatives.
What's the difference with the status quo if the main players are Microsoft, Meta and Google?
Even at today's bloated GPU prices (which won't last,) training rigs cost about the same as the servers and data pipes that serious dot-coms needed, and much less than the fabs that serious semiconductor outfits needed in ~1980, especially when adjusted for inflation. Many AI startups don't do their own primary training at all, making their costs identical to barebones dot-coms'.
> What's the difference with the status quo if the main players are Microsoft, Meta and Google?
You forgot OpenAI, Anthropic, Apple and Amazon. Of the seven, two are new, three emerged from the last revolution and two from the one before.
The incumbent companies in the 90s were the likes of IBM, DEC, HP, Unisys, SGI and Sun. Almost all of them smelled the coffee and invested heavily in various Internet plays. Only a few survived, most of them only barely. The previous revolution was even bloodier.
Today's incumbents are also smelling the coffee and have AI plays. None of them is certain they'll still be around 10 years from now.
The reason Meta opened up Llama is because they were convinced that the barriers to entry were minimal. In 2022, Altman famously claimed that if there was a Moore's Law in parameter count, it had already topped out. A large fraction of use cases can be handled with models under 10B. Training algorithms are increasing in efficiency faster than parameter counts are increasing. The only real barriers to entry are patents (and we won't know what that looks like until they break stealth) and the possibility of nefarious anticompetitive deals with the IP conglomerates that we're discussing here.
Many economists expect the disruption caused by AI to be much greater than that of PCs and the Internet. The usual parallel is electrification. The gentry from back East is in the covered wagons too.
If AI were only learning from scientific textbooks very few people would have problem with it. When AI learns from creative works like novels, it runs into the same problems we have as humans. We can read a book, but the minute we start distributing copies of it from memory we're in legal trouble.
I mean the entire industry is based on doing stuff without seeking consent.
All the major players seem to have used at some point at least the books dataset so def no regard for consent or copyright.
Bigger issue is mechanical limits like Twitter and Reddit - walled gardens of info. That could entrench existing players, whether via money (pay me for data), ethics (oh now suddenly consent matters) or just timing (more stuff mechanically restricted)
I doubt they can continue the progress at the same speed they have been at so far. Because the game is set to become more difficult.
.... says the article behind a paywall.
TBH I would not feel any sadness if LLM models plateaued, greatly decellerated development, or even -regressed- as they start to ingest all the garbage already being spewed by LLMs.
Consider this comment itself. How much do you think it is worth for an AI company? Maybe 0.00001 dollars? How would you handle the logistics of money that little money?
This is just a recent hype which relies on getting insane amounts of data to train, but we had and will have AI models that do not rely on training using data without consent.
And it makes sense that LLMs were the ones to trigger the hype. AI in general has been making steady progress but most usecases are really hard to explain to layman. Give them a chatbot and they naturally understand what's happening.
The only thing I'm seeing is that people are using them wrong. They're using them like all-seeing oracles, which is exacerbated by the confidence with which LLMs provide their answers, and the innate human idea that "it's a computer so it must be right".
But knowledge isn't really where LLMs shine, at least not without search engine integrations. It's rather generation, summarisation, translation (with context sensitivity), rewriting styles etc.