Those claiming AI training on copyrighted works is “theft” misunderstand key aspects of copyright law and AI technology. Copyright protects specific expressions of ideas, not the ideas themselves. When AI systems ingest copyrighted works, they’re extracting general patterns and concepts - the “Bob Dylan-ness” or “Hemingway-ness” - not copying specific text or images.
This process is akin to how humans learn by reading widely and absorbing styles and techniques, rather than memorizing and reproducing exact passages. The AI discards the original text, keeping only abstract representations in “vector space”. When generating new content, the AI isn’t recreating copyrighted works, but producing new expressions inspired by the concepts it’s learned.
This is fundamentally different from copying a book or song. It’s more like the long-standing artistic tradition of being influenced by others’ work. The law has always recognized that ideas themselves can’t be owned - only particular expressions of them.
Moreover, there’s precedent for this kind of use being considered “transformative” and thus fair use. The Google Books project, which scanned millions of books to create a searchable index, was ruled legal despite protests from authors and publishers. AI training is arguably even more transformative.
While it’s understandable that creators feel uneasy about this new technology, labeling it “theft” is both legally and technically inaccurate. We may need new ways to support and compensate creators in the AI age, but that doesn’t make the current use of copyrighted works for AI training illegal or unethical.
For those interested, this argument is nicely laid out by Damien Riehl in FLOSS Weekly episode 744. https://twit.tv/shows/floss-weekly/episodes/744
If they can base their business on stealing, then we can steal their AI services, right?
Here’s an experiment for you to try at home. Ask an AI model a question, copy a sentence or two of what they give back, and paste it into a search engine. The results may surprise you.
And stop comparing AI to humans but then giving AI models more freedom. If I wrote a paper I’d need to cite my sources. Where the fuck are your sources ChatGPT? Oh right, we’re not allowed to see that but you can take whatever you want from us. Sounds fair.
It’s not a breach of copyright or other IP law not to cite sources on your paper.
Getting your paper rejected for lacking sources is also not infringing in your freedom. Being forced to pay damages and delete your paper from any public space would be infringement of your freedom.
I mean, you’re not necessarily wrong. But that doesn’t change the fact that it’s still stealing, which was my point. Just because laws haven’t caught up to it yet doesn’t make it any less of a shitty thing to do.
It’s not stealing, its not even ‘piracy’ which also is not stealing.
Copyright laws need to be scaled back, to not criminalize socially accepted behavior, not expand.
I’ll train my AI on just the bee movie. Then I’m going to ask it “can you make me a movie about bees”? When it spits the whole movie, I can just watch it or sell it or whatever, it was a creation of my AI, which learned just like any human would! Of course I didn’t even pay for the original copy to train my AI, it’s for learning purposes, and learning should be a basic human right!
In the meantime I’ll introduce myself into the servers of large corporations and read their emails, codebase, teams and strategic analysis, it’s just learning!
learning should be a basic human right!
Education is a basic human right (except maybe in Usa, then it should be one there)
Yeah. A human right.
I am thrilled to see the output you get!
The problem with your argument is that it is 100% possible to get ChatGPT to produce verbatim extracts of copyrighted works. This has been suppressed by OpenAI in a rather brute force kind of way, by prohibiting the prompts that have been found so far to do this (e.g. the infamous “poetry poetry poetry…” ad infinitum hack), but the possibility is still there, no matter how much they try to plaster over it. In fact there are some people, much smarter than me, who see technical similarities between compression technology and the process of training an LLM, calling it a “blurry JPEG of the Internet”… the point being, you wouldn’t allow distribution of a copyrighted book just because you compressed it in a ZIP file first.
The problem with your argument is that it is 100% possible to get ChatGPT to produce verbatim extracts of copyrighted works.
Exactly! This is the core of the argument The New York Times made against OpenAI. And I think they are right.
The examples they provided were for very widely distributed stories (i.e. present in the data set many times over). The prompts they used were not provided. How many times they had to prompt was not provided. Their results are very difficult to reproduce, if not impossible, especially on newer models.
I mean, sure, it happens. But it’s not a generalizable problem. You’re not going to get it to regurgitate your Lemmy comment, even if they’ve trained on it. You can’t just go and ask it to write Harry Potter and the goblet of fire for you. It’s not the intended purpose of this technology. I expect it’ll largely be a solved problem in 5-10 years, if not sooner.
I agree. You can’t just dismiss the problem saying it’s “just data represented in vector space” and on the other hand not be able properly censor the models and require AI safety research. If you don’t know exactly what’s going on inside, you also can’t claim that copyright is not being violated.
It honestly blows my mind that people look at a neutral network that’s even capable of recreating short works it was trained on without having access to that text during generation… and choose to focus on IP law.
Right! Like if we could honestly further enhance that feature its an incredible increase in compression tech!
The problem with your argument is that it is 100% possible to get ChatGPT to produce verbatim extracts of copyrighted works.
What method still works? I’d like to try it.
I have access to ChatGPT 4, and the latest Anthropic model.
Edit: hm… no answers but downvotes. I wonder why that is.
ML techniques have been very useful in compression, yes, but it’s sort of nuts to say that a data structure that encodes only (sometimes overly so for certain regions of its latent space/embedding space/semantics space/whatever you want to call it right now) relationships between values rather than value sequences themselves as storing contiguous copyright protected works is storing partiularized creative works in particularly identifiable manner.
This would be a good point, if this is what the explicit purpose of the AI was. Which it isn’t. It can quote certain information verbatim despite not containing that data verbatim, through the process of learning, for the same reason we can.
I can ask you to quote famous lines from books all day as well. That doesn’t mean that you knowing those lines means you infringed on copyright. Now, if you were to put those to paper and sell them, you might get a cease and desist or a lawsuit. Therein lies the difference. Your goal would be explicitly to infringe on the specific expression of those words. Any human that would explicitly try to get an AI to produce infringing material… would be infringing. And unknowing infringement… well there are countless court cases where both sides think they did nothing wrong.
You don’t even need AI for that, if you followed the Infinite Monkey Theorem and just happened to stumble upon a work falling under copyright, you still could not sell it even if it was produced by a purely random process.
Another great example is the Mona Lisa. Most people know what it looks like and if they had sufficient talent could mimic it 1:1. However, there are numerous adaptations of the Mona Lisa that are not infringing (by today’s standards), because they transform the work to the point where it’s no longer the original expression, but a re-expression of the same idea. Anything less than that is pretty much completely safe infringement wise.
You’re right though that OpenAI tries to cover their ass by implementing safeguards. Which is to be expected because it’s a legal argument in court that once they became aware of situations they have to take steps to limit harm. They can indeed not prevent it completely, but it’s the effort that counts. Practically none of that kind of moderation is 100% effective. Otherwise we’d live in a pretty good world.
Y’all should really stop expecting people to buy into the analogy between human learning and machine learning i.e. “humans do it, so it’s okay if a computer does it too”. First of all there are vast differences between how humans learn and how machines “learn”, and second, it doesn’t matter anyway because there is lots of legal/moral precedent for not assigning the same rights to machines that are normally assigned to humans (for example, no intellectual property right has been granted to any synthetic media yet that I’m aware of).
That said, I agree that “the model contains a copy of the training data” is not a very good critique–a much stronger one would be to simply note all of the works with a Creative Commons “No Derivatives” license in the training data, since it is hard to argue that the model checkpoint isn’t derived from the training data.
a much stronger one would be to simply note all of the works with a Creative Commons “No Derivatives” license in the training data, since it is hard to argue that the model checkpoint isn’t derived from the training data.
Not really. First of all, creative commons strictly loosens the copyright restrictions on a work. The strongest license is actually no explicit license i.e. “All Rights Reserved.” No derivatives is already included under full, default, copyright.
Second, derivative has a pretty strict legal definition. It’s not enough to say that the derived work was created using a protected work, or even that the derived work couldn’t exist without the protected work. Some examples: create a word cloud of your favorite book, analyze the tone of news article to help you trade stocks, produce an image containing the most prominent color in every frame of a movie, or create a search index of the words found on all websites on the internet. All of that is absolutely allowed under even the strictest of copyright protections.
Statistical analysis of copyrighted materials, as in training AI, easily clears that same bar.
Equating LLMs with compression doesn’t make sense. Model sizes are larger than their training sets. if it requires “hacking” to extract text of sufficient length to break copyright, and the platform is doing everything they can to prevent it, that just makes them like every platform. I can download © material from YouTube (or wherever) all day long.
Model sizes are larger than their training sets
Excuse me, what? You think Huggingface is hosting 100’s of checkpoints each of which are multiples of their training data, which is on the order of terabytes or petabytes in disk space? I don’t know if I agree with the compression argument, myself, but for other reasons–your retort is objectively false.
Just taking GPT 3 as an example, its training set was 45 terabytes, yes. But that set was filtered and processed down to about 570 GB. GPT 3 was only actually trained on that 570 GB. The model itself is about 700 GB. Much of the generalized intelligence of an LLM comes from abstraction to other contexts.
Table 2.2 shows the final mixture of datasets that we used in training. The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl covering 2016 to 2019, constituting 45TB of compressed plaintext before filtering and 570GB after filtering, roughly equivalent to 400 billion byte-pair-encoded tokens. Language Models are Few-Shot Learners
*Did some more looking, and that model size estimate assumes 32 bit float. It’s actually 16 bit, so the model size is 350GB… technically some compression after all!
The issue isn’t that you can coax AI into giving away unaltered copyrighted books out of their trunk, the issue is that if you were to open the hood, you’d see that the entire engine is made of unaltered copyrighted books.
All those “anti hacking” measures are just there to obfuscate the fact that that the unaltered works are being in use and recallable at all times.
This is an inaccurate understanding of what’s going on. Under the hood is a neutral network with weights and biases, not a database of copyrighted work. That neutral network was trained on a HEAVILY filtered training set (as mentioned above, 45 terabytes was reduced to 570 GB for GPT3). Getting it to bug out and generate full sections of training data from its neutral network is a fun parlor trick, but you’re not going to use it to pirate a book. People do that the old fashioned way by just adding type:pdf to their common web search.
Again: nobody is complaining that you can make AI spit out their training data because AI is the only source of that training data. That is not the issue and nobody cares about AI as a delivery source of pirated material. The issue is that next to the transformed output, the not-transformed input is being in use in a commercial product.
The issue is that next to the transformed output, the not-transformed input is being in use in a commercial product.
Are you only talking about the word repetition glitch?
Look… All I have to say is… Support the Internet Archive!
(please)
Heh. Funny that this comment is uncontroversial. The Internet Archive supports Fair Use because, of course, it does.
This is from a position paper explicitly endorsed by the IA:
Based on well-established precedent, the ingestion of copyrighted works to create large language models or other AI training databases generally is a fair use.
By
- Library Copyright Alliance
- American Library Association
- Association of Research Libraries
tweet is good, your body argument is completely wrong
You know, those obsessed with pushing AI would do a lot better if they dropped the patronizing tone in every single one of their comments defending them.
It’s always fun reading “but you just don’t understand”.
On the other hand, it’s hard to have a serious discussion with people who insist that building a LLM or diffusion model amounts to copying pieces of material into an obfuscated database. And then having to deal with the typical reply after explanation is attempted of “that isn’t the point!” but without any elaboration strongly implies to me that some people just want to be pissy and don’t want to hear how they may have been manipulated into taking a pro-corporate, hyper-capitalist position on something.
I love that the collectivist ideal of sharing all that we’ve created for the betterment of humanity is being twisted into this disgusting display of corporate greed and overreach. OpenAI doesn’t need shit. They don’t have an inherent right to exist but must constantly make the case for it’s existence.
The bottom line is that if corporations need data that they themselves cannot create in order to build and sell a service then they must pay for it. One way or another.
I see this all as parallels with how aquifers and water rights have been handled and I’d argue we’ve fucked that up as well.
They do, though. They purchase data sets from people with licenses, use open source data sets, and/or scrape publicly available data themselves. Worst case they could download pirated data sets, but that’s copyright infringement committed by the entity distributing the data without the legal authority.
Beyond that, copyright doesn’t protect the work from being used to create something else, as long as you’re not distributing significant portions of it. Movie and book reviewers won that legal battle long ago.
Training data IS a massive industry already. You don’t see it because you probably don’t work in a field directly dealing with it. I work in medtech and millions and millions of dollars are spent acquiring training data every year. Should some new unique IP right be found on using otherwise legally rendered data to train AI, it is almost certainly going to be contracted away to hosting platforms via totally sound ToS and then further monetized such that only large and we’ll funded corporate entities can utilize it.
I don’t get your comment, are the pro corporate for AI or against it?
I have no personal interest in the matter, tbh. But I want people to actually understand what they’re advocating for and what the downstream effects would inevitably be. Model training is not inherently infringing activity under current IP law. It just isn’t. Neither the law, legislative or judicial, nor the actual engineering and operations of these current models support at all a finding of infringement. Effectively, this means that new legislation needs to be made to handle the issue. Most are effectively advocating for an entirely new IP right in the form of a “right to learn from” which further assetizes ideas and intangibles such that we get further shuffled into endstage capitalism, which most advocates are also presumably against.
I’m pretty sure most people are just mad that this is basically “rules for thee but not for me”, why should a company be free to pirate but I can’t? Case in point is the internet archive losing their case against a publisher. That’s the crux of the issue.
I get that that’s how it feels given how it’s being reported, but the reality is that due to the way this sort of ML works, what internet archive does and what an arbitrary GPT does are completely different, with the former being an explicit and straightforward copy relying on Fair Use defense and the latter being the industrialized version of intensive note taking into a notebook full of such notes while reading a book. That the outputs of such models are totally devoid of IP protections actually makes a pretty big difference imo in their usefulness to the entities we’re most concerned about, but that certainly doesn’t address the economic dilemma of putting an entire sector of labor at risk in narrow areas.
Though I am not a lawyer by training, I have been involved in such debates personally and professionally for many years. This post is unfortunately misguided. Copyright law makes concessions for education and creativity, including criticism and satire, because we recognize the value of such activities for human development. Debates over the excesses of copyright in the digital age were specifically about humans finding the application of copyright to the internet and all things digital too restrictive for their educational, creative, and yes, also their entertainment needs. So any anti-copyright arguments back then were in the spirit specifically of protecting the average person and public-interest non-profit institutions, such as digital archives and libraries, from big copyright owners who would sue and lobby for total control over every file in their catalogue, sometimes in the process severely limiting human potential.
AI’s ingesting of text and other formats is “learning” in name only, a term borrowed by computer scientists to describe a purely computational process. It does not hold the same value socially or morally as the learning that humans require to function and progress individually and collectively.
AI is not a person (unless we get definitive proof of a conscious AI, or are willing to grant every implementation of a statistical model personhood). Also AI it is not vital to human development and as such one could argue does not need special protections or special treatment to flourish. AI is a product, even more clearly so when it is proprietary and sold as a service.
Unlike past debates over copyright, this is not about protecting the little guy or organizations with a social mission from big corporate interests. It is the opposite. It is about big corporate interests turning human knowledge and creativity into a product they can then use to sell services to - and often to replace in their jobs - the very humans whose content they have ingested.
See, the tables are now turned and it is time to realize that copyright law, for all its faults, has never been only or primarily about protecting large copyright holders. It is also about protecting your average Joe from unauthorized uses of their work. More specifically uses that may cause damage, to the copyright owner or society at large. While a very imperfect mechanism, it is there for a reason, and its application need not be the end of AI. There’s a mechanism for individual copyright owners to grant rights to specific uses: it’s called licensing and should be mandatory in my view for the development of proprietary LLMs at least.
TL;DR: AI is not human, it is a product, one that may augment some tasks productively, but is also often aimed at replacing humans in their jobs - this makes all the difference in how we should balance rights and protections in law.
AI are people, my friend. /s
But, really, I think people should be able to run algorithms on whatever data they want. It’s whether the output is sufficiently different or “transformative” that matters (and other laws like using people’s likeness). Otherwise, I think the laws will get complex and nonsensical once you start adding special cases for “AI.” And I’d bet if new laws are written, they’d be written by lobbiests to further erode the threat of competition (from free software, for instance).
Bullshit. AI are not human. We shouldn’t treat them as such. AI are not creative. They just regurgitate what they are trained on. We call what it does “learning”, but that doesn’t mean we should elevate what they do to be legally equal to human learning.
It’s this same kind of twisted logic that makes people think Corporations are People.
I absolutely would download a car.
As others have said, it isn’t inspired always, sometimes it literally just copies stuff.
This feels like it was written by someone who invested their money in AI companies because they’re worried about their stocks
Sometimes I’ve noticed Google’s AI overview is a nearly word for word copy of the highest reddit result, or any result.
Lol
Generative AI does not work like this. They’re not like humans at all, it will regurgitate whatever input it receives, like how Google can’t stop Gemini from telling people to put glue in their pizza. If it really worked like that, there wouldn’t be these broad and extensive policies within tech companies about using it with company sensitive data like protection compliances. The day that a health insurance company manager says, “sure, you can feed Chat-GPT medical data” is the day I trust genAI.
Ha, ya know? I think I know some people who will just regurgitate whatever input they receive
…
:(
I feel you man lmao
I’ve just asked Gemini about cheese that slides off pizza, it didn’t recommend glue.
The last I had heard of this were articles months in saying it was still not fixed, but this doesn’t invalidate my point. It may have been retrained to respond otherwise, but it spouts garbled inputs.
It wasn’t Gemini, but the AI generated suggestions added to the top of Google search. But that AI was specifically trained to regurgitate and reference direct from websites, in an effort to minimize the amount of hallucinated answers.
Do you have a source for Search Generative Experience using a separate model? As far as I’m aware, all of Google’s AI services are powered by the Gemini LLM.
No mention of Gemini in their blog post on sge And their AI principles doc says
We acknowledge that large language models (LLMs) like those that power generative AI in Search have the potential to generate responses that seem to reflect opinions or emotions, since they have been trained on language that people use to reflect the human experience. We intentionally trained the models that power SGE to refrain from reflecting a persona. It is not designed to respond in the first person, for example, and we fine-tuned the model to provide objective, neutral responses that are corroborated with web results.
So a custom model.
This process is akin to how humans learn…
I’m so fucking sick of people saying that. We have no fucking clue how humans LEARN. Aka gather understanding aka how cognition works or what it truly is. On the contrary we can deduce that it probably isn’t very close to human memory/learning/cognition/sentience (any other buzzword that are stands-ins for things we don’t understand yet), considering human memory is extremely lossy and tends to infer its own bias, as opposed to LLMs that do neither and religiously follow patters to their own fault.
It’s quite literally a text prediction machine that started its life as a translator (and still does amazingly at that task), it just happens to turn out that general human language is a very powerful tool all on its own.
I could go on and on as I usually do on lemmy about AI, but your argument is literally “Neural network is theoretically like the nervous system, therefore human”, I have no faith in getting through to you people.
Even worse is, in order to further humanize machine learning systems, they often give them human-like names.
Kids pay for books, openAI should also pay for the material access used for training.
That would be true if they used material that was paywalled. But the vast majority of the training information used is publicly available. There’s plenty of freely available books and information that you only require an internet connection for to access, and learn from.
OpenAI like other AI companies keep their data sources confidential. But there are services and commercial databases for books that people understand are commonly used in the AI industry.
“This process is akin to how humans learn… The AI discards the original text, keeping only abstract representations…”
Now I sail the high seas myself, but I don’t think Paramount Studios would buy anyone’s defence they were only pirating their movies so they can learn the general content so they can produce their own knockoff.
Yes artists learn and inspire each other, but more often than not I’d imagine they consumed that art in an ethical way.
