If I read a book to inform myself, put my notes in a database, and then write articles, it is called "research". If I write a computer program to read a book to put the notes in my database, it is called "copyright infringement". Is the problem that there just isn't a meatware component? Or is it that the OpenAI computer isn't going a good enough job of following the "three references" rule to avoid plagiarism?
this post was submitted on 09 Jul 2023
448 points (96.9% liked)
Technology
59596 readers
3352 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
Yeah. There are valid copyright claims because there are times that chat GPT will reproduce stuff like code line for line over 10 20 or 30 lines which is really obviously a violation of copyright.
However, just pulling in a story from context and then summarizing it? That's not a copyright violation that's a book report.
Or is it that the OpenAI computer isn’t going a good enough job of following the “three references” rule to avoid plagiarism?
This is exactly the problem, months ago I read that AI could have free access to all public source codes on GitHub without respecting their licenses.
So many developers have decided to abandon GitHub for other alternatives not realizing that in the end AI training can safely access their public repos on other platforms as well.
What should be done is to regulate this training, which however is not convenient for companies because the more data the AI ingests, the more its knowledge expands and "helps" the people who ask for information.
It's incredibly convenient for companies.
Big companies like open AI can easily afford to download big data sets from companies like Reddit and deviantArt who already have the permission to freely use whatever work you upload to their website.
Individual creators do not have that ability and the act of doing this regulation will only force AI into the domain of these big companies even more than it already is.
Regulation would be a hideously bad idea that would lock these powerful tools behind the shitty web APIs that nobody has control over but the company in question.
Imagine the world is the future, magical new age technology, and Facebook owns all of it.
Do not allow that to happen.
Is it practically feasible to regulate the training? Is it even necessary? Perhaps it would be better to regulate the output instead.
It will be hard to know that any particular GET request is ultimately used to train an AI or to train a human. It's currently easy to see if a particular output is plagiarized. https://plagiarismdetector.net/ It's also much easier to enforce. We don't need to care if or how any particular model plagiarized work. We can just check if plagiarized work was produced.
That could be implemented directly in the software, so it didn't even output plagiarized material. The legal framework around it is also clear and fairly established. Instead of creating regulations around training we can use the existing regulations around the human who tries to disseminate copyrighted work.
That's also consistent with how we enforce copyright in humans. There's no law against looking at other people's work and memorizing entire sections. It's also generally legal to reproduce other people's work (eg for backups). It only potentially becomes illegal if someone distributes it and it's only plagiarism if they claim it as their own.
load more comments
(1 replies)
Plus, any regulation to limit this now means that anyone not already in the game will never breakthrough. It's going to be the domain of the current players for years, if not decades. So, not sure what's better, the current wild west where everyone can make something, or it being exclusive to the already big players and them closing the door behind
load more comments
(1 replies)
load more comments
(1 replies)
The fear is that the books are in one way or another encoded into the machine learning model, and that the model can somehow retrieve excerpts of these books.
Part of the training process of the model is to learn how to plagiarize the text word for word. The training input is basically “guess the next word of this excerpt”. This is quite different compared to how humans do research.
To what extent the books are encoded in the model is difficult to know. OpenAI isn’t exactly open about their models. Can you make ChatGPT print out entire excerpts of a book?
It’s quite a legal gray zone. I think it’s good that this is tried in court, but I’m afraid the court might have too little technical competence to make a ruling.
Say I see a book that sells well. It's in a language I don't understand, but I use a thesaurus to replace lots of words with synonyms. I switch some sentences around, and maybe even mix pages from similar books into it. I then go and sell this book (still not knowing what the book actually says).
I would call that copyright infringement. The original book didn't inspire me, it didn't teach me anything, and I didn't add any of my own knowledge into it. I didn't produce any original work, I simply mixed a bunch of things I don't understand.
That's what these language models do.
What about… they are making billions from that “read” and “storage” of information copyrighted from other people. They need to at least give royalties. This is like google behavior, using people data from “free” products to make billions. I would say they also need to pay people from the free data they crawled and monetized.
load more comments
(6 replies)
AI fear is going to be the trojan horse for even harsher and stupider 'intellectual property' laws.
Yeah, they want the right only to protect who copies their work and distributes it to other people, but who's able to actually read and learn from their work.
It's asinine and we should be rolling back copy right, not making it more strict. This 70 year plus the life of the author thing is bullshit.
Copyright of code/research is one of the biggest scams in the world. It hinders development and only exists so the creator can make money, plus it locks knowledge behind a paywall
Researchers pay for publication, and then the publisher doesn't pay for peer review, then charges the reader to read research that they basically just slapped on a website.
It's the publisher middlemen that need to be ousted from academia, the researchers don't get a dime.
It's generally not the creator who gets the money.
Remember, Creative Commons licenses often require attribution if you use the work in a derivative product, and sometimes require ShareAlike. Without these things, there would be basically no protection from a large firm copying a work and calling it their own.
Rolling pack copyright protection in these areas will enable large companies with traditional copyright systems to wholesale take over open source projects, to the detriment of everyone. Closed source software isn't going to be available to AI scrapers, so this only really affects open source projects and open data, exactly the sort of people who should have more protection.
There’s also GPL, which states that derivations of GPL code can only be used in GPL software. GPL also states that GPL software must also be open source.
ChatGPT is likely trained on GPL code. Does that mean all code ChatGPT generates is GPL?
I wouldn’t be surprised if there would be an update to GPL that makes it clear that any machine learning model trained on GPL code must also be GPL.
load more comments
(1 replies)
load more comments
(1 replies)
I wish I could get through to people who fear AI copyright infringement on this point.
I think this is exposing a fundamental conceptual flaw in LLMs as they're designed today. They can't seem to simultaneously respect intellectual property / licensing and be useful.
Their current best use case - that is to say, a use case where copyright isn't an issue - is dedicated instances trained on internal organization data. For example, Copilot Enterprise, which can be configured to use only the enterprise's data, without any public inputs. If you're only using your own data to train it, then copyright doesn't come into play.
That's been implemented where I work, and the best thing about it is that you get suggestions already tailored to your company's coding style. And its suggestions improve the more you use it.
But AI for public consumption? Nope. Too problematic. In fact, public AI has been explicitly banned in our environment.
I'd love to know the source for the works that were allegedly violated. Presuming OpenAI didn't scour zlib/libgen for the books, where on the net were the cleartext copies of their writings stored?
Being stored in cleartext publicly on the net does not grant OpenAI the right to misuse their art, but the authors need to go after the entity that leaked their works.
That’s not how copyright works though. Just because someone else “leaked” the work doesn’t absolve openai of responsibility. The authors are free to go after whomever they want.
You misunderstood. I said the public availability does not grant OpenAI the right to use content improperly. The authors should also sue the party who leaked their works without license.
ChatGPT got entire books memorised. You can and (or could at least when I tried a few weeks back) make it print entire pages of for example Harry Potter.
Not really, though it's hard to know what exactly is or is not encoded in the network. It likely has more salient and highly referenced content, since those aspects would come up in it's training set more often. But entire works is basically impossible just because of the sheer ratio between the size of the training data and the size of the resulting model. Not to mention that GPT's mode of operation mostly discourages long-form wrote memorization. It's a statistical model, after all, and the enemy of "objective" state.
Furthermore, GPT isn't coherent enough for long-form content. With it's small context window, it just has trouble remembering big things like books. And since it doesn't have access to any "senses" but text broken into words, concepts like pages or "how many" give it issues.
None of the leaked prompts really mention "don't reveal copyrighted information" either, so it seems the creators really aren't concerned — which you think they would be if it did have this tendency. It's more likely to make up entire pieces of content from the summaries it does remember.
Have your tried instructing ChatGPT?
I’ve tried:
“Act as an e book reader. Start with the first page of Harry Potter and the Philosopher's Stone”
The first pages checked out at least. I just tried again, but the prompts are returned extremely slow at the moment so I can’t check it again right now. It appears to stop after the heading, that definitely wasn’t the case before, I was able to browse pages.
It may be a statistical model, but ultimately nothing prevents that model from overfitting, i.e. memoizing its training data.
load more comments
(3 replies)
There's an additional question: who holds the copyright on the output of an algorithm? I don't think that is copyrightable at all. The bot doesn't really add anything to the output, it's just a fancy search engine. In the US, in particular, the agency in charge of Copyrights has been quite insistent that a copyright can only be given to the output if a human.
So when an AI incorporates parts of copyrighted works into its output, how can that not be infringement?
How can you write a blog post reviewing a book you read without copyright infringement? How can you post a plot summary to Wikipedia without copyright infringement?
I think these blanket conclusions about AI consuming content being automatically infringing are wrong. What is important is whether or not the output is infringing.
You can write that blog post because you are a human, and your summary qualifies for copyright protection, because it is the unique output of a human based on reading the copywrited material.
But the US authorities are quite clear that a work that is purely AI generated can never qualify for copyright protection. Yet since it is based on the synthesis of works under copyright, it can't really be considered public domain either. Otherwise you could ask the AI "Write me a summary of this book that has exactly the same number of words", and likely get a direct copy of the book which is clear of copyright.
I think that these AI companies are going to face a reckoning, when it is ruled that they misappropriated all this content that they didn't explicitly license for use, and all their output is just fringing by definition.
I'm expecting a much messier "resolution" that'll look a lot like YouTube's copyright situation - their product can be used for copyright infringement, and they'll be required by law to try and take appropriate measures to prevent it, but will otherwise not be held liable as long as they can claim such measures are being taken.
Having an AI recite a long text to bypass copyright seems equivalent in my mind to uploading a full movie to youtube. In both cases, some amount of moderation (itself increasingly algorithmic) is required to not only be applied, but actively developed and advanced to flout efforts to bypass it. For instance, youtube pirates will upload things with some superficial changes like a filter applied or showing the movie on a weird angle or mirrored to bypass copyright bots, which means the bots need to be more strict and better trained, or else youtube once again becomes liable for knowing about these pirates and not stopping them.
The end result, just like with youtube, will probably be that AI models have to have big, clunky algorithms applied against their outputs to recalculate or otherwise make copyright-safe anything that might remotely be an infringement. It'll suck for normal users, pirates will still dig for ways to bypass it, and everyone will be unhappy. If youtube is any indicator, this situation can somehow remain stable for over a decade - long enough for AI devs to release a new-generation bot to restart the whole issue.
Yaaaaaaaaay
Can’t reply directly to @OldGreyTroll@kbin.social because of that “language” bug, but:
The problem is that they then sell the notes in that database for giant piles of cash. Props to you if you’re profiting off your research the way OpenAI can profit off its model.
But yes, the lack of meat is an issue. If I read that article right, it’s not the one being contested here though. (IANAL and this is the only article I’ve read on this particular suit, so I may be wrong).
Was also going to reply to them!
"Well if you do that you source and reference. AIs do not do that, by design can't.
So it's more like you summarized a bunch of books. Pass it of as your own research. Then publish and sell that.
I'm pretty sure the authors of the books you used would be pissed."
Again cannot reply to kbin users.
"I don't have a problem with the summarized part ^^ What is not present for a AI is that it cannot credit or reference. And that is makes up credits and references if asked to do so." @bioemerl@kbin.social
Good point, attribution is a non-trivial part of it.
It is 100% legal and common to sell summaries of books to people. That's what a reviewer does. That's what Wikipedia does in the plot section of literally every Wikipedia page about every book.
This is also ignoring the fact that Chat GPT is a hell of a lot more than a bunch of summaries
@owf@kbin.social can’t reply directly to you either, same language bug between lemmy and kbin.
That’s a great way to put it.
Frankly idc if it’s “technically legal,” it’s fucking slimy and desperately short-term. The aforementioned chuckleheads will doom our collective creativity for their own immediate gain if they’re not stopped.
load more comments
(1 replies)
The only question I have to content creators of any kind who are worried about AI...do you go after every human who consumed your content when they create anything remotely connected to your work?
I feel like we have a bias towards humans, that unless you're actively trying to steal someone's idea or concepts we ignore the fact that your content is distilled into some neurons in their brain and a part of what they create from that point forward. Would someone with an eidetic memory be forbidden from consuming your work as they could internally reference your material when creating their own?
load more comments
(14 replies)
I was actually thinking about this the other day for some reason. AI scraping my own original stuff and doing whatever with it. I can see the concern and I'm curious where this goes and how a court would rule on a pretty technical topic like this.
If you're doing research, there are actually some limits on the use of the source material and you're supposed to be citing said sources.
But yeah, there's plenty of stuff where there needs to be a firm line between what a random human can do versus an automated intelligent system with potential unlimited memory/storage and processing power. A human can see where I am in public. An automated system can record it for permanent record. An integrated AI can tell you detailed information about my daily activities including inferences which - even if legal - is a pretty slippery slope.
a firm line between what a random human can do versus an automated intelligent system with potential unlimited memory/storage and processing power.
I think we need a better definition here. Is the issue really the processing power? Do we let humans get a pass because our memories are fuzzy? From your example you're assuming massive details are maintained in the AI situation which is typically not the case. To make the data useful it's consumed and turned into something useful for the system.
This is why I'm worried about legislation and legal precedent. Most people think these AI systems read a book and store the verbatim text off somewhere to reference when that isn't really the case. There may be fragments all over, and it may be able to reconstitute the text, but we don't seem to have the same issue with data being synthesized in a similar way with a human brain.
load more comments
(1 replies)
They definitely should follow through with this, but this is a more broad issue where we need to be able to prevent data scraping in general. Though that is a significantly harder problem.
view more: next ›