AI companies required to disclose copyrighted training data under new bill | Hacker News Reader

AI companies required to disclose copyrighted training data under new bill

theverge.com

97 pointsrntn45 comments

I can already see a future where a watchdog group gets an AI to generate something very close to a copyrighted work and accuses the AI company of having included that work in the training data without declaring; the AI company denying that and explaining the generation with "parallel evolution" or spontaneous inspiration of the AI or something; and we're being left to guess who's right and who's wrong...

Say a Dan Brown book enters the public domain. Then every Dan Brown book enters the public domain, as he uses more or less the same structure for every book, with different details.

Gonna be a messy time when we catch up to that can down the road.

This. Many people act as if copyright is eternal and never expiring and that fair use does not exist, but that is not (or was not) the intention. They have time limits, and fair use mean you can use content without permission and in ways the creator "did not intend".

And to your point, we cannot copyright style, only the exact content itself.

Can patent rounded corners though.

Hey, that technology took 10 years to develop.

That’s just so dumb though. You can write a book “like Dan brown” and it’s fine. It’s absolute nonsense to imply that ai can’t generate something unless there’s something similar enough that’s open. Anyone can claim that a book is like another book because that’s a dumb and pointlessly subjective metric.

I'm not sure how you inferred "metric" from my comment, only that Dan Brown's scaffolding was apparent when I read three or four of his novels in quick succession. Such patterns are equivalent to artists' signature techniques and lend themselves to transfer learning, like generating an image in the style of starry night. This is more one-shot examples versus measurement or a loss function.

You said if one book enters public domain then every book effectively enters public domain from a plausible deniability angle of using the one relevant book. But that makes no sense at all. Because anyone can write a book in the style of Dan Brown without infringing on his IP. You don’t need Starry Night to teach a model to make works like Starry Night. You just need a work like Starry Night. And if you want to argue that the machine can learn style better than humans for text, then how is this ever going to hold up in court as “too similar to the style”?

Got it. So we are effectively agreeing!

I do not think so? For one I think there is no can to kick down the road at all because the effect you say isn’t real

The copyright system has been broken for a long time. AI is just making the system’s inadequacy more obvious.

Just look at YouTube copyright strike and understand how screwed up the enforcement of a law like this is going to be.

Tom Scott did a piece on it that’s worth watching:

https://www.youtube.com/watch?v=1Jwo5qc78QU

Unlike in the real world, when it's sonor extremely difficult to prove that you have never seen a given a copyrighted work, is sort of questions seems quite practical to resolve here? Keeping complete records of what went into training an LLM sounds like what we already do?

I doubt this once Bloomberg, Netflix, and Hollywood fully jump abroad the AI train.

They'll probably just lobby to change the laws, seeing this happen already.

It’s not so hypothetical. They’re already stealing copyrighted works and putting them in the AI’s. I put a list of them with their licenses in the “Proving Wrongdoing” section of my article:

http://gethisword.com/tech/exploringai/index.html

I also put a copyright proposal in there which should work for all the competing interests. I mean, people will still be greedy. Yet, I think it’s a win to allow all copyrighted works to be used for training while restricting only the outputs to only the degree we do for people.

Personally, I really hope the law allows GPT4-style AI’s to be built legally and affordable. If so, there’s so many benefits to be gained in businesses we can start, augmenting workers to improve work/life/productivity balance, and non-profit projects.

Plus, whichever countries don’t do it will be left behind in this area by the countries that make AI training easier. Japan is the only one I’ve heard made it legal to use copyrighted works for data-science models. I don’t know that you can distribute them like The Pile or Common Crawl, though. If you scraped or bought it yourself, Japan is the closest thing I know to being legal under copyright law.

> The bill also says AI developers must report efforts to “red team” the model to prevent it from providing “inaccurate or harmful information” around [...] elections, policing, [...] public services, [...]

Pressure to suppress accurate but "harmful" information about the government?

That seems like a pretty clear cut first amendment violation, no?

This is interesting with the recent work showing that models using different architectures, but the same dataset converge to similar capabilities. It makes the case that the data set itself and the compute over it is the secret sauce for these models.

It'll be especially interesting with the ruling indicating that the output of these models isn't copyrightable as that and internally generated (from scratch) data will be the only bits omitted from this reporting. I'll be curious if the gap gets bridged on including the disclosure of data generated by a model containing copyrighted data.

The dataset is really 75% of it. Just look how runway "fixed" stable diffusion 1.4.

> This is interesting with the recent work showing that models using different architectures, but the same dataset converge to similar capabilities. It makes the case that the data set itself and the compute over it is the secret sauce for these models.

Yes, I go as far as putting 99% the merit on the dataset, given a compute budget. Humans with similar cultural exposure are also remarkably close in intelligence, even though brains are very different at micro level.

If language data is actually the source of most of our acquired intelligence, if intelligence is a collective process, if it has an evolutionary drive, then isn't it silly to discuss so much about models or brains while forgetting about its crystallized form - language and text.

It is a repository of past experience spanning millennia, and a self replication medium for ideas. All our important knowledge is encoded in language. We (and more recently LLMs) draw heavily on this recorded experience. It cost a lot to be earned in the first place. If we lost all this knowledge, it would take us millennia to recover. It's smarter than all of us.

If I were in OpenAI, I’d wonder if this hack works. We already have millions if not billions of archived chats between GPT-4 and users. Also possibly a thumbs up and thumbs down accompanying them. Once these chats reach a large number, I’d try training a new language model only on these chats. It’s very likely you’d get a GPT-4 capable language model that is not trained on no copyrighted data (but might be trained on second hand copyrighted data since it’s dataset contains completions from a model trained on copyrighted data). Wonder how you’ll change the law to account for this?

Oh god, when I build a dataset I manually go a lot of it to make sure it is in good shape. I will go through thousands of rows of data otherwise I don't get the result I need.

Think of people going through all of these private conversations they've had in ChatGPT. I mean we know that it isn't fully private but it can still be a bit delicate. Hopefully they anonymize it at least.

I have zero knowledge of AI training, but is it possible that, soon, the line between training per se and fine-tuning will get technically so narrow that some groups can feed tens (or hundreds ?) of GBs of Copyrighted datasets [1] into the best open source AI models (LLMs or otherwise) ? I guess you could call this piracy 3.0.

[1] https://annas-archive.org/llm

Corporate models are already going that direction. Adobe’s image model is only trained on Adobe Stock, for example, and they and others are getting volumes of new images that artists have explicitly consented for use in an AI model to quell the goal post moving arguments about stock creators not imagining AI. Other models without that will exist too by then there is one world where this bill will just be providing protectionist favor to the former models, and the latter models will still exist in some more anonymous or offshore capacity.

Everything that happened in 2022 and 2023 is over, all of those discussions are outdated already, it’s too late to try to make levels of working groups and information gathering.

The future of training seems to, at least partly, be in synthetic data. I can imagine systems where a “data synthesizer” LLM is trained on open data and probably some licensed data. The synthesizer then generates data “to spec” to train larger models. MOE type models will likely have different approaches in so far as something like a Mathematical expert likely gets a long way with training data from out of copyright works by Newton, Euler, et al.

It's already how we fine-tune open source LLMs. All of them live off data exfiltrated from GPT-4. And it seems to help closing the gap fast. Microsoft had a whole family of papers on this idea: TinyStories, Phi-1, Phi-1.5, Phi-2...

Synthetic data has many advantages - it is free of copyright issues, the downstream models can't possibly violate copyright if they never saw the copyrighted works to begin with.

It is also more diverse and we can ensure higher average quality and less bias. It can also merge information across multiple sources. Sometimes we can filter using feedback from code execution, simulations, preference models or humans. If you can "execute" the LLM output and get a score, you're on to a self improving loop. LLMs can act as agents, collecting their own experiences and feedback.

I think GPTs are a ploy by OpenAI to collect synthetic data with human-in-the-loop and tools, to improve their datasets. This would also be in-domain for users and for LLM errors. They would contain LLM errors and the feedback. Very good data, on-policy. My estimations for 100M users at 10K tokens per month per user is 1T synthetic tokens per month. In a year they double the size of the GPT-4 training set. And we're paying and working for it.

But fortunately 12 months after they release GPT-5 we will recover 90% of its abilities in open source models.

> Synthetic data has many advantages - it is free of copyright issues, the downstream models can't possibly violate copyright if they never saw the copyrighted works to begin with.

I feel like we don't know if this is true or not. If we decide models trained on copyrighted data aren't fair game, it's possible we'll decide "laundered" data also isn't.

I mean, maybe that's not feasible. And I hope we don't decide training on copyrighted material is bogus anyway. But I don't think we know yet.

But also - you can totally violate copyright of something you never saw.

It's a matter of ensuring the synthetic content is different enough from the referenced content. We can filter.

Sure, but what matters for copyright is output, not input. For now.

If we make the (poor, imo) decision to prevent training on copyrighted data, that's a restriction on the training process, not on its result.

And in the world where we're making bad decisions to put legal restrictions on the training process, "can't train on data obtained by models that were trained without these restrictions" seems on the table.

If I'm a college student writing an essay I have to "put the ideas into my own words" to show that I understand them or face accusations of plagerism. LLM's will similarly have to be careful to use their own language to express the underlying ideas. Original ideas are extremely rare (e.g. theory of relativity) and any author will have a tough time proving that they had the idea first. The vast majority of writing is reusing ideas that have already been expressed. Even stories that appear unique are similar to stories that people have been telling for ages as Joseph Campbell explained in "The Hero with a Thousand Faces." The burden will be on the copyright holder to prove that the LLM steals their work after LLM's become as advanced as a college student putting into their own words.

The problem artists have with LLMs is not that they routinely output copyrighted works (it can happen, but isn't common), but that they copy human-expressed knowledge and art styles on an industrial scale, thus destroying revenue streams that supported or encouraged creative expression in the past. I doubt this is going to be solved via existing copyright laws. I suspect that the need for disclosure of copyright on the source material is to build a case for more novel legal protections for artists down the line.

> Copyright doesn't apply to ideas. It applies to works, which are specific, complete expressions of ideas.

I was attempting to say that in my first sentence.

> they copy human-expressed knowledge and art styles on an industrial scale, thus destroying revenue streams that supported or encouraged creative expression in the past

I currently have no opinion on art works as my focus is on knowledge and LLMs. I just did a search on Amazon for "child care" and found 40,000 books on the subject. I doubt the world needs 40,000 books on child care so if LLMs remove the revenue stream from these authors maybe they will spend their time writing books the world needs.

Only if you ignore case law. OpenAI seems to have done a lot of work around music—I now find it much harder to trick GPT-4 into outputting song lyrics verbatim.

So if it's a nonprofit that uses BOINC the law does not apply?

In a landscape ubiquitous with AI, what incentives do these companies have to remain separate rather than merge into a giant conglomerate?

Replace “AI” with “printing presses”, “electricity”, “general computing”, “off-shoring”, “online marketplaces”, etc. Variants of this sentiment have been expressed throughout history, but have always been looked upon as foolish in hindsight.

There will always be pressure to merge, but there is also pressure to not do so. A single technology or business strategy will not disrupt this dynamic.

If LLMs didn't have a nasty habit of directly regurgitating training data this would be a non-issue. Do I have this right?

And then companies outside the US or wherever these laws are passed will train without "copyright" burden.

Nothing stopping individuals from training on their own hardware either. HuggingFace Ceo's just said open source gpt4 level models coming this year.

True, I honestly think this is largely media companies making sure they (not artists) get their cut of everything, Spotify style.

Once data and logic get separated this will become less important. I mean there is no need to train data into the model, like it's done today. Ideally it should be a small generic model, or several of them working together, with quick access to external preprocessed data. This approach has a lot of advantages.

Sounds like what they tried back in the 60s.

There is nothing truly new in this world. It goes in spiral and that will be the next loop on upper level. Think of it, you can't train in the latest news in LLM. Once you finally understand it you need a solution. ChatGPT is using Bing. It's a clear separation, right? Is it the best possible solution? Obviously not. What is better.. it's a research area right now.

This will just create fly by night companies.

Not passed yet.

I wonder if there are provisions around synthetic materials generated from models trained on copyright data.

Heck, it hasn't even been assigned to a committee yet.

It doesn't even have a bill number assigned.

Reply on news.ycombinator.com