Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The shady world of Brave selling copyrighted data for AI training

stackdiary.com

Post reply on HN
Print view View original

91–100 of 127 posts

Re: The shady world of Brave selling copyrighted data for AI training

#91

Post by luma »

twoodfin wrote: post #56
_fbpp wrote: post #40

Earlier quoted context omitted.

The entire fair use claim is derived not from any legal basis, but rather, that "it has to be fair use" because it would be legally catastrophic for OpenAI et al if it weren't true. If you look at the core argument in favour of fair use, it's that "LLMs do not copy the training data", yet this is obviously false. For Github copilot and ChatGPT examples of it reciting large sections of training data are well known. Pl…

It actually doesn’t even matter if LLMs reproduce copyrighted data from their training. The issue is that a human copied the data from its source into memory for use in training, and this copy was likely not fair use under cases like MAI Systems . The Supreme Court hasn’t ruled on a software case like this, as far as I know. But given the recent 7-2 decision against Andy Warhol’s estate for his copying of photographs…

How does that align with Google Books scanning libraries full of copyrighted text, offering full reproductions of sections of the work, and then having the supreme court declare it all to be Fair Use? I think that is a far more relevant precedent here: https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Googl....

Re: The shady world of Brave selling copyrighted data for AI training

#92

Post by twoodfin »

luma wrote: post #91
twoodfin wrote: post #56

Earlier quoted context omitted.

It actually doesn’t even matter if LLMs reproduce copyrighted data from their training. The issue is that a human copied the data from its source into memory for use in training, and this copy was likely not fair use under cases like MAI Systems . The Supreme Court hasn’t ruled on a software case like this, as far as I know. But given the recent 7-2 decision against Andy Warhol’s estate for his copying of photographs…

How does that align with Google Books scanning libraries full of copyrighted text, offering full reproductions of sections of the work, and then having the supreme court declare it all to be Fair Use? I think that is a far more relevant precedent here: https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Googl... .

The Supreme Court declined to hear the case on appeal, which is a shade different from endorsing the decision after a hearing.

That being said, it doesn’t take a lot of effort to differentiate these cases. Google was indexing copyrighted works and providing access to limited extracts. They weren’t transforming them into new works and then selling access to those new works over APIs.

Re: The shady world of Brave selling copyrighted data for AI training

#93

Post by BeFlatXIII »

_fbpp wrote: post #49
kodah wrote: post #42

Unpopular opinion: the next iteration of privacy laws needs to factor in AI. If AI is allowed to slurp up PII or derogative works and the people defending it defend it with the zeal of cryptobros then we're in for a decade of real pain in terms of both copyright law, PII, and IP exposure.

The fun part is that the GDPR already does. The answer is you're not allowed to use personal data for AI . (And "personal data" here covers things like all public social media posts ) Facebook recently got told by the CJEU that, no, they can't use people's posts to target advertisements. Even if those ads are what's paying for the platform. That you can't claim such processing as "part of the contract" unless it is a…

End result: EU AI enjoyers use a VPN plus a US-based credit card borrowed from a friend.

Re: The shady world of Brave selling copyrighted data for AI training

#94

Post by richk449 »

kmeisthax wrote: post #83
_fbpp wrote: post #40

Earlier quoted context omitted.

The entire fair use claim is derived not from any legal basis, but rather, that "it has to be fair use" because it would be legally catastrophic for OpenAI et al if it weren't true. If you look at the core argument in favour of fair use, it's that "LLMs do not copy the training data", yet this is obviously false. For Github copilot and ChatGPT examples of it reciting large sections of training data are well known. Pl…

OpenAI's bias research on DALL-E revealed that most examples of regurgitation come from repeated copies of the same image in the training set. When they filtered out duplicates, DALL-E stopped drawing training examples. The problem is that filtering the training set is naively O(n^2) and n is already extremely large for DALL-E. For LLMs, it's comically huge, plus now you have to do substring search. I've yet to hear…

> A judge willing to commit the Butlerian Jihad[0] might even say that regurgitation does not matter and that all AI outputs are derivative works of the entire training set[1].

A judge can’t “commit” the butlierian jihad. A jihad is a mass event caused by some fraction of the population believing in some cause.

Which kinda gets to a point that seems to be missed. Copyright law is not “intrinsic” - nobody thinks that copyright is a natural law - it is just a pragmatic implementation which balances various public and private goods. If the world changes such that the law no longer does a good job of balancing the various goods, then either the law will get changed or people will ignore the law.

Re: The shady world of Brave selling copyrighted data for AI training

#95

Post by everforward »

ethanbond wrote: post #36
brookst wrote: post #33

Earlier quoted context omitted.

Do you think a human learning something from reading is fair use? Or are we all copyright violators because reading that article altered our connectomes, and we may recall parts of it later?

The point being raised is quite specific. Not sure if you’re willingly ignoring it or what? The answer is no, because you reading the article didn’t dramatically degrade its market value. An AI ingesting all content on the internet and then being ultra-effective at frontrunning that content for a large number of future readers does degrade its market value (and subsumes it into the model’s value).

This applies to so many things, though.

The most obvious parallel to me is YouTube. There are a ton of people ingesting books, then transforming that information into a roughly paraphrased video for people to watch for free (ish). That devalues the books they read and paraphrased, because other people don't need to read them.

Spark Notes devalue actual books in a way, because a lot of high schoolers read those instead of buying the actual book.

Search engines have also supplanted books in large part, because I don't need a whole book to answer a specific question. I don't know anyone that owns an encyclopedia anymore.

This is the next iteration of these processes. Non-novel information's market value has been degrading for decades now. A series of questions that would have cost thousands of dollars in books to answer in the 70's/80's is now free, with or without AI.

Re: The shady world of Brave selling copyrighted data for AI training

#96

Post by boondoggle16 »

nicce wrote: post #89
boondoggle16 wrote: post #85

Earlier quoted context omitted.

I do not use BAT or any crypto. Brave just works, and it blocks ads automatically when I tell friends to install it on their computers. I used to recommend Firefox, but Mozilla has totally jumped the shark (privacy violations [multiple], wastes too much money, blocks APIs that are useful with no real security risks while approving APIs with little use that do have security risks, etc, very user hostile). Chromium is…

> Chromium is obviously not trustworthy at this point, let alone Chrome. So that leaves like, Safari and Opera Brave is like 99% of Chromium + uBlock…

Right, I should have said "the main Chromium branch is obviously not trustworthy". It is possible to remove the untrustworthy bits, however, and there are a variety of de-googled Chromium builds.

Chromium is a great browser, unfortunately the official branch has been poisoned by Google.

Re: The shady world of Brave selling copyrighted data for AI training

#97

Post by pedrocr »

niemandhier wrote: post #69
pedrocr wrote: post #63

Earlier quoted context omitted.

It's not clear that "data mining" covers this use. These models are huge, big enough that they can just contain direct copies of copyrighted works. They've been shown to reproduce them relatively easily. The argument is that they've actually generalized enough or learned enough that they're now no longer the sum of the dataset. I can definitely see that being possible but the way the technology works it's really hard…

My reading of the relevant laws would actually lead me to believe that this is not a problem, as long as those reproductions are not returned and the eights holder did not opt out. But courts might decide differently. Regarding the copyright of returned material here is a good discussion: https://copyrightblog.kluweriplaw.com/2023/05/09/generative-...

That's clearly not enough. There's a continuum between producing exact input copy and having genuine creativity because the model actually learned something. A model that just reformats code and changes all the variable names would pass your test and yet be clearly a copyright violation. This whole argument requires that the neural network weights do something creative because they learned from the code instead of just transforming it. We're even careful about this with humans with things like clean room reimplementations to make sure.

Re: The shady world of Brave selling copyrighted data for AI training

#98

Post by kmeisthax »

richk449 wrote: post #94
kmeisthax wrote: post #83

Earlier quoted context omitted.

OpenAI's bias research on DALL-E revealed that most examples of regurgitation come from repeated copies of the same image in the training set. When they filtered out duplicates, DALL-E stopped drawing training examples. The problem is that filtering the training set is naively O(n^2) and n is already extremely large for DALL-E. For LLMs, it's comically huge, plus now you have to do substring search. I've yet to hear…

> A judge willing to commit the Butlerian Jihad[0] might even say that regurgitation does not matter and that all AI outputs are derivative works of the entire training set[1]. A judge can’t “commit” the butlierian jihad. A jihad is a mass event caused by some fraction of the population believing in some cause. Which kinda gets to a point that seems to be missed. Copyright law is not “intrinsic” - nobody thinks that…

Copyright is a unique case in which the law represents a bargain struck in the 1970s that hasn't been updated since. Everyone ignores it because it's nearly impossible to actually enforce copyright on individual infringers. But that doesn't mean copyright is meaningless: any activity which is large enough to be legible[0] to the state will be forced to bend itself to fit within the copyright bargain.

And AI training is extremely legible. This is not like a bunch of people downloading stuff off BitTorrent. All of the large foundation models we use were trained by a large corporation with a source of venture capital funding which could be easily shut off by a sufficiently motivated government. Weights-available and liberally licensed models exist, but most improvements on them are fine-tuning. Anonymous individuals can fine-tune an LLM or art generator with a small amount of data and compute, but they cannot make meaningful improvements on the state of the art.

So our sufficiently motivated copyright judge could at least effectively freeze AI art in time until Big Tech and the MAFIAA agree on how to properly split the proceeds from screwing over individual artists.

"Butlerian Jihad" is a term from a book, so you don't need to take "jihad" literally. However, I will point out that there is a significant fraction of the population that does want to see AI permanently banned from creative endeavors. The loss of ownership over their work from having it be in the training set is a factor, but their main argument is that they specifically want to keep their current jobs as they are. They do not want to be replaced with AI, nor do they want to replace their existing drawing work with SEO keyword stuffed text-to-image prompts.

[0] https://en.wikipedia.org/wiki/Seeing_Like_a_State

Re: The shady world of Brave selling copyrighted data for AI training

#99

Post by llamaimperative »

everforward wrote: post #95
ethanbond wrote: post #36

Earlier quoted context omitted.

The point being raised is quite specific. Not sure if you’re willingly ignoring it or what? The answer is no, because you reading the article didn’t dramatically degrade its market value. An AI ingesting all content on the internet and then being ultra-effective at frontrunning that content for a large number of future readers does degrade its market value (and subsumes it into the model’s value).

This applies to so many things, though. The most obvious parallel to me is YouTube. There are a ton of people ingesting books, then transforming that information into a roughly paraphrased video for people to watch for free (ish). That devalues the books they read and paraphrased, because other people don't need to read them. Spark Notes devalue actual books in a way, because a lot of high schoolers read those instea…

LLMs are attracting so much positive attention because they are likely to be a huge, huge step change improvement than all those methods you mention.

For that same exact reason, it’s totally reasonable they’re attracting unique amounts of negative attention too.

You can’t have it both ways: yes LLMs are going to change information retrieval the way nothing else has before, but no it’s actually just like all the other things in terms of their impact on incentive structures.

FWIW I don’t really know where I land on this issue. I just find it totally incoherent to believe in the bull case of “this will transform everything” while also portraying it all as par for the course when discussing potential negatives.

Just because Spark Notes didn’t obviously manage to kill valuable parts of our information ecosystem and economy does not mean that Spark Notes x 10,000,000 will not.

Re: The shady world of Brave selling copyrighted data for AI training

#100

Post by ethanbond »

cma wrote: post #87
ethanbond wrote: post #36

Earlier quoted context omitted.

The point being raised is quite specific. Not sure if you’re willingly ignoring it or what? The answer is no, because you reading the article didn’t dramatically degrade its market value. An AI ingesting all content on the internet and then being ultra-effective at frontrunning that content for a large number of future readers does degrade its market value (and subsumes it into the model’s value).

> The answer is no, because you reading the article didn’t dramatically degrade its market value. How about if you read a news article to write a competing one rewording and possibly citing it (one of the most common practices in news)?

How about it? Do you not think it incurs a lot of negative effects?
Post reply on HN