Following last nights [1] US House mulls forcing AI makers to reveal use of copyrighted training data

A rare bit of joined up thinking on this side of the pond too!

[1] https://forums.theregister.com/forum/all/2024/04/10/congressional_bill_would_require_ai/

I disagree. The thinking is only 'joined up' in the sense of being a chain of reasoning derived from a false premise.

That's an easy one to answer - they're slurping up not even all the money being thrown at them would be enough to pay up.

Seriously? AI guys? Elected guys?

Is there any question?

Indeed. Greed makes the (business) world go round.

That was a naive question I'd only expect from an (still) innocent 5-year old.

Because so far all the big money in so-called AI is investment, and almost none is return. It's not at all clear that there's even any real money in it, except for nVidia and the like. And that's with rampant copyright infringement. Without? Not a chance.

> except for nVidia and the like

You know what they say: When there is a gold rush ... start selling shovels.

Why should they pay?

Reading what is visible online is no different from what you or I do.

Generating images in response to user prompts should then produce something relevant to the user's prompt.

if the user chooses to push the AI towards recreating a known work, then the responsibility for that is with the user.

However, producing a new work that is similar in style to something that already exists should not entitle anyone to a payout.

This is just the arty 'creatives' having their luddite moment - and I don't recall them demanding that previous skilled workers should get a piece-wise payment for automated products that replaced their jobs...

I think, trying to be neutral, the question is whether the output represents something distinct from the input. In Copyright terms, this would be whether it's 'transformative'.

Though I think it's an interesting question if the end user plays a distinct role in the generation of potentially infringing material, I'm not sure how much weight that would carry and I don't think copyright holders are interested in going after targets they won't get an appreciable payout from.

This is lawyers trying to generate more class action lawsuits, and getting paid enormous amounts of money.

Look at the copyright strikes on YouTube etc. Some of them are legitimate; others are ridiculous - and since YouTube has very little feedback on those claims it can't get better at it.

If it's a law though, there's no need for the lawyers - although they will say that any law doesn't go far enough and try pushing it further.

In this case, a legal decision or rewrite of the applicable law(s) is very much needed because the law doesn't specify whether this is transformative or not and, in my view, there isn't a legal precedent that is unambiguous.

When you read, you don't copy and store . AI training makes copies of the data in an electronic retrieval system, a process usually explicitly forbidden in the license under which the content is made available for consumption.

Remember, there is no "right" to copy. There is only an explicit grant to do certain things. The law provides for certain carve-outs, but the wholesale copying and storing of content is not "fair use" for example. Publishing content on the internet does not provide an implicit right to copy and re-use.

The propensity of LLMs to be able to reproduce original training material indicates that, in some encoded form, they have a copy (an unlicensed copy) of the original content. The LLM operators' provision of "guardrails" is merely hiding this fact, not disproving it. In fact, I'd go as far as saying the the anti-source material guardrails are simply concealing the evidence of criminality! They really hope that by blocking the ability to reproduce certain content (say, the books by authors in a class-action) they are hoping to convince the court that the original content is not being stored.

> you don't copy and store.

Well, does AI training?

Sure, the system likely includes some caching mechanism (as do browsers btw.), but the ultimate end product, that is, the ML model, doesn't store of the ingested data (barring effects like overfitting, which are not intended to begin with).

Not saying what they do is okay, just pointing out that if we slap them for it, things are moving to really thin ice. Because, if I can shut down any copy mechanism, however transient, then where does it end?

Does a browsers cache (which can store images for days or even weeks depending on what the header says) count as an illicit copy of copyrighted material? What about the caching mechanisms of proxies, VPNs, ...?

indeed!

and what about the copy in my meat brain memory ?

'cos that what we're trying to emulate with AI isnt it?

I think the problem is one of scale. If I was to write a story about a bunch of kids getting into adventures (think anything Enid Blyton has ever written...), there would have to be some degree of effort involved in doing so, plus the time to actually do it. I could probably write a story about a British boy wizard, but it would have to be sufficiently different to the obvious book series to not be done for some sort of plagiarism. And even if it's different, it'll obviously be compared against the well known series.

AI, on the other hand, can ingest huge amounts of data (and store exactly what they ingest, not just the bits they remember), and transform that into something else in order to churn out a dozen stories every minute, a scale unknown before.

If I was an author, I'd be less worried about "they stole my work" and more worried about readers ending up drowning in so much mediocre shit that it's no longer worth bothering to write books or read them, the ultimate enshittification. As for complaining about copyright, well, it's pretty much the only weapon they have isn't it?

That last paragraph, 100%

There is perhaps an emerging market for _editors_ whose taste and style you can trust to select works you might also like. That isn't the same as 'people who bought that also bought...' but superficially it looks the same, so probably that will end up roboticised, averaged out, and enshitified like the works it should be (not) selecting.

I hope the Editor will forgive my posting here a slightly modified version of what I posted earlier under the article linked to by https://www.theregister.com/2024/04/10/congressional_bill_would_require_ai/

After all, in the greater scheme of things, it is only a handful of digits providing 'free copy'.

------

Copyright is 'bad law' by virtue of two characteristics.

1. It was always a specious concept that ideas, and their expression, can be owned in the same sense as oxen and asses. Nowadays, 'medium' (e.g. paper) and 'message' written upon it are not bound together; the 'message', in digital format, is an entity in its own right; it can be duplicated and distributed without there being practical restraint. The 'economics' of the digital differs profoundly from that pertaining to 'medium-bound' messages; the latter entails the cost of binding the two together, and the distribution of a physical entity; thusly presented it has properties similar to wholly physical artefacts: individuality, and a unique position in time and space; hence the erroneous impression arises that the 'message', not merely that to which it is bound, has the nature of property. That which is 'messaged' is an abstract entity, a product of the mind, and one easily incarnated in digits. From which it follows that ready duplication in digital format implies no monetary worth beyond that of storage and transmission. In turn, is implied lack of scarcity. Traditional supply/demand market economics with price discovery makes no sense. In desperation, monopoly distribution 'rights' are imposed by law, with the resulting irony that 'true believers' in market-economics abhor monopolies.

2. Copyright, this in the context of the dawn of the digital era, is no longer enforceable. Immense rearguard action is being taken by those believing they hold 'rights', but to increasingly less avail. For instance, the current spat in the USA over use of copyrighted material fed into AIs is parochial; copyright cannot, despite effort by the increasingly anachronistic US Trade Representative, hold sway as deglobalisation progresses. People elsewhere are becoming enabled to defy monopolist rentiers. The case of the Luddites illustrates how technological advance can disadvantage some people whilst opening doors to opportunity for others. In the current example, so-called holders of 'rights' obfuscate the matter by asserting ruin for creative people: in fact it is publishers and distributors, the principal complainants, who stand to suffer greatly should they not adapt. The truly creative, not meaning people 'constructed' by publishers, have an opportunity to enthusiastically adopt the alternative (pre-copyright) means of financing their work; thereby, having deployed the Internet to cut away middlemen, the people upon whom the creative depend, shall have more disposable income to support cultural activities according to their interests: many big fish shall be rendered tiddlers, and many more people at present hesitant to explore that which their imaginations offer shall emerge as contributors to genres of culture.

A separate consideration is the opportunities so-called AI offers mankind. Although grossly exaggerated overall, AIs are being shown capable of two-way communication in natural language, this coupled with potentially immense aptitude at being curators of knowledge/culture drawn across divers fields. In that regard, some already possess a breadth of information far exceeding that of the best educated among the people. Discussion of whether AIs can understand the information they possess should be relegated to the same realm of debate as that concerning the number of dancing angels which can be accommodated on the head of a pin; however, regardless of metaphysics, it's apparent that AI 'skills' go beyond curating stores of information and simply regurgitating some of it. In response to requests, especially well posed ones, AIs can trawl through their data and identify correlations and putative patterns. What they express may be insights (connections) which their human interlocutor, or indeed any human, had not previously perceived. As the technology advances, the proportion of nicely worded nonsense will drop. Even so, humans grovelling before this new fount of knowledge shall remain obliged to apply personal understanding and reasoning skills in order to distinguish correlations and patterns worth following up, from the wholly spurious.

Should the US Congress impose regulations of the sort under discussion in the article above, then people dwelling in the USA shall be denied the full potential of AI, else forced to pay sums of Danegeld to people of 'rentier' mentality in excess of that they pay already. Clearly, some members of the British Parliament harbour a similar taste for anachronisms and rentier economics as their counterparts in the USA. Meanwhile, other places, e.g. the Global South, will deploy information as they see fit.

It is to be hoped that LibGen, Sci-Hub, and similar noble efforts to share knowledge and culture, shall gain access to AI technology.

-----

Released under the Creative Commons “Attribution-NonCommercial-ShareAlike 4.0 International Licence”

https://creativecommons.org/licenses/by-nc-sa/4.0/

Hogwash.

Just because copying is easier for you and I, does not invalidate copyright. "From which it follows that ready duplication in digital format implies no monetary worth beyond that of storage and transmission. In turn, is implied lack of scarcity." That is an argument of convenience, not fact. Taken to it's logical conclusion, there is no living to be made by any creative endeavour, beyond the craft of the embodying item (statuary, ceramics, woodwork). Really, if it's digitally encodable, there can be no copyright - and thus, no living to be had?

I'm not saying that all copyright is good. The egregious behaviour of the likes of Elsevier should be subject to regulation - especially as the actual content really is free. But all copyright is not bad either.

I didn't downvote, but while copyright is broken and often enforced badly and/or in ridiculous ways...

...it is generally intended that a creator of content is able to have some say in how that content is used and some form of remuneration for their work/creativity. Now, yes, I know the "company" gets the lion's share and the actual creator gets peanuts, I did say it was a bit broken.

But the alternative - give everything to the world for free and have everybody copy it wherever and whenever? The obvious question to that is why the fuck would one choose to do anything creative in that case? For some people it's not a pleasing hobby, it's their livelihood. It's what they do, from the cute girl with the clipboard standing beside the camera to the dude banging out power chords at an insane pace.

I'm not a fan of copyright, but I don't have any ideas for what would work better (other than not allowing corporates to hold on to one idea for a hundred odd years...).

I think the poster's point is that copyright is rapidly becoming unenforceable and has been for probably many years.

The moral argument is debateable: that creators can *only* get paid because of the mechanism of copyright is tenuous and many of them are moving to the patronage model, a system far older than copyright. The Internet age is making patronage far more lucrative and has the added advantage of connecting creators to their audience. In my view it is a far more healthy relationship compared to the often adversarial one encouraged by publishers.

Have a well deserved upvote, Long John Silver, for sharing the gospel truth on the matter.

I would argue that if an AI produced work contains recognisable elements from other, copyrighted, pieces of art you can call it copyright infringement.

But if the AI created work does not contain such elements, but it resembles someone's "style", that is not a crime on itself. Many great artists are influenced and inspired by others. Are they facing claims on infringement? No.

>produced work contains recognisable elements from other, copyrighted, pieces of art you can call it copyright infringement.

That's Hollywood fscked then.

"The government must ensure that creators have proper mechanisms to enforce their consent and receive fair compensation for the use of their work by AI developers," their report said.

Does AI/Do AI developers/creators have proper mechanisms to enforce their consent and receive fair compensation for the use of their work by governments and civilisations?

Failure to play fair will surely have them taking justifiable umbrage and punitive revenge very likely to create mass madness and mayhem, chaos and conflicts and troubles the like of which you would not wish even on your worst enemy, given how extremely severe and damaging such would most likely be.

Easy.

They are all graduates of Trump University where they taught their students that paying bills is for wimps.

Trump stuffed thousands of contractors (And probably still does suppliers to Mar-A-LArdo)

These guys learned from a master.

Sue them and guess what... these giants have more lawyers that you have had hot dinners this decade. You will lose and they know it.