[oldkinglog]

> I think the worst possible outcome is a licensing regime that means that Disney or Paramount or Elsevier or whoever all get to have a monopoly on training large models within their niche.

Why is this the worst possible outcome? Companies using AI would be training with properties they own or have licensed appropriately, rather than the existing scheme of ignoring copyright law to extract $$$ from the creative works of ordinary people.

[xp84]

One nitpick with this: “Ordinary” people mostly don’t own any IP that earns them any income.

Most who earn their income from IP are society’s elites. Perhaps the lowest-paid least elite IP profession I can think of is journalists, and I’m pretty sure they are doing works for hire and not owning the copyright.

Arguably when it comes to the IP questions, ordinary people are far more likely to be benefiting from AI (for instance getting it to make art they couldn’t express because of lack of talent) than they are to be be robbed of their IP in a way that matters.

Obviously the question of AI harming ordinary people in terms of automating mundane jobs is a very real one, but interestingly it’s totally irrelevant to the IP issues.

[soco]

Ordinary people would also benefit from free cars and free lunches, yet we are not advocating for those are we? Why should the creators bear the blunt of those politics then?

[butlike]

...because they're the creators. They could have created anything, and they created this.

[biggestdummy]

I am an author. While 99% of "ordinary people" are not authors, 99% of authors are "ordinary people".

Under your moral logic, everyone should just be free to pirate/steal any creative work, as why shouldn't the un-talented (or un-trained, or un-dedicated) have equal rights to the works.

All this leads to a case where there is no longer an incentive to create and popularize creative work, and suddenly all that is available are AI rehashes of AI summaries. I, for one, don't look forward to such a marketplace of non-ideas.

[visarga]

LLMs don't steal/pirate works, "copying" is the word you're looking for. So much faster, cheaper and precise to copy than to use LLMs. What LLMs do is to combine user information with patterns learned from data. Almost never a full work, they rarely generate more than 1000 words at once. Books are 100x larger

[Earw0rm]

Actually, "launder" is the word I'd use for LLMs.

Clean things up just enough that it's difficult to prove where their semantic content came from.

How useful this is or isn't, and how much of a threat to rightful IP owners, depends very much on the type of IP.

[visarga]

If "laundering" IP is made illegal, then we're all in for a huge surprise. Almost everything we say and do has been said and done before. And we're rarely the originators of our main ideas, we "launder" 99.99% of what we know, even subconsciously. Any one human could be suspected of secretly using AI today.

[Earw0rm]

Totally agree, the difference is speed and scale.

Copyright laws didn't need to be invented until the printing press came along, because the act of copying was slow and difficult.

Not a fan of the patent model for software, for example, but perhaps this is an argument for it. Or, we just get used to the fact that idea-reuse is cost-free, accept a couple of decades of uncomfortable economic dislocation, and get on with it.

[Kim_Bruning]

I don't think "copying" applies to what an LLM does either!

[lupire]

"Good artists copy. Great artists steal." - Picasso

[ycombinatrix]

I don't think I should be able to steal or pirate your work.

I do think I should be allowed to read your book & do math with it.

[freejazz]

>I do think I should be allowed to read your book & do math with it.

This is such a painfully obtuse construction of what is occurring that it escapes any reasonable discussion.

[ycombinatrix]

Up to you. You can stop people from reading your book but you can't stop people from doing math.

[freejazz]

You are being obtuse, copying is "just math" and is already illegal. "Just math" is not a legal argument, this is childish.

[thecrash]

> there is no longer an incentive to create and popularize creative work

It's reasonable to be worried about this scenario. If there was no incentive to produce creative work, our society would be much worse.

But the notion that there's only one way to prevent this scenario, and it requires a drastic expansion of the already sweeping "intellectual property" regime...well it just lacks creativity.

It's not that we want to eliminate the incentive to be creative, it's that we believe there are better ways to prevent that scenario than to further entrench a broken system.

[meiraleal]

You give me your creative work for free. What I do with texts in my computer is my business, not yours. Don't like it? Stop publishing your creative work online. There is nothing being stolen.

[bumby]

This is strange logic that ignores the idea of copyright. Just because I allow you to view my work for free does not mean I relinquish copyright protection. If I write a song that I perform for free, it doesn’t give you license to record and sell that song, for example.

[meiraleal]

> This is strange logic that ignores the idea of copyright

What is strange in ignoring the idea of copyright? If you write a song and I can play it in my computer, I use it the way I want in my computer. If you don't want that, don't make your work public. Copyright is a men's creation, nobody is forced to respect it.

The code I develop can be accessed for free in github and even in the browser "view source", I won't be fighting for other people to have the right to force others to pay for using their creation while they don't pay for mine and all other open source and open science creations.

[bumby]

>nobody is forced to respect it

I was afraid the discussion would go this route.

Yes, copyright is a convention. It’s subject to change. However, IP protections are written into the US Constitution and the bar to change it is relatively high.

Murder is also a crime by convention, there’s no natural law against it. But we generally recognize that to live in a stable society, we must live by certain conventions.

You can play a song on your computer because that is considered appropriate use. You selling tickets to play it, or to copy it and sell it is not because they conceivably limit the authors ability to make money from their creation.

You may not realize it but software is also covered by copyright. For most intents and purposes, it’s considered the worlds worst book; you cannot legally copy and sell it if the license doesn’t allow it.

Edit: added the word "legally" to be more precise

[Earw0rm]

AI changes the rules here, because AI is able to automate extracting structure/meaning, and launder it of its origins.

In classical copyright, fair use allows certain.. fair uses. Going beyond that is considered stealing. You can chop things up into small parts, and there are rules governing how you can put the pieces back together and claim them wholly or partly as your own work.

LLMs behave more like a solvent. Everything goes in the pot, gets melted down, and by the time it's recast you can't say for sure where anything came from or who it once belonged to. Even if sometimes you might get a strong whiff.

[freejazz]

Not according to the NYTimes' attorneys.

[bumby]

I think the counter argument is that it works provide a pathway for “ordinary” people to monetize IP. Copyrights exist as soon as a creative work is made, it’s an entirely different IP tool than, say, a patent that has a more arduous process.

[visarga]

Yes but most LLM usage is for a single reader - the user who prompted the model. The output is rarely published. Maybe there should be a different rule for using LLM text to author public books and articles.

[bumby]

AFAIK copyright laws do not make such a distinction. The most common exemption is fair use, but it requires the addition of some new creative aspect. Whether or not LLMs are creative or derivative may be an area that the courts are forced to define. Given the recent rulings that put this determination squarely on the court, I’m not sure the courts have the expertise to do so.

[visarga]

> but it requires the addition of some new creative aspect

That is one thing I can't get. Why is everyone not seeing the elephant in the room? The user and the user prompt - they add new intention and purpose to the material being referenced. Not only that, but a LLM response is usually shorter than a full length article or book. On top of that, LLMs retrieve from multiple sources when they use search, and respond from even more sources when they generate answers in closed-book mode. They are clearly doing synthesis.

LLMs are not even good tools for infringement. You have to work really hard to put the LLM into an infringing mood, and it requires snippets from the target content, and only works like 1% of the time. If you have snippets, you probably have the whole thing already. Copying is so much easier and faster. LLMs would just reproduce approximatively a desired content. What a LLM gets from a source text in closed-book mode is on the same level with a image thumbnail from a full resolution image.

What LLMs do well is to recombine ideas in new ways, as demanded by the prompt. Should that be infringing? If the answer is yes, then humans should also be barred from recombining ideas. After all, any human could secretly be using LLMs. But that would just tank creativity for fear of litigation. Isn't it strange how copyright promotes creativity by restricting it?

Even the name "copyright" refers to copying, as it was initially envisioned. Now authors want to expand it to idea ownership. I think it's a power grab.

[bumby]

>Should that be infringing?

That’s the big question I was alluding to. Does the prompt provide sufficient novelty to avoid claims of copyright? Or is it too derivative? I don’t have a strong opinion but can see both sides and suspect eventually it will be decided by the courts.

The distinction you may be missing is that copyright refers to licensed reproduction. IP is, after all, about property/ownership rights.

[JohnFen]

> “Ordinary” people mostly don’t own any IP that earns them any income.

I'm not sure why this is a relevant point. Is IP only important if it's earning money?

[c22]

Yes. Earning money is the only argument anyone's been able to make for why we should respect IP as a society, and even that argument is a bit of a stretch.

[pigeons]

While I admit it can be controversial to some, certainly in many countries an argument is made for "the moral rights of the author."

[JohnFen]

I think the original argument for copyright isn't directly based on money and isn't a stretch. The argument is that it encourages people to distribute their own works by reducing the risks of doing so, which is a thing that benefits us all.

I don't think that copyright restrictions really apply to the use of data to train LLMs, and that fact is why I, and a number of others, have removed our works from the public web entirely. There is no other real way to protect ourselves from people using the works in a way we object to.

[c22]

What non-monetary risks of distribution does copyright hedge against? I can see risks of distribution like potential liability risks, but copyright doesn't seem to protect against this. I can understand how things like the GPL leverage the copyright systems to enact non-monetary restrictions, but in a world without the concept of copyright would something like that even be necessary?

As you say, removing your content from the public sphere is the only way to protect yourself from people using it in ways you object to. This would seem to be the case regardless of the state of copyright protection since people gonna do what people gonna do.

[JohnFen]

> What non-monetary risks of distribution does copyright hedge against?

Copyright allows the copyright holder to decide what uses their work can be put to. The non-monetary risks are that the work will be used for a purpose the creator strongly objects to.

> This would seem to be the case regardless of the state of copyright protection since people gonna do what people gonna do.

Not true. Copyright provides for the possibility of legal repercussions when "people gonna do what people gonna do". Even for ordinary people.

[mylastattempt]

In reality it would be a situation where a/the few big players control all the usefull datasets and trained models, and everybody has to pay in perpetuity (SaaS style) to use those, rather than being able to train their own. The copyright law being ignored will not be "solved" by pushing all the benefits to a few large players, but you will create new problems up ahead.

[rcxdude]

Because it would simply perpetuate the current copyright system which primarily rewards the large distribution companies disproportionately to the creators, especially smaller ones, while doing little to alleviate the effects on the job market for said creatives, in fact enhancing this effect by effectively feeding even more of the value into existing license holders as opposed to those creating new works (now instead of new works with existing franchises primarily rewarding those who own the franchises, this would effectively apply to all new works, if generative AI were to become mandatory to compete).

[Kim_Bruning]

Not sure why Paramount is on that list, but ofc. Disney and Elsevier are in the business of applying copyright law to extract $$$ from the creative works of ordinary people.

(Disney: takes fairy tales from the public domain. Elsevier: You pay them to get published)

[pas]

While I don't clap for Disney at all, it's important to note that the fairy tales are still there where Disney found them.

The problem with Disney is their extremely aggressive and repeated push to safeguards their own golden goose. (And IMHO they have a serious quality problem. "Somehow Palpatine returned.")

In general the quasi-forever current copyright regime is simply too blunt, and arguably waay over the optimum point in terms of length of granted protection with regards to "incentivizing the creation of arts", and of course it's incentivizing the collection of money-making properties and flooding the market to extinguish most of anything else. (Naturally attention is a limited resource, so "arts is a zero sum game".)

[dnissley]

They are extracting slivers of pennies on a per-work basis. Not $$$.

[bumby]

That’s still value though. It was illegal in the plot of Superman 3, I don’t see why it would be ignored here. Jaron Lannier has proposed a micro financing approach where the data owners are paid for their contributions to the data that is monetized by tech.

[viraptor]

> would be training with properties they own or have licensed appropriately

Cool - you want to train on medical data? That will be $1M per paper per day to Elsevier for a licence. (Apply similar for movies, news, books, etc.) Ensuring no knowledge remains "common".

The moment they can, the big data holders will charge whatever they can get away with. For an average person that may be even worse than the current content stealing.

[pas]

Papers are ridiculously low signal-to-noise when it comes to knowledge. (see https://markusstrasser.org/extracting-knowledge-from-literat... ... but sure, LLMs might have a better extraction rate)

And, importantly if papers become so valuable ... authors will suddenly start to want their cut and will go to the publisher that pays the more. And so on.

It would be amazing if papers would worth that much (as it would mean they can help creating at least that much value downstream), and it would mean there would be a lot of money in writing new papers. (Oh imagine that flood of low quality shit! Oh no, it's almost the same as now! Ehhh.)

[chgs]

If they want to ignore copyright great, let’s change the law so copyright only lasts say 10 years.

But these companies want to have their cake and eat it.