Hacker News Reader

Comment by Animats | Hacker News Reader

Probably not. Humans are trained by reading books and writing something similar, but not identical. Unless what comes out looks a lot like a specific thing that went in, it's probably not copyright infringement in the US.

Artists making copies of famous pictures is a standard part of artist training. Tribute bands are a thing. Elvis impersonators are a mini-industry. The lawsuits so far tend to involve "passing off" new work as someone else's work, as with that ML-generated popular song passed off as a new work by a major artist. That's not a copyright issue. That's ordinary fraud.

The bitching from writers and musicians reflects not that they're being ripped off, but that they're being out-produced. Authors didn't expect to be in the position of John Henry vs. the steam hammer. Now they are.

View on HN

If I listen to 1000 pop songs and then write my own that's different from all of them, that's not copyright infringement. If I listen to 1000 pop songs and then write one that's nearly exactly the same as one of them, that absolutely is copyright infringement and I would most likely lose a lawsuit. The fact that I listened to 999 other songs does not absolve me of the infringement.

Most modern neural nets, and certainly LLMs, are overparameterized, which means they are completely capable of regurgitating exact training instances. They can and do repeat copyrighted training data verbatim. It's absolutely copyright infringement when that happens.

I anticipate AI researchers will soon come up with a way to ensure that outputs are a sufficient distance from all training inputs -- to essentially "block" direct regurgitation as valid output, with a tunable "max similarity" set to a safe distance. At that point we can start being more confident that our AI output is genuinely transformative and safe from (valid!) lawsuit.

> The bitching from writers and musicians reflects not that they're being ripped off, but that they're being out-produced.

This comes across as quite heartless. It will be interesting to see whether your view on this changes when AI starts making obsolete whatever you have spent your life specializing in.

I do not mind being made redundant by AI. I'll go become an electrician, general contractor, or something else you're not going to automate in my lifetime. The problem isn't the automation, its what happens if the usual Capital subjects capture all the gains vs them being distributed as excess across the populace.

EDIT: We should welcome automation and destruction of work where a machine can do it better. We fear it because of a broken socioeconomic system. Agree we should probably put breaks on anything so disruptive unless people impacted have some safeguards, because if you don't, what good is progress for progress' sake? It then becomes no different than bullshit jobs.

This is a critical distinction. Like 'toomuchtodo, I can go pick up a nailer and build a house if the computery jobs go away. But I cannot effectively participate in the governance of the society of which I am in theory an equal member if the benefits of automation continue inexorably to accrue to a very small group.

(e: and his edit says it as eloquently as I can!)

Highest praise. Appreciate the kind words.

This. My full-time job is automating myself out of a job. Arguably, that's one of the indicators of a quality software engineer versus one who can follow directions but not decrease the problem space.

The problem is not automation... Automation is great. It frees up people's time to do more activities. The problem is how we distribute the spoils of that freed time.

Your suggestion to someone whose multi-decade career was ended by AI is to play career whack-a-mole, try to guess which careers won't be next to be killed off, go back to school for several years, and pray?

Yes.

Telephone operators. Typing pools. Grooms. TV repair technicians. Weavers

To this list we may be adding law clerks. Drafters. Copy writers.

This is an ancient process. Some societies and cultures cope and maintain a civil society, some don't

So do not play "whack'a'mole", campaign for social change

I guess we should get rid of bulldozers and give more people shovels.

While you're at it, take computers always from people. They automate way too much stuff.

Btw, I am in one of those careers that people think is at the risk of being automated away. I don't really care about it. If I become obsolete, maybe my skills were not that useful. I'll try working in something else.

And if all jobs are automated away, I'll just go starve with the rest of the peasantry. The last to starve will be the first to suffocate.

Actually, a lot of songs are incredibly similar to older songs (see ed sheeran v marvin gaye), and they're not infringing.

Things will look very similar because there are only so many things that sound good to the mass appeal.

> I anticipate AI researchers will soon come up with a way to ensure that outputs are a sufficient distance from all training inputs

This is already a thing.

If you are basing your argument on a comparison between Ed Sheeran and Marvin Gaye, then you are talking about something completely different.

A better example is that Vanilla Ice had to pay royalties to Queen and David Bowie for using the bass hook from Under Pressure in Ice Ice Baby.

> a lot of songs are incredibly similar to older songs (see ed sheeran v marvin gaye), and they're not infringing

They're really not that similar. The Marvin Gaye family lawsuits are absolutely frivolous and a stain on his legacy. Suing over the same chord progression (ignoring tons of small variations that actually matter) would be like suing an author because they also had a character who was an angsty teenager with overbearing parents. Suing over a mostly-kinda-similar riff is like suing an author because they also used the sentence "Terrified, he ran." The songs only sound similar when you ignore all elements of the whole song and focus exactly on a few ways in which they overlap out of context.

They're nothing like GitHub CoPilot literally outputting comments that include copyright notifications from their original author. It's much more like suing someone who directly samples your song as a core element of their own (the Vanilla Ice example in the other comment).

> 1000 pop songs and then write one that's nearly exactly the same as one of them

Alright, but how do you define this when talking about most of modern "pop" music (using a broad definition that also includes things like techno or rock sub-genres)?

Maybe even only consider what's commercially relevant.

What about this aspect?

Creativity is also often about subtlety, e.g. polished songwriting and sound design.

Already without AI, rent-seeking is prevalent in popular music, maybe it's even an integral part?

For a stupid example of reusing themes in pop music, take this especially spicy example: https://en.wikipedia.org/wiki/Kernkraft_400

Or listen to the radio maybe. As a 90s and 2000s dance music lover, there's plenty of rehashing old content, sometimes licensed, other times it's just what music is.

Songs are memes after all

yeah, basically write an AI which would remove watermarks before training on data. Or after, if you've already trained (sarcasm). Great programmatic solution indeed.

Pop music is perhaps the worst example you could have chosen. It is highly derivative. https://www.youtube.com/watch?v=QpB_40hYjXU

> They can and do repeat copyrighted training data verbatim. It's absolutely copyright infringement when that happens.

My brain can do this too with music. In my 20's I used to be able to hear songs crystal clear in my head as if they were coming from the stereo itself.

Did my brain copy the song (yes!). Did it receive copyright approval from the artist (no!).

If you put that information into a fixed medium of expression can you be found guilty of copyright infringement? (Yes!)

> Most modern neural nets, and certainly LLMs, are overparameterized, which means they are completely capable of regurgitating exact training instances.

No they're not.

GPT-3: Model size: 175 GB, training set size 300 billion tokens.

GPT-4: Not disclosed.

LLaMA: Model size: 65GB, training set size, 1 trillion tokens.

These things condense the data. They don't just index it, like a search engine.

[1] https://lambdalabs.com/blog/demystifying-gpt-3

Also, having the same number of parameters as the training set doesn't mean you can memorize it

This is a bad take that doesn't account for the legal definition of copyright. The first step of training an LLM, going from the source to a tokenized input array is creating a derivative work. Unless that's something that can be shown as fair use (the most likely exception, but there are others), then ingesting items into an GPT is probably a copyright violation.

https://www.copyright.gov/title17/92chap1.html#106

106. Exclusive rights in copyrighted works

Subject to sections 107 through 122, the owner of copyright under this title has the exclusive rights to do and to authorize any of the following:

(2) to prepare derivative works based upon the copyrighted work;

https://www.copyright.gov/title17/92chap1.html#107

107. Limitations on exclusive rights: Fair use41

Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include—

(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;

(2) the nature of the copyrighted work;

(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

(4) the effect of the use upon the potential market for or value of the copyrighted work.

The fact that a work is unpublished shall not itself bar a finding of fair use if such finding is made upon consideration of all the above factors.

===

What will the courts decide? Who knows. It's probably going to be a fairly massive case regardless.

>ingesting items into an GPT is probably a copyright violation.

Copyright violation is generally on the output side of things, not the input. It's not a copyright violation to watch a movie.

I also would nit pick that a model is not BASED UPON a copywritten work, just because consuming the copywritten work changed the models parameters slightly. The model itself is not a derivative of any specific work. If one work represented over 50% of a model, it might be a different story. People are getting a little over focused on the word derivative and ignoring the rest of the sentence.

> Copyright violation is generally on the output side of things, not the input. It's not a copyright violation to watch a movie.

Copyright infringement is doing anything that violates a exclusive right of a copyright holder[1]. There are a bunch of exclusive rights one of which involves the output side. The Southern District of New York Federal court just sided with publishers against the Internet Archive specifically in deciding that scanning books was a copyright violation. [2]. Particularly of note is page 14 of the decision [3]

---

> The Publishers have established a prima facie case of copyright infringement. First, the Publishers hold exclusive publishing rights in the Works in Suit pursuant to 17 U.S.C. § 106, and the Works were timely registered with the Copyright Office. Pls.’ 56.1 ¶¶ 210-214. Second, IA copied the entire Works in Suit without the Publishers’ permission. See id. ¶¶ 242-245. Specifically, IA does not dispute that it violated the Publishers’ reproduction rights, by creating copies of the Works in Suit, 17 U.S.C. § 106(1); the Publishers’ rights to prepare derivative works, by “recasting” the Publishers’ print books into ebooks, id. § 106(2); the Publishers’ distribution rights, by distributing ebook copies of the Works in Suit to IA’s users, id. § 106(3); the Publishers’ public performance rights, through the “read aloud” function on IA’s Website, id. § 106(4); and the Publishers’ display rights, by showing the Works in Suit to users through IA’s in-browser viewer, id. § 106(5).5

---

The generation of the model literally requires a step which converts the source material into tokens. That step is infringing. It's no different to taking photocopies of a photo, or whipping my phone out during a movie at the theater to record. Whether it's infringing is not the issue. It is. It's whether the act is fair use that is the problem.

I guess what you're trying to argue here is that using the resulting model is non-infringing. That's a whole 'nother thing I guess.

---

[1]: https://www.copyright.gov/title17/92chap5.html

[2]: https://en.wikipedia.org/wiki/Hachette_v._Internet_Archive

[3]: https://www.eff.org/files/2023/03/29/188_opinion.pdf

The internet archive case was about distributing the copies. The output side.

Had they archived the books and not lent out unlimited copies, nothing would have come of the case.

Probably yes. It still holds the book inside the model (or parts). For generative art it spits out watermarks, for example.

If it was just some code, included into a compiled EXE, without permission - it would be a major violation. But including books inside of an AI model - it is different (not really).

Reply on news.ycombinator.com