Przejdź do głównej zawartości


When it comes to AI art (or "art"), it's hard to find a nuanced position that respects creative workers' labor rights, free expression, copyright law's vital exceptions and limitations, and aesthetics.

--

If you'd like an essay-formatted version of this thread to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:

https://pluralistic.net/2024/05/13/spooky-action-at-a-close-up/#invisible-hand

1/
An old woodcut of a disembodied man's hand operating a Ouija board planchette. It has been modified to add an extra finger and thumb. It has been tinted green. It has been placed on a 'code waterfall' backdrop as seen in the credit sequences of the Wachowskis' 'Matrix' movies.
An old woodcut of a disembodied man's hand operating a Ouija board planchette. It has been modified to add an extra finger and thumb. It has been tinted green. It has been placed on a 'code waterfall' backdrop as seen in the credit sequences of the Wachowskis' 'Matrix' movies.
Ten wpis został zedytowany (2 lata temu)

Ostrzeżenie o treści: Long thread/2

Ostrzeżenie o treści: Long thread/3

Ostrzeżenie o treści: Long thread/4

Ostrzeżenie o treści: Long thread/5

Ostrzeżenie o treści: Long thread/6

Ostrzeżenie o treści: Long thread/7

Ostrzeżenie o treści: Long thread/8

Ostrzeżenie o treści: Long thread/9

Ostrzeżenie o treści: Long thread/10

Ostrzeżenie o treści: Long thread/11

Ostrzeżenie o treści: Long thread/12

Ostrzeżenie o treści: Long thread/13

Ostrzeżenie o treści: Long thread/14

Ostrzeżenie o treści: Long thread/15

Ostrzeżenie o treści: Long thread/16

Ostrzeżenie o treści: Long thread/17

Ostrzeżenie o treści: Long thread/18

Ostrzeżenie o treści: Long thread/19

Ostrzeżenie o treści: Long thread/20

Ostrzeżenie o treści: Long thread/21

Ostrzeżenie o treści: Long thread/22

Ostrzeżenie o treści: Long thread/23

Ostrzeżenie o treści: Long thread/24

Ostrzeżenie o treści: Long thread/25

Ostrzeżenie o treści: Long thread/26

Ostrzeżenie o treści: Long thread/27

Ostrzeżenie o treści: Long thread/28

Ostrzeżenie o treści: Long thread/29

Ostrzeżenie o treści: Long thread/30

Ostrzeżenie o treści: Long thread/31

Ostrzeżenie o treści: Long thread/32

Ostrzeżenie o treści: Long thread/33

Ostrzeżenie o treści: Long thread/34

Ostrzeżenie o treści: Long thread/35

Ostrzeżenie o treści: Long thread/36

Ostrzeżenie o treści: Long thread/37

Ostrzeżenie o treści: Long thread/38

Ostrzeżenie o treści: Long thread/39

Ostrzeżenie o treści: Long thread/40

Ostrzeżenie o treści: Long thread/41

Ostrzeżenie o treści: Long thread/42

Ostrzeżenie o treści: Long thread/43

Ostrzeżenie o treści: Long thread/44

Ostrzeżenie o treści: Long thread/45

Ostrzeżenie o treści: Long thread/46

Ostrzeżenie o treści: Long thread/47

Ostrzeżenie o treści: Long thread/48

Ostrzeżenie o treści: Long thread/49

Ostrzeżenie o treści: Long thread/50

Ostrzeżenie o treści: Long thread/51

Ostrzeżenie o treści: Long thread/52

Ostrzeżenie o treści: Long thread/eof

@timoreilly @internetarchive @AuthorsAlliance Many thanks for this elaborate discussion which is painful to grok in its consequences. Interestingly the article only deals with copy**right**. The hotter case for me is copy**left**. I choose #GPL or #LGPL because I want derivative work to maintain the same conditions. So despite recent rulings about copyright failure, how scrapers use my copylefted code does not explicitly transcend the required copyleft into the derivative work.
@timoreilly @internetarchive @AuthorsAlliance This also extends to the Creative Commons siblings of GPL/LGPL. The issue I see here is that those large corporations do not need to win the court case. It is sufficient they have enough money to keep the thing long enough in court to have me financially exhausted. This is what happened to patents. They are useless for a startup by now.
@markuswerle @timoreilly @internetarchive @AuthorsAlliance

No, copyleft licenses don't trump fair use,, de minimis, and other limitations and exceptions.

CC, GPL, etc are licenses for things you need permission to do.

Fair use are things you don't need permission to do, so you don't need to license those uses, so you aren't bound by sharealike or other clauses.
@timoreilly @internetarchive @AuthorsAlliance So your opinion is that GPL code should be allowed to be used for training models that are used in commercial products? Can you elaborate on this from 2 perspectives?

1. Current legal situation. Do you think the legal setting is clear and covered by fair use? Note that my complete code is scraped.
2. Possible future legal situation that pursues the idea that you must not make money with derivatives of my work
@markuswerle @timoreilly @internetarchive @AuthorsAlliance

Neither of these apply. The first third of my essay explains in detail why model training isn't infringing, so I won't rehearse that here.

If it's not infringing, then it doesn't require a license to undertake.

If a user need not license a work to make some use of it, they need not abide by license terms, either.

You're having a labor issue, and you're trying to solve it with copyright, and it won't work well or at all.
@internetarchive@archive.org @AuthorsAlliance The point is that in my perception it is not "some use of it", but the whole of my copylefted work is transcended into the derivative work. So despite my work only changing a fraction of the AI model 100% of my work was used as a contribution. I still have to understand how this can be fair use.
Ten wpis został zedytowany (2 lata temu)
Busting someone who takes 0.0000033% one gazillion times in a row might still be a criminal. Imagine someone who is hacking into an online bank and stealing 0.0000033% of each bank account.
@markuswerle The law regarding criminal theft isn't copyright law. Every word in every book appears in another, copyrighted book. If taking enough single words from other books added up to infringement, no more books would ever be possible.
I am not totally on the same page here. Extracting single words is the extreme case, so I do not buy the argument yet, especially not for more complex artwork. I may have to think more about it.

We may face the similarity problem envisioned by Kevin Kelly in "The library of Forms" https://kk.org/mt-files/outofcontrol/ch14-a.html
I think you miss two key points in here.
1. No mention of "commercial impact" prong of fair use - LLMs directly diminish the market for the original works (esp. stuff like promoting midjourney with "in the style of X").
2. The memorization problem isn't easy - it's impossible. LLM providers today are accounting for it on a vanishingly small number of potential outputs, and it's not by altering the models (they basically can't). It's by slapping output filters on at the end.
@bensomers

Commercial impact wouldn't apply to the model, it would apply to the *use* of the model.

In Betamax, SCOTUS held that a tool doesn't infringe provided it can "sustain a substantial non-infringing use." There is no question that an LLM can sustain a substantial noninfringing use, e.g., analyzing police reports for New Orleans Innocence Project.

A user of an LLM might infringe copyright, but the LLM itself doesn't.
@midnightfire You're missing two things:

1. It's either impossible or nearly impossible to make a copyright principle that bans AI and doesn't break collage; and

2. Such a law wouldn't prevent artists' bosses from trying to fire us and replace us with AI, because they want to fire us and they have huge corpuses of copyrighted works they're *already* licensing for AI training *for the express purpose* of firing us.
@midnightfire The *very* best we could hope for from a new copyright law that bans training is:

a) An enormous amount of collateral damage inflicted on unrelated activity, including, but not limited to collage;

b) A minor shift in the relative distribution of misappropriated creative wages, from the tech companies that hate us and want us to starve to the entertainment companies that hate us and want us to starve.