Copyright Office suggests AI copyright debate was settled in 1965
arstechnica.com
arstechnica.com
What this article doesn't talk about at all is the far more controversial AI copyright debate, the one most people will think of given the headline: whether training a model is fair use. That's the one everyone is actually concerned about, and they're definitely not claiming it was settled already.
To the best of my knowledge, there is no direct precedent from any federal circuit addressing this issue and certainly no USSC opinions dealing with the issue. Additionally, any analogies drawn from precedent focused on other areas of intellectual property law is easily distinguishable. This is truly fresh legal ground and the next 10 years of jurisprudence will go a long way towards building the precedent that your comment would imply to already exist.
Just to be explicit, the above, while a legal opinion, IS NOT legal advice.
Arguably the material has been learned and not copied. Maybe in some cases learned with an uncanny ability or photographic memory, but learned. (People with photographic memories also cannot reproduce content in an unlimited fashion).
There's nothing special about an LLM, there's no learning, and they regurgitate verbatim text too.
May as well say curl + images in a db are learned as well, so thus I can use Mickey Miuse as I please in my php web page.
Let’s say that the model “is influenced by” the copyrighted material. That seems hard to argue against.
So, now that we aren’t using the word “learned”, why would we say that the way the models are influenced by the copyrighted works that appear in the “training set” (not to imply that “training” in the usual sense is happening) counts as a copyright violation?
Or, perhaps the claim is that the outputs of the model are violating copyright?
If the output is substantially similar to some particular copyrighted work that is in the “training set”, and could work as a substitute for that work, and if the output resembling the work is in part due to the influence that the work had on the model, I think in this case it would be a clear case of violation of copyright.
However, if it doesn’t have substantial similarity to any particular copyrighted work that influenced it, only similarity to the style common to many of the works that influenced it (even if all by the same author), my impression is that this would not constitute copyright infringement because styles are not protected by copyright, only individual works.
(Now, is this unjust, in the case of it copying the style of some particular author/artist? Idk, maybe? But my impression is that copyright doesn’t protect styles, and that it probably shouldn’t protect styles in general… so I guess maybe if we had a law making a special case forbidding the (deliberate?) copying of a person’s particular style via some kind of machine learning model? Idk.)
The model is 100-500x smaller than its training set. That is something hinting at learning, as direct storage is impossible.
Neither your comments to HN nor anyone else's strenuous assertions that there's anything to debate are going to change anything.
If you want to treat LLMs as a special case—which is what you want, since there is an entire history of jurisprudence that you have to contend with here—then you need to get Congress to write legislation that says so.
More than that, a constitutional amendment. See Feist vs. Rural Telephone.
The US does not have database copyright or "sweat of the brow" copyright. There has to be human originality.
The AI itself can be construed as a special kind of a database, given that it can be queried to reproduce at least part of its training dataset with precision...
No.
(That’s what the NYT tried to argue/show. I’m not sure if that case is ongoing or settled.)
Similarly with patents.
Even when used there should be a timeout, possibly per clause to avoid overly broad stuff.
But then a few IP trolls and lawyers would have to find another job.
---
"Congress shall make no law respecting an establishment of religion, or prohibiting the free exercise thereof; or abridging the freedom of speech, or of the press; or the right of the people peaceably to assemble, and to petition the Government for a redress of grievances."
---
The rest of the bill of rights looks similarly. Now a days it'd be thousands of pages long trying to elaborate endlessly upon every single scenario. But the more important thing is that this idea of trying to codify every scenario still doesn't work because you end up with a zillion loopholes in just about every single law with some clever clown going 'ah hah! you didn't cover this!' So all you really do is end up with laws that are not only excessively fragile and subject to exploitation, but also completely indecipherable by just about anybody, certainly including the people voting on their passage.
Basically (speaking from a US perspective) we’re doomed either way.
I'd also add that the more simple laws add transparency. If a e.g. judge does something not in the apparent spirit of a older law, it's apparent to anybody with a mind. But when instead an offender (or judge) relies on a nuanced, if esoteric, interpretation to a loophole in section 274, subsection 74 paragraph 13a, nobody has any clue what's happening.
And if we're screwed either way, it'd be nice if people could at least visibly and explicitly see that they're being screwed.
Fair enough but does that help settle the other question, which is whether weights are considered derivative works of any material used in the training?
I want to highlight that training the model is only one part of the copyright questions going on, the other is how they are making and keeping direct long-term copies in giant training datasets.
Imagine what would happen if a regular person bought and then immediately resold thousands of books, CDs, and movies, taking just enough time to make a copy of each one and building out their own library/movie-theater for friends and coworkers. You think the powers-that-be would let you or I get away with that?
There is no (non-evil) reason to hold multijillion dollar corporations with professional legal advice on-tap to a lower standard than regular people.
Yes, training does impart some fraction of an article into the weights. NYT famously “demonstrated” this by like typing whole paragraphs of their articles into GPT and having the model produce some of the following sentences. However, this substitutes for the article in zero ways, since if you need the article to summon the article…who cares?
We should admit that nothing about our copyright law intentionally weighed in about LLMs. It’s simply nuts to apply a law to a situation when its drafters had no idea of the positive or negative implications of such an application. It would be like applying a law that calculated prison time based on number of horses stolen to someone who stole a Honda Civic with 100hp and saying that clearly they should get 100 years because it is equal to 100 horses.
Now, I get that we have a useless legislative branch which even if they actually passed intentionally applicable laws, they’d be stupid ones, but I think making simplistic analogies like that do not help anyone, other than the Luddites. Look, the cat is out of the bag and even if say, the US government effectively killed all Gen AI by forcing any training to be done with material you own the copyright of, countries like China (and criminals, who can simply use the tech in secret) will happily just pull ahead of us and economically demolish us - like any country would have done if a competing country had banned electricity 100 years ago.
We need something better than the ridiculously unfit 200-year old paradigm for this.
This policy is sensible. Most AI generated works should be uncopyrightable, except where a substantial human contribution is in the output.
Simply describing a picture and letting AI generate it shouldn't be enough for the same reason that dictating what you want to a painter isn't enough to earn you copyright over the resulting painting.
I would be wary about integrating too much AI output into works one wants to enforce copyright over without some level of documentation. The nightmare scenario is having your copyright stripped away because of evidence one used AI extensively.
They just can't get it from the government because they are not, in any sense, the author of the creative work.
But what about the other direction, can distributing such AI generated content VIOLATE somebody else's copyright?
If output of AI cannot be copyrighted, can it violate copyright?
Interesting take, and I've heard this many times. I'm curious to explore this further and see why you think that is, and where do you draw the line?
Is it the "low effort"? Is it the "automated" stuff? Is the process of setting it up, prompting it and choosing a result not enough "creative input"? If so, why?
Let's take a "real world" example as analogue. Say I setup a camera on a tripod. I set it to take pictures every 1 second, and leave it there. Come back 1hr later, and go through the pictures. I select one of the sunset I like, and post it. Would I not have copyright on that picture? I wasn't there when it was taken. But I did setup the camera and selected the end result. How is that different?
Taking it back to genAI, say I build/train/finetune my own model. Would it now have enough "effort" from me that I can use those generations? Is this an effort thing or is it more? Or is it just that someone else did the work?
What about random "art"? As in art based on random numbers. Say I write a script in python to use random math formulas to "draw" on a canvas. I let it run for a couple of hours, come back, look at the results and select one. Do I not get copyright on the resulting "art" because it was randomly generated by a script? Does it matter if the script was written by me? Would it be different if you download my script and generate the art yourself? Would you not have copyright?
I guess what I'm trying to say is "where do we draw the line?". It's not clear to me why people say "simply prompting and selecting isn't creative enough". This distinction wasn't there before. Plenty of "art" out there based on random processes + curation. Why the sudden change?
Other kind of binaries which are fully generated by a machine like private keys aren't copyrightable.
1. Can I get sued by a 3rd party when using AI generated work in my project?
2. Can I sue a 3rd party when they use my AI generated work in their project?
So any book or story on Amazon that was generated substantially via prompting should now have to be removed based on this guidance from the copyright office.
But in the US, Feist v Rural explicitly disavowed the sweat of the brow doctrine, and said that facts have no copyright value--a work requires a quantum of original creative spark to be copyrightable (it was discussed in the context of phone books--a phone book does still have some residual "thin copyright", but the listing of phone numbers is not copyrightable, and it is actually difficult to infringe on the thin copyright of a phone book). In the US, your example would easily be found to be not infringing, if the only similarity were reproducing the same 100 facts.
"Elephants are enormous mammals, usually grey in color, with significant intelligence and social habits. They are native to portions of Africa and Asia, with different elephant species found in each region. They are famous for their strong and dexterous trunks, which can also be used to communicate something like a trumpet. Humans, especially in South Asia, have long admired elephants and used them for transportation and various kinds of work; African elephants are famous for having been ridden to war by the Carthaginians, to the dismay of their Roman opponents. Today elephants are significantly threatened by various human activities, both those intentionally directed at the elephants (like killing them for meat or for the ivory derived from their tusks) and those not deliberately meant to affect them (like deforestation). While our word for elephants comes from the Greek, it was probably borrowed by the ancient Greeks from another language family."
I just wrote this paragraph about elephants based on my own knowledge. There is no copyrightability of the substantive information here (e.g. if you learned something new about elephants, you can tell other people) but there is probably some copyrightability of the paragraph based on things like the creativity of my word choices. That distinction can sometimes create confusion in discussions about "facts", and I'm not positive that legal standards that are meant to clarify it have always given a clear and workable rule.
Largest Land Animals By Size:
African Bush Elephant
Asian Elephant
African Forest Elephant