I think I have said on a few occasions that for instance a class-action copyright lawsuit against Copilot might not bear directly on AI art tools like Midjourney, to the extent that CoPilot apparently does tend to copy from its training set verbatim (and unattributed) whereas (I thought at the time) Midjourney doesn’t.
Well, it turns out that Midjourney does, maybe, to an extent. For maybe a few works?
The one that’s gotten the most attention is the 1984 photograph of Sharbat Gula by Steve McCurry, popularly known as “Afghan Girl“. The strings “afghan girl” and (haha) “afgan girl” are prohibited in Midjourney prompts at the moment. (“The phrase afghan girl is banned. Circumventing this filter to violate our rules may result in your access being revoked.”) And this is apparently because that phrase all by itself elicits what are arguably just slight variations of the original.
There’s a Twitter post that claims to show this, but I’m not certain enough it’s real to link to it. Also it’s on Twitter. But I can say that entering similar non-banned phrases like “young Afghan woman” also produce images that are at least quite similar to the photo of Gula, more similar than I would have expected. Given the size of the Midjourney training set, that image in association with those words must occur a lot of times!
(Update: it seems likely that the most widely-circulated image purporting to show Midjourney spontaneously generating close copies of the Gula “Afghan Girl” picture, is not actually that: it was made by giving the AI a copy of the original photo (!) and the prompt “afghan girl, digital art”. That the AI can make a copy of a work, given a copy of the work, is no surprise! Evidence, on a link probably usable only if you’re logged into Midjourney, is here. Given the further examples below, this doesn’t entirely undercut the point, but it’s interesting.)
The other example that I know of is “Starry Night”, which brings up variations of the van Gogh piece. This one’s out of copyright :) so I have no qualms about posting what I got:
Pretty obviously derivative in the usual sense. Derivative Work in the legal sense? I have no idea, and copyright law is sufficiently squishy and subjective that there is probably not a correct answer until and unless explicitly litigated, or the legal landscape otherwise changes significantly.
Are there other short phrases that will home in on a particular famous image? “Mona Lisa” (also out of copyright) certainly seems to:
Interesting and/or hideous variations, but still instantly recognizable.
What else might we try? “Migrant Mother” produces images that I think are clearly not derivative works:
Striking perhaps, ruined by the bizarre hands perhaps, in the same general category as the photo by Lange, but clearly of different people, in different positions, and so on. It’s not “plagiarizing” here, at any rate.
What if we tried harder? Let’s explicitly prompt with like “Migrant Mother photo, Dorothea Lange, 1936“. Whoa, yipes! Is this out of copyright? Well, if not it’s probably Fair Use in this posting anyway, so here:
Definitely derivative, and possibly Derivative. How about “Moon and Half Dome, Ansel Adams, 1960“? Well:
This is definitely not the picture that that search will get you in Google Images; if nothing else, the moon is way too large, and the top of Half Dome is a bizarre penguin-bill sort of shape. I’m guessing that this is because there are lots of other Ansel Adams pictures in the training set associated with words like “moon” and “half dome”, and mushing them all together quasi-semantically gives this set. The origin of the penguin-bill I dunno.
Maybe “Beatles Abbey Road cover, crossing the street“?
Crosswalk, front guy in white, roundish car to the left, check. Derivative in various senses, for sure. More specific prompting could presumably increase the exactness.
So I think we’ve established, to the extent of the tiny number of experiments I have the energy to do, that Midjourney (and, I would wager, other AI art tools, mutatis mutandis; I could get a Starry Night easily out of NightCafe, but not a Migrant Mother) can in fact produce images, the production of which arguably violates one or more of the rights of the copyright holder. It is most likely to do it if you explicitly try to do it (giving the most famous name of the image along with the artist and ideally the year and anything else that might help), but can also do it by accident (innocently typing “afghan girl”).
This doesn’t mean that these tools usually or typically do this; the fact that you can get a copy of an image from a tool that looks like it involves copyright laws doesn’t mean that other images made with it also involve copyright laws. To use the usual comparison, you can easily violate copyright using Photoshop, but that doesn’t suggest that there aren’t non-infringing uses of Photoshop, nor does it provide evidence that any particular image from Photoshop is infringing.
The easiest way to think about the blocking of “afg{h}an girl” from Midjourney prompts is that they have made a tool, realized that it could be used to violate copyright, and taken action to make it more difficult to use it that way in some cases.
This all bears on the question of whether images made with AI tools violate copyrights; the question of whether making the AI tools in the first place involves an infringing use is a somewhat different question, and we might talk about it some other time, although I’m still feeling kind of burnt out on the legal issues. But I did want to update on this one particular thing.
More updates: More stuff has happened in this area! For my latest deep and otherwise thoughts on the subject, I recommend the copyright tag on the weblog here.
ceoln 