19

Generative AI can create images. Assume that these images are totally original.

Would that mean that we would have the full rights to the images created by the AI? Could we use these images in a product that we sell?

UPDATE

"AI Art" is possible thanks to Latent Diffusion Models, the most popular being DALL-E, Craiyon, Latitude Voyage, Midjourney and Stable Diffusion.

As of this writing, only Stable Diffusion is open source. Here is their declaration admitting that all work created with the tool grants all rights to the user of the tool (assuming you don't do illegal things).

enter image description here

Cybernetic
  • 467
  • 4
  • 11
  • Comments have been moved to chat; please do not continue the discussion here. Before posting a comment below this one, please review the purposes of comments. Comments that do not request clarification or suggest improvements usually belong as an answer, on [meta], or in [chat]. Comments continuing discussion may be removed. – Dale M Mar 27 '23 at 10:48
  • How it be otherwise? Do you not own the AI? Failing ownership, do you not a contract? – Robbie Goodwin Apr 05 '23 at 18:45
  • 2
    @RobbieGoodwin AI uses training data, and that training data is based on images that have copyright. So the question is, can the AI-generated image be considered an original work of art to which I now have ownership? This is new territory, and so far nobody really knows the answers. It will take a court case to produce the precedent that everyone else relies on going forward. – Cybernetic Apr 05 '23 at 19:21
  • @Cybernetic To suggest that AI comes into this is to misunderstand the idea of Copyright.

    If your outcome is based on copyright images, you fall foul of copyright and that's all there is to it.

    How could it matter what was achieved through people, dumb algorythms or AI?

    This territory isn't new; it's been well trodden since 'copyright' emerged, how long ago?

    FYI every published image has two forms of copyright; one for the content, one for the presentation. Whether they conflict is quite separate.

    – Robbie Goodwin Apr 05 '23 at 20:13
  • 1
    @RobbieGoodwin ALL inventions, of any kind, sit on top of things that already exist. There is no such thing as a purely original invention. The question is, does the combination of things used constitute an original work of "art." This is dealt with regularly in patent law. In the case of AI, it is not merely blending existing images, it is CREATING a never-before-seen image using elements of existing images. Imagine an artist looking at a photo for inspiration of their painting. Does the artist still own the painting? Yes. Understand now? – Cybernetic Apr 05 '23 at 23:42
  • With AI, exactly what do you believe is different from anything human?

    Where did you get the idea any invention must, let alone that all do sit on what already exists? Have you met anyone else who believes that, including your own teachers or students?

    Are you suggesting the wheel sits on the ski or travois? The sail on the oar?

    If right, you'd be describing patent, not copyright… neither the same nor, here, similar.

    Do you not accept '… works of art…' are outside patent law and solely to do with copyright?

    Original is not derivative. How is that hard?

    No? Understand yet?

    – Robbie Goodwin Apr 10 '23 at 21:36

2 Answers2

33

Per a decision of the US Copyright Office last month, AI generated images are not subject to copyright. That means you can use the generated images for any purpose you want1, but so can anyone else. However, the specific usage of a given image might be protected - so if you put a caption on the image and arrange it in the form of a comic (as the artist in that example did), that specific text and arrangement can be protected, but the underlying image can't be.

Laws may differ elsewhere in the world, but that's the current stance in the US.


1 Subject to any appropriate laws, including any copyright laws which the new image itself may violate. Just because the image isn't protected itself doesn't mean that it can't infringe on someone else's copyrights. See the other answer for more details.

Bobson
  • 655
  • 5
  • 12
  • 17
    Hwoever, the art can still infringe on copyrighted works - such as "draw iron man" would violate the copyrights of Marvel – Trish Mar 26 '23 at 20:44
  • 29
    @Trish No, it wouldn't. Only specific expressions of a character are protected by copyright, and the AI would produce entirely new ones. It might infringe the one of trademarks of Marvel, though, if they've trademarked Iron Man's likeness. – nick012000 Mar 27 '23 at 08:43
  • 11
    @Trish in addition to content, independent artists have voiced issues that their distinct style was reproduced by an AI tool. One instance of this was an artist who was banned from a subreddit because their original art was assumed to be AI generated. That's in addition to other controversies over how ethical it even is for the AI tool to have been trained on copyrighted image data or otherwise data that should not be publically available. – VLAZ Mar 27 '23 at 09:05
  • 19
    There is another controversy where the training image data includes private medical information. Overall, I find that a clear-cut answer "sure it's fine" is very problematic considering the multitude of issues including copyright issues like reproducing the Getty Images watermark. – VLAZ Mar 27 '23 at 09:10
  • 2
    @Trish I swear I remember characters being copyrightable but don't remember the details. The words "iron man" are not copyright (they would be a trademark) but the design of a guy in a robot suit with a big electric thing on his chest and ...etc... is already a copyrighted creative work, even if you draw him yourself in a different pose. IIRC. – user253751 Mar 27 '23 at 09:46
  • 2
    A potentially relevant addendum to this answer might be if editing the generated image (human contribution) results in a copyrightable product (a derived work?). And if so, what degree of human contribution or change would be required to achieve that result? – anjama Mar 27 '23 at 13:01
  • 1
    @anjama - That's a good question I don't know if there's been an answer to yet. – Bobson Mar 27 '23 at 14:34
  • 1
    I've added a caveat about the possibility the new image may violate copyright, but I don't think that users of the AI would be liable for any violations related to the training data. – Bobson Mar 27 '23 at 14:43
  • @VLAZ the original question included the direction to assume that these images are totally original, so the issue isn't with the answer, but with whether we can ever assume that. But still, given the current state, it does seem fair to remind here despite the fact this answer correct applied the question – JeopardyTempest Mar 28 '23 at 00:44
  • By italicizing artwork, you are indicating that there are (or may potentially be) other categories, like technical figures or such? Which I'm guessing may not have the same ruling clarity (or even a sharp distinction on what exactly classifies as artwork)? – JeopardyTempest Mar 28 '23 at 00:45
  • 2
    @JeopardyTempest - As opposed to other forms of media. This ruling was specifically about image generation - I don't know if there's a legal distinction between artwork and other images. I suppose the original question does ask about "images", though, so I ought to update my wording. – Bobson Mar 28 '23 at 05:37
  • @user253751: Titles, names, and ideas are not copyrightable, but characters somehow are, and a work can be found to be infringing even if the only points of similarity, other than the name, between an infringing character and the original are also shared by unrelated characters in myriad other unrelated works. – supercat Mar 28 '23 at 20:05
  • @VLAZ unless the law changes or courts change their minds, it's indeed fine. Someone's medical photos being in the dataset is a red herring in all of this. – JonathanReez Mar 29 '23 at 01:13
  • @nick012000 That is incorrect. Marvel’s copyright of the Iron Man comics includes an exclusive right to any derivative works—and a picture of the character by someone else is a clear derivative work, which means Marvel has the right to prevent the distribution of that image, and potentially to seek damages from any distribution it does receive. – KRyan Mar 29 '23 at 14:42
  • @supercat yes, a digital work can infringe copyright of another digital work even if the only similarities are the numbers 1 and 0 which are also shared by many unrelated works. This is not a productive line of thought. – user253751 Mar 30 '23 at 06:39
  • @user253751: In cases where digital works infringe other digital works, there would be an identifiable characteristic of one which is also present in the other, but absent from 99.9999999% of other digital works. – supercat Mar 30 '23 at 16:37
  • @supercat is that characteristic the 1 or the 0? – user253751 Mar 30 '23 at 16:46
  • Chat seems to think no such move has been made. Either way, what makes the Patent Office believe it matters whether the images are the product of human authorship? If that view were true, would it not suggest both that what my pet chimp paints cannot be copyright and, more seriously, that a company can own the copyright to its employees' works only by purchase and never, not being a human author, by mere right of contract? – Robbie Goodwin Apr 10 '23 at 21:54
  • 1
    @RobbieGoodwin Something your pet chimp paints is explicitly uncopyrightable. To quote from this page: "The animal's work was not an intellectual creation of the humans, and copyrights can only be held by legal persons—which an animal is not... In August 2014, the United States Copyright Office clarified their rules to explicitly state that items created by a non-human cannot be copyrighted" The company, on the other hand, is a legal person, and so can hold copyrights – Bobson Apr 11 '23 at 20:08
  • Do you not think anything the US Copyright Office says is subject to review by the Courts?

    Do you see, the 'monkey selfie' link has nothing to do with what any animal might paint?

    To show the real point, why not consider what happens if you or I copy (by any means) something our pet/slave has created and exhibit it as our own work?

    You've clearly stated, the creature had no copyright. Why, then would your or my publication not be considered legally original?

    – Robbie Goodwin Apr 11 '23 at 20:23
  • @RobbieGoodwin Comments aren't long enough for this debate, and aren't intended for it. Please ask a new question so you can get a proper answer. – Bobson Apr 11 '23 at 20:34
  • Good for you. Please consider the original Question or go to Chat or preferably, both… – Robbie Goodwin Apr 11 '23 at 20:36
10

The new work would not be protected by copyright law in the US. A different concern would be whether a created image might leave you liable in a copyright-infringement lawsuit. Although the output might be unique in that it algorithmically mashes two protected images into a new image, you might get sued for copying those underlying works. The input might just be text, not an image file, but that text is used by the program to search for relevant images, then copies the files (without permission). A person who unleashes a content-copying program on the internet is not immunized from an infringement lawsuit because "it wasn't me, it was my program". This is an issue that is being addressed in the courts right now (see this complaint).

On the other hand, if an image is created in a different manner, analogous to how you might verbally describe a desired image to a human artist, then there is literally no copying, and probably no legal infringement. I say "probably no infringement" because the output could end up resembling a protected work, so even if there was no actual copying, there could be legal infringement when the result is "strikingly similar" to a protected work.

user6726
  • 214,947
  • 11
  • 343
  • 576
  • 5
    "Although the output might be unique in that it algorithmically mashes two protected images into a new image" That's not how AI image generation works. It doesn't mash images together; it derives a set of statistical predictions based on a corpus of images it's trained on that are then used to create new images. – nick012000 Mar 26 '23 at 20:41
  • 2
    in a sense this is no different than patent infringement law. You can change a few things and call it your own, but the trick legally is to figure out what constitutes copying, or if it is indeed original. – Cybernetic Mar 26 '23 at 20:52
  • 9
    A big difference between copyright and patent law is that violating copyright requires copying. That implies access to the original. The AI program was presumably fed the original as input at some point. Patent infringement occurs whether or not the infringer has any knowledge of the original. – George White Mar 27 '23 at 01:15
  • 6
    @nick012000 in what sense is this not just a really advanced way of mashing images together? and if it is... here's a fun fact: every output is derived from every input. – user253751 Mar 27 '23 at 09:48
  • @user253751 the AI can both interpolate (which you might describe as "mashing together", though honestly that is a poor descriptor for interpolation) and extrapolate, which is emphatically not mashing together. – Clumsy cat Mar 27 '23 at 14:51
  • 3
    @Clumsycat Its extrapolation quality is stunningly poor compared to its interpolation quality, though. If you've got something visually-pleasing that isn't abstract and psychedelic, it's probably interpolated. – wizzwizz4 Mar 27 '23 at 15:09
  • @user253751 Because the images themselves aren't involved, just a statistical representation of the set of images, which is simply an un-copyrightable set of facts about those images. – nick012000 Mar 27 '23 at 20:23
  • Actually, Stephen Wolfram argues that since chatGPT uses randomness we should expect different outputs for the same inputs. This of course makes sense considering how it works. https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/ – Cybernetic Mar 27 '23 at 23:21
  • 1
    @nick012000 The pixel values of an image is also a set of facts about the image, yet it's copyrightable – user253751 Mar 28 '23 at 07:33
  • 1
    @user253751 It's complicated. The neural network doesn't store information about any particular image - it's more of storing which colors, line directions, juxtapositions, tend to correspond to which words. Which is science rather than art. – Therac Mar 28 '23 at 15:16