Imagine That IP Law Blog

Newly developed artificial intelligence systems have been used to generate new inventions (e.g. Dr. Stephen Thaler’s “Device for the Autonomous Bootstrapping of Unified Sentience” or DABUS), works of authorship (e.g. Alice and Sparkle, created by Ammaar Reshi using the ChatGPT large language model), and works of art (e.g. Zarya of the Dawn, created by Kristina Kashtanova using the Midjourney image generator). The U.S. Patent and Trademark Office and the U.S. Copyright Office have each held that artificial intelligence systems cannot qualify as inventors or authors for the purpose of patent or copyright registration (See, e.g., “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” 88 FR 16190 (U.S. Copyright Office, March 16, 2023) and “In re Application No. 16/524,350” (USPTO, April 22, 2020), affirmed in Thaler v. Hirshfeld, 558 F.Supp.3d 238 (E.D. Va., Sep. 2, 2021), subsequently affirmed in Thaler v. Vidal, 2021-2347 (Fed. Cir. Aug. 5, 2022)). While the Copyright Office’s new guidance states that works “containing AI-generated material [may] also contain sufficient human authorship to support a copyright claim,” and that the work as a whole may be registrable, “[w]hen an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright...” (88 FR at 16192).

So, the AI system can’t be the author or inventor. What about the person who set up the system? Artificial intelligence pioneer Dr. Stephen Thaler argues no in his Supreme Court petition appealing the Federal Circuit’s decision in Thaler v. Vidal, noting that 35 U.S.C. §115 requires that one named as an inventor on a patent application must “believe[] himself or herself to be the original inventor or an original joint inventor of a claimed invention in the application,” and that someone who merely “trained [an AI system] and provided it with information” could not be recognized as an inventor. Petition for Writ of Certiorari at 29-30. Similarly, in its letter reviewing Zarya of the Dawn, the Copyright Office noted that “the images in the Work that were generated by the Midjourney technology are not the product of human authorship” because although Kashtanova “‘guided’ the structure and content of each image,… it was Midjourney—not Kashtanova—that originated the ‘traditional elements of authorship’ in the image.” (“Letter re: Zarya of the Dawn (Registration # VAu001480196)” at 1, 8 (U.S. Copyright Office, Feb. 21, 2023).

If neither the AI system nor its operator hold rights over the material, are these works simply public domain? Not necessarily. While the systems and their operators might not be able to claim ownership of their creative output, they’re not the only ones with a potential interest.  A new challenger has entered, in the form of the original authors and artists on whose work these AI tools were trained. In a class action suit filed in January, Andersen et al. v. Stability AI, Inc. et al., artist plaintiffs have raised claims including copyright infringement, rights of publicity, and unfair competition (Complaint at 30-42, 3:23-cv-00201 (N.D. Cal., Jan. 13, 2013)). Similarly, in February, Getty Images filed suit against Stability AI, Inc,. alleging copyright and trademark infringement, as well as trademark dilution (Complaint at 23-34, Getty Images Inc. v. Stability AI Inc., 1:23-cv-00135-UNA (D. Del., February 3, 2023)). These claims rest on several novel and, of yet, untested theories.

Artificial intelligence-based text-to-image generators such as Stable Diffusion and Midjourney use images and associated captions scraped from webpages to build a training set, frequently including billions of images. A model is trained from the training set, and when provided with tokens from text prompts and an initial randomly generated noise image, the model repeatedly filters the noise to create a coherent image. The Andersen plaintiffs characterize this process as one of building a collage from source images (Complaint at 14-15, 20), and also compare it to lossy compression techniques used in media encoding, including for MP3 audio files or JPEG image files (Id. at 17). This isn’t quite correct: while the training data set, frequently petabytes in size, includes the source images and associated text, the trained model is typically only a few gigabytes at most and sometimes smaller. The model itself is a dense array of “latent vectors,” encoding the training data in a high-dimensional space, and does not include any of the original images. Even equating diffusion encoding with compression (and with a near-magical algorithm with a compression ratio greater than a million to one), the original images cannot be recovered. However, similar images can be created from properly crafted text prompts.

While creating new images from a trained generative AI model does not require copying any source image or “creating a collage” of parts of existing images, as discussed above, the training data for the model is scraped from billions of web pages – literally copied into a training database. However, this act of copying may not constitute infringement under the fair use doctrine (codified at 17 U.S.C. §107). In Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015), the court found that Google’s “Google Books” project that involved adding digital copies of millions of books to a publicly searchable database was a transformative use protected by the fair use doctrine, and therefore was not infringement. In particular, the first factor considered under 17 U.S.C. §107 is the “purpose and character of the [accused] use”. In Campbell v. Acuff–Rose Music, Inc., the Supreme Court described this factor as a determination “whether the new work merely ‘supersede[s] the objects’ of the original creation… (‘supplanting’ the original), or instead adds something new, with a further purpose or different character… [I]t asks, in other words, whether and to what extent the new work is ‘transformative.’” (510 U.S. 569, 579 (1994) (internal citation omitted)). The Authors Guild court noted that “the purpose of Google's copying of the original copyrighted books is to make available significant information about those books, permitting a searcher to identify those that contain a word or term of interest, as well as those that do not include reference to it…,” leaving them “no doubt that the purpose of this copying is the sort of transformative purpose described in Campbell as strongly favoring satisfaction of the first factor.” (804 F.3d at 217 (emphasis in original)).

However, the court also noted that this first factor was not dispositive, and particularly considered the fourth factor under §107, “the effect of the use upon the potential market for or value of the copyrighted work,” stating that “[e]ven if the purpose of the copying is for a valuably transformative purpose, such copying might nonetheless harm the value of the copyrighted original if done in a manner that results in widespread revelation of sufficiently significant portions of the original as to make available a significantly competing substitute.”  (804 F.3d at 223 (emphasis in original)). Nonetheless, the court determined that Google’s presentation of “discontinuous, tiny fragments” of the original works would not adversely affect their market value, and that the fair use factors leaned in the favor of non-infringement (Id. at 224).

This may not be true for AI-based image generators. While they may not necessarily affect the market value of particularly famous works, such as where a user wants to present a copy of Banksy’s Flower Thrower or Jackson Pollock’s Convergence, image generators are being widely used as a replacement for stock photos such as those provided by Getty Images (See, e.g., “How generative AI will help power your presentation in 2023,” Sharon Goldman (VentureBeat, January 24, 2023) and “Forget clipart! These websites generate AI art instantly,” Rich DeMuro (KTLA, February 15, 2023)  As these tools improve in quality, they will almost certainly have a negative effect on the commercial art economy.

The primary counterargument is that image generators merely provide additional competition in the marketplace: rather than providing infringing copies of the original works, these systems provide new images, albeit similar to the originals. In this, they’re similar to a diligent art student that studies a wide variety of artwork and is able to create new works on request in a named style. While such an artist may have a negative effect on another artist’s commercial success, copyright was never intended to give creators a commercial monopoly on the entire industry.

Asking an art student – or an AI image generator – to create a new work in a particular artist’s style may have other legal implications beyond copyright. The plaintiffs in Andersen also raised common law and statutory right of publicity claims (the latter under Cal. Civ. Code §3344) due to the generation of art with their distinctive styles. For example, plaintiff Sarah Andersen is the creator of the webcomic Sarah’s Scribbles, which is characterized by a black-and-white, hand drawn and sparse style with characters with big heads and distinctive bulging eyes. Thousands of Andersen’s comics were apparently included in training data for Stable Diffusion, and users quickly discovered that image generation prompts including “in the style of Sarah Andersen” or similar text caused the system to generate images that could easily be mistaken for her original work. Similar prompts have been found for artists as diverse as Bill Watterson (“Calvin and Hobbes”), Edward Gorey (“Gashlycrumb Tinies”), and Rob Liefeld (“Deadpool”). In an opinion piece for The New York Times, Andersen stated, “[t]he notion that someone could type my name into a generator and produce an image in my style immediately disturbed me.” (“The Alt-Right Manipulated My Comic. Then A.I. Claimed It,” Sarah Andersen (N.Y. Times, December 31, 2022) The Andersen plaintiffs contend that they have “invested considerable energy, effort, ingenuity, and creativity into the development of their distinct artistic identities and have successfully built careers as artists,” and that “the value of Plaintiffs’ name recognition— and thus the value of their art itself—is diluted in a market flooded with AI-generated copies associated with Plaintiffs’ names and artistic styles.” (Complaint at 36-38, 3:23-cv-00201).

This may ultimately be a more compelling argument than copyright infringement. One can easily see users of AI systems passing off machine-created works as unknown originals by famous artists, and while the current generation of these tools frequently leave visual evidence of their origin (version 4 of Midjourney has been particularly noted for poorly drawing hands), future generations may have output that is indistinguishable from that of humans (Midjourney’s version 5 already shows incredible improvement). While the individual users creating works “in the style of” famous artists are not confused or misled into thinking they’re actually obtaining an original, secondary viewers may be. As the 6th Circuit noted in Esercizio v. Roberts in a case applying the Lanham Act (15 U.S.C. §1125), unfair competition claims are not limited to confusion at the “point of sale,” but can also apply where third parties, viewing a counterfeit work, may be confused as to its legitimacy: “‘[t]he fact that an immediate buyer of a $25 counterfeit watch does not entertain any notions that it is the real thing has no place in this analysis. Once a product is injected into commerce, there is no bar to confusion, mistake, or deception occurring at some future point in time…’ [T]his interpretation [of the statute] was necessary to protect against the cheapening and dilution of the genuine product, and to protect the manufacturer's reputation.” (944 F.2d 1235, 1244 (6th Cir. 1991)). Similarly, a court could find that using AI generators to create works in the style of given artists would ‘cheapen and dilute the genuine product.’

Pandora’s Box is opened and cannot be shut again: creative AI systems are not going to go away. These tools have been released “into the wild” with no real control or oversight; efforts to contain them may be inadequate, if not impossible. While training models may require petabytes of computer memory and thousands of high-end processors, the actual trained AI tools may be downloaded and run by anyone on commodity hardware. Although Congress and the courts may struggle to regulate commercial use of their output via copyright and trademark law, failure to do so could mean a huge transformation, if not the complete demise of the literary and professional art worlds in the near future.  But that’s not to say that there’s no hope. Other new technologies, from immutable blockchain records of provenance such as NFTs to reverse AI-detection and watermarking, can work hand in hand with a legal framework to provide a coherent and fair treatment of the underlying rights in AI-generated content.

Reprinted with permission from the April 26, 2023 issue of The Legal Intelligencer ©2023 ALM Media Properties, LLC. Further duplication without permission is prohibited. All rights reserved.