Artificial Imagination Unleashed: Navigating Copyright Frontiers in the Age of Generative AI

By: Yihua Jin

INTRODUCTION

The break of OpenAI and subsequent development of the revolutionary ChatGPT caused an explosion of generative AI use throughout the world. Nearly every area of business is experimenting with AI to improve or even revolutionize their way of working — from grammar checks to image generation, productivity tools have become an increasingly popular place to incorporate AI to improve their services.

Image, text, video, and sound generation are currently some of the most popular AI tools. These tools have a widespread application in all sorts of industries, as they can help generate anything from emails and computer code to legal documents and posters. Of all these industries, it is probably no surprise that these tools significantly impact the entertainment industry, an industry that depends on content production and creativity.

 A 2023 study by Variety reported that by September 2023, 16% of U.S. entertainment professionals used generative AI in some way, either within their team or their company. Another 14% of entertainment professionals did not use generative AI at the time, but planned to use it in the future. Generative AI has the potential for use in all areas of entertainment. In film and TV production, AI generates scripts, storylines, marketing campaigns, and movie and TV backdrop images. In music, voice cloning enables collaborations between world-famous artists and re-recordings of classic songs, without the artist’s physical presence. 

However, with the seemingly endless possibilities generative AI brings to the entertainment industry comes serious threats. One of the primary concerns of AI usage is job displacement by generative AI tools. The aforementioned 2023 Variety study showed that 42% of U.S. entertainment professionals are at least somewhat concerned about the impact generative AI will have on job security over the next two to three years. Scholars and professionals alike worry that generative AI may replace human writers, artists, and other creative professionals, as companies are incentivized to use generative AI to cut costs.

Another concern is copyright infringement. Generative AI programs are generally trained by ingesting a large amount of data, in the scale of billions or more. For example, Stable Diffusion and Midjourney, two text-to-image AI generation programs, were trained on a data set containing 5.8 billion images. Many of these copyrighted images were scraped from third-party websites such as Pinterest and art shopping sites without the original artists’ consent. This has created a new battleground for litigating issues related to artist rights. Currently, there is ongoing debate and litigation about whether scraping images from the internet without the artist’s consent constitutes copyright infringement.

Despite these concerns, AI tools have already been used within the entertainment industry, and it may not be long before the use of generative AI in blockbuster movies. For example, the visual effects team in “Everything Everywhere All At Once” relied on an AI tool from Runway, though not a generative one, to remove backgrounds from images. A team member also commented that had he discovered generative AI tools while making the movie, he would have used it to generate photos of aliens, to save the time of handmaking 300 shots that would be flipped through in seconds. Elsewhere, filmmakers have shown great interest in using Cuebric, the first generative AI system to produce and edit images for film and television, to create backdrops.

In general, it is fair to say there is a lot of uncertainty in terms of the impact of generative AI in the entertainment industry. Since the launch of ChatGPT in 2022, both the debate and the hype around generative AI have only increased. This note seeks to examine the developments of the debate around one particular aspect of generative AI: the use of image generator AI in the entertainment industry.

FROM PROMPTS TO PIXELS: UNDERSTANDING IMAGE GENERATOR AI 

Image generator AI is one of the most popular, yet most fiercely questioned types of generative AI. Image generator AI generally performs text-to-image generation, meaning it lets the user type a request in natural language, and then generates images based on the prompt the user enters. While the user has little control over the initial output beyond the prompt, many image generator AI programs allow users to pick from several results, and modify the output in various ways. For example, Midjourney allows the user to cut, zoom in, and zoom out of the image output. It also has a “variation” function, which produces variations that can “retain the main composition of the original image but introduce subtle changes to its details,” or “change the composition, number of elements, colors, and the type of details within the image” more drastically. Another image generator AI created by DeepAI allows the user to choose a desired image art style through the initial prompt.

Generative AI needs to be trained on huge datasets — and image generator AI is no exception. Extensive training on massive datasets is required to make the AI image generation models understand, improve, and master the relationship between language concepts and visual representations. “Massive,” in this context, means most developers or companies are not able to prepare the number of images necessary to train and perfect an image generator AI. As a result, developers and companies rely on scraping images from the internet. While these images are readily accessible, many are copyrighted. However, the scraping is most often done without permission and proper attribution to artists. As briefly introduced above, for example, Stable Diffusion was trained on a data set compiled by a German non-profit: LAION-5B. An analysis of 12 million of the 600 million images in the data set revealed that a large chunk of them came from third-party websites such as Pinterest, ArtStation, and Fine Art America – all sites where artists upload online portfolios or sell artwork. In other words, while the artists may have uploaded the images themselves, they did not intend to give up rights to the images. This has created, and will continue to create, novel copyright issues and claims of copyright infringements.

ARTIFICIAL CREATIONS, REAL DISPUTES: CURRENT LEGAL ISSUES INVOLVING GENERATOR AI

Image generator AI programs are among the first generative AI models to test the copyrightability of its outputs, and among the first to face copyright infringement liabilities. We’ve only scratched the surface of the questions raised by the potential copyrightability of image generator AI outputs. Recently, both a computer scientist and a comic artist attempted to obtain copyright protection for separate works generated with the help of image generator AIs by way of lawsuit, and were unsuccessful. On another front, Getty Images recently filed suit against Stability AI, and a group of visual artists have sued AI companies Stability AI Ltd., Midjourney Inc., and DeviantArt Inc., all for copyright infringement. Below will separately examine the above two sets of lawsuits, and the legal issues raised, resolved, or left open.

Copyrightability of Work Produced with the Help of Image Generator AI Models

Computer Scientist Stephen Thaler developed and owned a computer system he called “Creativity Machine.” Thaler sought to register an image created by the computer system, named “A Recent Entrance to Paradise,” with the Copyright Office. Notably, Thaler did not claim to be the author of the image, but named Creativity Machine as the author, claiming the work was “autonomously created by a computer algorithm running on a machine.” Subsequently, Thaler claimed copyright over the computer-generated work himself under the theory that it should be a “work-for-hire to the owner of the Creativity Machine.”

Comic artist Kristina Kashtanova sought to register a comic book created with the help of Midjourney, a text-to-image generator AI, under a drastically different theory. The comic book consisted of images generated by Midjourney, which Kashtanova then arranged and added text to. Kashtanova initially applied for and obtained a copyright registration for the comic book, without disclosing her use of Midjourney. She claimed to be the only author of the work. After the Copyright Office became aware of Kashtanova’s use of Midjourney, the Office intended to cancel her registration, deeming her application “incorrect, or at a minimum, substantively incomplete.” Kashtanova then timely provided additional information, claiming her registration should not be canceled because she either “authored every aspect of the work, with Midjourney serving merely as an assistive tool,” or at the very least, that “portions of the work are registrable because the text was authored by Ms. Kashtanova and the Work is a copyrightable compilation due to her creative selection, coordination, and arrangement of the text and images.” 

While the two individuals seeking copyright protection of images generated by image generator AIs alleged different theories, the Copyright Office and the US District Court for the District of Columbia’s holdings are essentially the same: the images lacked human authorship, and are therefore not copyrightable. 

Both courts applied the same standard to determine whether a work is copyrightable. The Copyright Act allows registration of work if it qualifies as an “original work[] of authorship fixed in any tangible medium of expression.” The term “original” has two components: (1) independent creation, and (2) sufficient creativity. Independent creation requires the work to be independently created by the author. Sufficient creativity is not a high bar; past case law holds that “only a modicum of creativity is necessary.” However, courts state that a work is not copyrightable when “the creative spark is utterly lacking or so trivial as to be virtually nonexistent.”

Moreover, courts have interpreted the phrase “works of authorship” to only allow creations of human authors to be copyrightable. In both Thaler’s and Kashtanova’s scenarios, the Office and the court found the image generator AI to be the author of the images. Therefore, in both cases, the images are not copyrightable, because they are not created by a human author.

At the appeals stage, Thaler presented the easier case. Thaler asserted new facts, stating that he “provided instructions and directed his AI to create the Work,” and that “the AI only operates at [his] direction.” However, the district court did not consider these facts when making the decision, because they directly contradicted the administrative record. In the administrative record, Thaler informed the Register that the image was “‘[c]reated autonomously by machine,’ and that his claim to the copyright was only based on the fact of his ‘[o]wnership of the machine.’” The court therefore simply followed precedent and held that the image was not copyrightable, because it lacked human authorship.

The district court also discussed reasons for the human authorship limitation. The court noted that:

The act of human creation—and how to best encourage human individuals to engage in that creation, and thereby promote science and the useful arts—was [] central to American copyright from its very inception. Non-human actors need no incentivization with the promise of exclusive rights under United States law, and copyright was therefore not designed to reach them.

Image generator AI and other forms of generative AI are certainly non-human actors, and the court made it clear that they will not give copyright protections to these actors because of public policy. However, the court has an interest in incentivizing creative works involving AI. While the court does not explicitly make any new holdings beyond reconfirming the necessity of human authorship, the court stated:

Undoubtedly, we are approaching new frontiers in copyright as artists put AI in their toolbox to be used in the generation of new visual and other artistic works. The increased attenuation of human creativity from the actual generation of the final work will prompt challenging questions regarding how much human input is necessary to qualify the user of an AI system as an “author” of a generated work, the scope of the protection obtained over the resultant image, how to assess the originality of AI-generated works where the systems may have been trained on unknown pre-existing works, how copyright might best be used to incentivize creative works involving AI, and more.

In other words, the court recognized that the copyrightability of artworks made with the help of an image generator AI program is unprecedented and undecided. How much human input is necessary to qualify the user of an image generator AI, not the AI, as an author, remains an open question. The court seems unwilling to declare that all images created with the help of generative AI are uncopyrightable. The court, as well as the United States Patent and Trademark Office, and the United States Copyright Office, seem interested in “‘jointly establish[ing] a national commission on AI’ to access, among other topics, how intellectual property law may best ‘incentivize future AI-related innovations and creations.’”

In Kashtanova’s copyright application, the Copyright Office took on the question of how much human input is necessary to qualify the user of an image generator AI as the author. While the Office did not provide a complete and definitive guide of what kinds of usage of image generator AIs keep the end product copyrightable, the Office’s analysis still provides helpful guidance and insight into the question.

Unlike Thaler, Kashtanova claimed that Midjourney was merely “an assistive tool” in her creation of the images in the comic book and that she authored the images. The Office rejected her argument. The argument went into great detail to examine how Midjourney worked, taking administrative notice that users have several ways to guide Midjourney’s output, including text describing what Midjourney should generate, “a URL of one or more images to influence the generated output,” and “parameters directing Midjourney to generate an image in a particular aspect ratio or providing other functional directions.” Nevertheless, the Office came to the conclusion that Midjourney, not Kashtanova, was the author of these images, because Kashtanova did not control the process of creating those images. The court stated, “While additional prompts applied to one of these initial images can influence the subsequent images, the process is not controlled by the user because it is not possible to predict what Midjourney will create ahead of time.” Additionally, despite Kashtanova’s assertion of influencing structure and content of each image, the Copyright Office “makes clear that it was Midjourney—not Kashtanova—that originated the ‘traditional elements of authorship’ in the images.”

The Office seemed to equate authorship with sufficient control over the final product—probably rightfully so. The Office determined that the usage of Midjourney is more like hiring an artist to create an image than using a tool, because the user cannot predict the specific output of Midjourney. The Office reasoned that when artists use other editing or assistive tools, “they select what visual material to modify, choose which tools to use and what changes to make, and take specific steps to control the final image such that it amounts to the artist’s ‘own original mental conception, to which [they] gave visible form.” Meanwhile, the text prompts do not guarantee any particular visual output. Therefore, prompts “function closer to suggestions than orders, similar to the situation of a client who hires an artist to create an image with general directions as to its contents.” In that situation, the author would be the artist hired, instead of Kashtanova. The Office also rejected the argument that Kashtanova’s significant time and effort in working with Midjourney indicated any control or authorship. The Office reasoned that “sweat of the brow” cannot be a basis for copyright protection in otherwise unprotectable material.

In sum, the Office seems to set a high bar that users of image generator AIs must overcome to gain copyright protections for the outputs. Text prompts or other ways to influence the outputs are not enough to assert sufficient control over the image, no matter how much effort a user puts into tweaking the output or choosing from a large number of outputs. The Office seems to suggest that direct outputs by image generator AI programs are not copyrightable, because the user is simply commissioning the AI to create an image. Therefore, similar to situations when the user commissions human artists to create an image, the author of the image is the artist, not the user.

However, the Office does not go so far as to say all images created with the help of image generator AIs are uncopyrightable. To the contrary, the Office confirms that it “will register works that contain otherwise unprotectable material that has been edited, modified, or otherwise revised by a human author.” However, they stated they would only do so “if the new work contains a ‘sufficient amount of original authorship’ to itself qualify for copyright protection.” In Kashtanova’s case, the court found changes to a portrait’s upper lip are “too minor and imperceptible to supply the necessary creativity for copyright protection.” Kashtanova claimed to have altered the image using photoshop to “show[] aging of the face, smoothing of gradients[,] and modifications of lines and shapes.” It is possible that had Kashtanova described in more detail what contributions were from her use of Photoshop as opposed to Midjourney’s generation, her substantive edits could have constituted human authorship. 

In the end, the Office granted Kashtanova’s comic book copyright protection to the text and the “selection, coordination, and arrangement of text created by the author and artwork generated by Midjourney.” However, what the copyright protects against potential copyright infringers is yet to be tested. For example, had the characters of the comic book been original, are the rights to use them reserved because Kashtanova had copyright to the text? Or, is anyone free to use the same characters? Moreover, while the Office’s opinion makes it extremely difficult to claim control over outputs of text-to-image generative AI, it seems possible that outputs of other kinds of images by generator AIs may still get copyright protection if the user has more control over the output.

Some tools newly employed by the video game industry seem to raise this question. For example, Meshy, a generative AI, lets users convert 2D images into 3D models. By providing the 2D image, does the user have sufficient control over the outputted 3D model? The user can theoretically control the color, look, and texture of the 3D model. Therefore, the argument that the user, not Meshy, is the author seems plausible. These tools bring us into unknown territory, and may force the courts and the legislature to clarify the amount of control necessary to warrant copyright protection in the future.

Copyright Misuse by Generative AI Developers

Early last year, both Getty Images and a group of visual artists filed lawsuits against generative AI companies. Both lawsuits allege that AI companies misused copyrighted work to train their generative AI systems. More specifically, Getty Images alleged that Stability AI committed copyright infringement, providing false copyright management information, removal or alternation of copyright management information, and trademark infringement. The group of visual artists alleged that Stability AI, Midjourney, and DeviantArt committed direct copyright infringement by downloading and storing copies of their works, using them to train AI image products. They also alleged vicarious copyright infringement when third parties used the companies’ products to create “fakes,” defined by them as “images that can pass as original works by that artist.” Moreover, they alleged companies violated the Digital Millenium Copyright Act by removing copyright management information from the works.

As of now, the courts have not made a complete decision on either case. However, the District Court of the Northern District of California dismissed most of the group of artists’ claims last October, leaving only the direct copyright infringement claim asserted by plaintiff Sarah Anderson against Stability to move forward. The court did allow plaintiffs to amend their complaint, and they did so timely.

Regarding the group of artists’ direct infringement claim, most of the problems with the claim were with the artists not registering their works. The court dismissed the plaintiffs’ Copyright Act claims related to unregistered works, and limited Anderson’s Copyright Act complaint to registered collections. While Anderson does not pinpoint which of her registered works were used as Training Images, she uses “ihavebeentraiend.com” to justify her belief that her works were used in LAION datasets and for Stable Diffusion training. The court sided with Anderson and stated “[t]hat is a sufficient basis to allow [Anderson’s] copyright claims to proceed at this juncture.” The court specifically took notice of the special nature of the case, namely the difficulty of identifying which specific copyrighted work was scraped illegally, as “LAION scraped five billion images to create the Training Image datasets.”

The court also held that Anderson’s primary theory of direct copyright infringement should withstand Stability AI’s motion to dismiss. The court stated:

Plaintiffs have adequately alleged direct infringement based on the allegations that Stability “downloaded or otherwise acquired copies of billions of copyrighted images without permission to create Stable Diffusion,” and used those images (called “Training Images”) to train Stable Diffusion and caused those “images to be stored at and incorporated into Stable Diffusion as compressed copies.”

The court seems to be receptive to the artists’ core claim, at least at the pleading stage, that the generative AI company’s scraping and usage of copyrighted images to train their generative AI constitutes direct copyright infringement. Since almost all generative AI models are trained by scraping copyrighted images, should the court ultimately rule in the artists’ favor, it would essentially confirm that all generative AI companies have been committing copyright infringement. The decision may force those companies to drastically change their business practices. For example, it may force these companies to shell out large amounts of money in damages to artists, or an injunction could bring the use of generative AI to a halt until companies alter their business practices.

Internationally, a UK court recently ruled that Getty Images’ lawsuit against Stability AI for copyright infringement can proceed to trial. While the outcomes of these trials are far from certain, they will undoubtedly have important implications for the future developments of generative AI programs, and the relationship between artists and technology. Noteworthily, both Getty Images and the prior mentioned group of artists seek injunctive relief and damages against developers of generative AI. In the most extreme scenario, this may force companies that have developed image generators to disregard their current models and force future developers to either develop their own database for training models, or pay for images. How that can and will realistically happen remains an unresolved question. Regardless, this may render generative AI economically unfeasible.

Legal Gray Areas and Unanswered Questions In AI and Copyright Law

As of yet, it seems that no artists have tried to sue users of image generator AIs for copyright infringement. Realistically, it is probably much more efficient to target generative AI companies than to target individual users. However, this does not mean that users cannot be found to have infringed copyrighted images by using image generator AIs, and we may start to see suits targeting users as generative AI becomes increasingly more prevalent.

Consider the following situation: the user of an image generator AI asks the model to produce a certain image in a certain artist’s style, and the AI does so with satisfactory quality. Is the user directly infringing the artist’s artwork? To prevail on a copyright action in this scenario, the artist would have to prove ownership of the copyright, and infringement by the user. Establishing copying requires proof that there is substantial similarity between the copyrighted and accused works. The definition of “substantial similarity” is “whether an average lay observer would recognize the alleged copy as having been appropriated from the copyrighted work.” Copying “need not be of every detail so long as the copy is substantially similar to the copyrighted work.” However, the copying of an idea is not copying for purpose of determining copyright infringement, because ideas are not protectable. Only “the particular expression of an idea” is protectable.

One particular case illustrates this point. While a copyright infringement case between artist Saul Steinberg and producers, promoters, distributors, and advertisers of the movie "Moscow on the Hudson," Steinberg v. Columbia Pictures Indus., does not involve generative AI, it provides an example of when the copying of style, composition, rendering of details and such amounted to “substantial similarity.” In the case, the artist claimed the defendants’ movie poster copied his illustration. The court found that there was substantial similarity between the copyrighted illustration and accused poster, because:

Both illustrations represent a bird’s eye view across the edge of Manhattan and a river bordering New York City to the world beyond. Both depict approximately four city blocks in detail and become increasingly minimalist as the design recedes into the background. Both use the device of a narrow band of blue wash across the top of the poster to represent the sky, and both delineate the horizon with a band of primary red.

Moreover, the court also noted the strong similarity in “the rendering of the New York City blocks. Both artists chose a vantage point that looks directly down a wide two-way cross street that intersects two avenues before reaching a river.” In sum, while the copying of an artist’s style alone would probably not constitute copyright infringement, if the image generator AI in doing so also copied the artist’s composition, rendering of details, and such, the user may be subject to copyright infringement liability.

While this issue has not yet been litigated, this scenario is already a reality. Greg Rutkowski is a Polish digital artist who creates dreamy fantasy landscapes with classical painting styles. His distinctive style has been one of the most commonly used prompts in Stable Diffusion. A news reporter, while reporting on this phenomenon, opened Stable Diffusion and entered “Wizard with sword and a glowing orb of magic fire fights a fierce dragon Greg Rutkowski,” and the image generator AI produced something that looks arguably similar in style, color and some details to Rutkowski’s work. This scenario begs the question: does a user inputting these prompts infringe on Rutkowski’s work when the output is arguably substantially similar to one of his copyrighted paintings? It seems to be a possibility and a legal risk for the user. We have yet to see how a court would rule in such a case, but the outcome would have huge implications for artists and AI users.

CONCLUSION

There is still much uncertainty around the laws governing generative AI. The outcome of current lawsuits disputing whether scraping and using acquired images to train generative AI models constitutes copyright infringement may significantly impact how future generative AI companies operate, and how those companies interact with artists. If these companies are found to have committed copyright infringement, they may have to figure out how to compensate millions of artists. To avoid future litigation, companies may be forced to strike deals with artists to use their work. On the other hand, future attempts to obtain copyright protections for works created with the help of generative AIs will refine the tests for when a generative AI program is a tool, and when it is the actual author.

We have yet to see when and how these legal disputes will eventually be resolved, but the debate around generative AI may not be over until the sharp divide between technology and art is reconciled. The resolution of these legal disputes by the courts may not be satisfactory for artists and may fail to adequately resolve the issues that AI has created. The lack of legal liability for generative AI companies can lead to artists taking matters into their own hands. In fact, some artists have already used newly developed data poisoning tools to protest generative AI companies’ scraping practice. A tool named Nightshade lets artists add invisible changes to the pixels in their art. If the image is scraped into an AI training set, it will damage the future iterations of image-generating AI models by rendering some of their outputs useless. For example, the model will output a cat instead of a dog, and cows instead of cars. 

Until the fear of job displacement is eradicated, and artists no longer feel used by generative AI companies, the courts and Copyright Offices will struggle to truly promote both human creativity and the progressive of this innovative technology.