Shannon Bjorklund

On January 29, the U.S. Copyright Office released Part 2 of its planned 3-part report on the legal and policy issues related to copyright and artificial intelligence (AI). Part 1 of the report, which was published in July 2024, addressed the topic of digital replicas. Part 2 now addresses issues related to the copyrightability of generative AI model outputs. The highly anticipated Part 3 is slated to address copyright infringement and fair use issues involving generative AI. Generative AI Models (Generally) Generative AI models are machine learning models that, when given an instruction or request (a prompt), can generate new content such as text, images, videos, and sounds. Prompts are often textual, but can be other input, such as images. These models learn the underlying patterns and structures of their training data and use them to produce new data based on the prompt. Examples include Meta’s Llama models, OpenAI’s ChatGPT models, and Google’s Imagen 3 model. The creation of prompts themselves has developed into its own category of computer technology—many in the field have begun using generative models to develop large amounts of prompts or to develop optimized prompts. Due to the rising use of these generative AI models, the Copyright Office is publishing guidance addressing issues of copyright law uniquely applicable to AI models. Historical Copyright Law Gives Some Guidance The Copyright Office began Part 2 of its report by explaining how copyright law has kept up with changes in technology, particularly with the inventions of photography, movies, and video games. As technology evolves, the Copyright Office and courts must decide to what extent technology can be involved in works of authorship that can still claim copyright protection. Because technology has always been permitted to be used in generating copyrighted works, the Copyright Office explained that there cannot be a bright-line rule “that registration will be denied merely because a computer may have been used in some manner in creating the work.” Instead, the analysis involves a case-specific sliding scale: “Where AI merely assists an author in the creative process, its use does not change the copyrightability of the output. At the other extreme, if content is entirely generated by AI, it cannot be protected by copyright.” As discussed by the TMCA in the past, the U.S. Supreme Court and the Copyright Office have long held that authorship requires the involvement of human creativity. The crucial question is “whether the ‘work’ is basically one of human authorship, with the computer [or other device] merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.” U.S. Copyright Office, Report to the Librarian of Congress by the Register of Copyrights (1965). The Copyright Office also analogized the generative AI model issues to other topics already decided in the law. For example, cases involving commissioning works, where courts have analyzed what level of contribution is necessary to qualify as authorship, seem applicable. A person who merely describes to an author what the commissioned work should do or look like is not a joint author for purposes of the Copyright Act, because such contributions constitute unprotectible ideas. Likewise, a person who provides high-level guidance to a generative AI model would also have difficulty proving sole or co-authorship. Specific Examples Provided by the Copyright Office The Copyright Office offered a few examples to illustrate the boundaries of copyright protection for material produced (at least in part) by generative AI models. Prompts On one end of the spectrum, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output. “Prompts essentially function as instructions that convey unprotectible ideas.” The Copyright Office provided its own generative example to illustrate its point: While this prompt is very descriptive, and the resulting image reflects some of the instructions, it does not reflect all of them (there is no “highly detailed wood”). The generative AI model also uses its own content to fill in gaps (the cat’s breed, clothes beneath the robe, human hands, etc.). The Copyright Office noted that identical prompts can generate multiple different outputs, which further underscores the lack of human control. This reality may change as technology evolves. “There may come a time when prompts can sufficiently control expressive elements in AI-generated outputs to reflect human authorship. If further advances in technology provide users with increased control over those expressive elements, a different conclusion may be called for.” But for now, the Copyright Office’s position is clear: prompts are not copyrightable. Expressive Inputs Expressive inputs are inputs of copyrighted works into generative AI models with instructions to modify those works. For example, a human author may create a drawing, input that work into an AI system, and instruct the system to modify color or scale. Or a user could upload a copyrighted story written in the first person point of view, and instruct the model to re-write the story from that of a third person. These inputs are different from prompts because they contribute more than just an intellectual idea. The Copyright Office again provided its own example: Here, the Copyright Office appears more inclined to provide protection, though that protection would not cover the entire output. Instead, “copyright in this type of AI-generated output would cover the perceptible human expression.” Modifying or Arranging AI-Generated Content Finally, the Copyright Office addressed situations where a generative model is used as an initial step, but the author further modifies the output. The Copyright Office noted that whether or not such modifications rise to the minimum standard of originality, that is, whether the work is the product of creative choices with respect to the selection of the images and/or text that make up the work and the placement and arrangement of the images and/or text on each of the work’s pages. The copyright would extend to the material the human author contributed but would not extend to the underlying AI-generated content itself. The inclusion of elements of AI-generated content in a larger human-authored work does not affect the copyrightability of the larger work as a whole. “For example, a film that includes AI-generated special effects or background artwork is copyrightable, even if the AI effects and artwork separately are not.” Key Takeaways The key takeaways from Part 2 of the report include how the Copyright Office interprets the human authorship requirement as it relates to the use of models. The Copyright Office explained that “[c]opyright protects the original expression in a work created by a human author, even if the work also includes AI-generated material,” but “[c]opyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements.” Whether the human interaction with the generative AI model is sufficient to constitute authorship should be decided on a case-by case basis. In the words of the Copyright Office: “prompts do not alone provide sufficient control.”

How would you pronounce #law? Is it “hashtag law,” as the PTO examiner suggested would be the most common understanding, or “pound law,” as alleged by the unsuccessful registrant of #law as a service mark. And to what does it refer: a hashtag for a social media topic, or a number that can be dialed from a mobile phone to reach an attorney referral service? Law firm Pound Law, LLC filed a lawsuit against the PTO and its director, arguing that its application for #law was improperly denied. According to Plaintiff, “the #LAW mark has become widely recognized by the relevant public as a distinctive brand and as a mnemonic/vanity telephone number mark uniquely associated with the services of Morgan & Morgan,” because Plaintiff spent “hundreds of millions” of dollars advertising it since 2006. Notably, this #law apparently only applies to telephone calls made on a mobile phone; it does not work on landlines, and the firm uses the website “www.forthepeople.com” and not either Pound Law or Morgan & Morgan. The examiner, and TTAB, decided that “consumers would not perceive [Plaintiff’s] use of #LAW as a source indicator but rather only as a means to contact [Plaintiff], based on the manner of use.” The history of this mark illustrates how quickly technology has unfolded over the past 15 years. Morgan & Morgan allegedly began using #law as a mnemonic and vanity phone number in 2006—the same year Twitter was founded. A few years later, Twitter took off: it had 300,000 tweets per day by the end of 2008, which ballooned to over 50 million tweets per day by early 2010. In 2022, there were over 500 million tweets per day, a non-negligible number of them using #law to identify the subject. A Twitter search renders countless tweets from law firms to law schools, legal consultants, protestors, and coaches using #law. Is Morgan & Morgan’s use really as “unique” and “distinctive” in 2023 as it may have been in 2007? The treatment of this mark calls to mind the Booking.com saga, covered extensively in this blog. The TTAB determined that the mark “Booking.com” was not registrable on the basis that it was generic, described here. The TTAB decision was overturned by Judge Brinkema in the Eastern District of Virginia, which held that Booking.com had acquired distinctiveness, described here, but awarded fees to the PTO because the case had been brought in district court instead of appealed directly to the Federal Circuit, as described here. The Fourth Circuit affirmed the District Court’s decision that the mark is registrable, but ordered the plaintiff to pay the PTO’s fees on appeal, described here. The PTO petitioned for certiorari to the U.S. Supreme Court, which was granted as described here, and the Supreme Court ultimately ruled that Booking.com is registrable, rejecting the premise that a generic.com term is ineligible for trademark protection, described here. This also calls to mind a different source of trademark obsolescence: trademarks that have been lost due to genericide. Among these are the yoyo, the flip phone and the escalator – all of which were marks associated with one brand until they became so widely used that they no longer identify a specific brand. Will #law go the way of the yoyo, or will it follow the winding path of the Booking.com decisions? Only time will tell.