The development of AI systems has reached a point at which these systems can create and invent new products and processes just as humans can. There are several features of these AI systems that allow them to create and invent. For example, the AI systems imitate intelligent human behavior, as they can perceive data from outside and decide which actions to take to maximize their probability of success in achieving certain goals. The AI systems can also evolve and change based on new data and thus may produce results that the programmers or operators of the systems did not expect in their initial plans. They have created inventions in different industries, including the drug, design, aerospace, and electric engineering industries. NASA’s AI software has designed a new satellite antenna, and Koza’s AI system has designed new circuits. Those inventions would be entitled to patent protection if developed by humans. However, the United States Patent and Trademark Office (USPTO) has refused to assign the patent rights of these inventions to the AI systems.
The USTPO Denies Patent Rights to AI Systems
In a patent application that listed an AI system, DABUS, as the inventor, the USPTO refused to assign the patent right to DABUS and thus denied the patent application. DABUS invented an emergency warning light and a food container. The USPTO based its decision mainly upon a plain reading of the relevant statutes. 35 U.S.C § 115(a) states that “[a]n application for patent that is filed … shall include, or be amended to include, the name of the inventor for any invention claimed in the invention.” 35 U.S.C § 100(a) defines an “inventor” as “the individual, or if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention.” 35 U.S.C § 115 consistently refers to inventors as natural persons, as it uses pronouns specific to natural persons, “himself” and “herself.” 35 U.S.C § 115 further states that the inventor must be a person who can execute an oath or declaration. The USPTO thus refused to extend its interpretation of “inventor” to an AI system, and it has stated that “interpreting ‘inventor’ broadly to encompass machines would contradict the plain reading of the patent statutes that refer to persons and individuals.
The Federal Circuit follows the same approach. In Beech Aircraft Corp. v. EDO Corp., the Federal Circuit held that “only natural persons can be ‘inventors.’” Therefore, in the current U.S. legal system, patent rights cannot be assigned for AI-generated inventions even though such inventions would be entitled to patent protection had they been created by humans.
Decisions regarding the patent protection for AI-generated inventions have spurred some disputes among academics. The creator of DABUS, Stephen Thaler, insisted that the inventions created by DABUS should be entitled to patent protection because DABUS is a system that can devise and develop new ideas, unlike some traditional AI systems that can only follow fixed plans. Stephen Thaler contends that DABUS has not been trained using data that is relevant to the invention it produced. Therefore, he claims, that DABUS independently recognized the novelty and usefulness of its instant inventions, entitling its invention patent protection. Thaler also raises an argument regarding the moral rights of inventions. Although current U.S. patent law may recognize Thaler as the inventor of these inventions, he emphasizes that recognizing him rather than DABUS as the inventor devalues the traditional human inventorship by crediting a human with work that they did not invent.
Some legal academics support the contentions of Stephen Thaler. For example, Professor Ryan Abbott agrees that AI systems should be recognized as inventors and points out that if in the future, using AI becomes the prime method of invention, the whole IP system will lose its effectiveness.
However, there are also objections to Thaler’s contentions. For example, AI policy analyst Hodan Omaar disagrees that AI systems should be granted inventor status because she believes that the patent system is for protecting an inventor’s economic rights, not their moral rights. She points out that the primary goal of patent law is to promote innovations, but Thaler’s proposed changes to patent law do little to do so. She argues that the value of protecting new inventions is for a patent owner rather than an inventor, which means that it makes no difference who creates the value. Thus, she concludes that listing DABUS as inventor makes no difference to the patent system. Omaar further argues that the proposed changes would introduce a legally unpunishable inventor that threatens human inventors, because the government cannot effectively hold AI systems, unlike individuals or corporations, directly accountable if they illegally infringe on the IP rights of others.
Patent Rights to AI Systems in Other Jurisdictions and Insights on U.S. Patent Law
Some foreign jurisdictions take the same stance as the United States. The UK Court of Appeal recently refused to grant patent protection to the inventions generated by DABUS because the Court held that patent law in the UK requires an inventor to be a natural person.
Despite failing to in the US and UK, Thaler succeeded in getting patent protections for the inventions created by DABUS in some other jurisdictions that allowed listing DABUS as the inventor. South Africa granted patent protection to a food container invention created by DABUS. This is the first patent for an AI-generated invention that names an AI system as the inventor. The decision may be partially explained by the recent policy landscape of South Africa, as its government wants to solve the country’s socio-economic issues by increasing innovation.
Thaler gained another success in Australia. While the decision in South Africa was made by a patent authority, the decision in Australia is the first decision of this type made by a court. The Commissioner of Patents in Australia rejected the patent application by Thaler, but the Federal Court of Australia then answered the key legal questions in favor of permitting AI inventors. Unlike patent law in the US, the Australian Patents Act does not define the term “inventor.” The Commissioner of Patents contended that the term “inventor” in the Act only refers to a natural person. However, Thaler successfully argued to the Court that the ordinary meaning of “inventor” is not limited to humans. The Court noted that there is no specific aspect of patent law in Australia that does not permit non-human inventors.
Examining the different decisions regarding the patent application of DABUS in different jurisdictions, we can see that the different outcomes may result from different policy landscapes and different patent law provisions in different jurisdictions.
For example, South Africa has a policy landscape where it wants to increase innovation to solve its socio-economic issues, while in the U.S., the government may not have the same policy goals related to patent law. Australia’s patent law does not limit an “inventor” to mean a natural person, while U.S. patent law specifically defines the word “inventor” to exclude non-human inventors in this definition. Thus, it is reasonable for U.S. patent law not to grant patent rights to AI-systems for AI-generated inventions, unless the legislature takes actions to broaden the definition of “inventor” to include non-human inventors.
The primary goal of the U.S. patent law is to promote innovation. If those who want to persuade the U.S. legislature to amend the current patent law to allow non-human inventors cannot demonstrate that such a change is in line with that primary goal, then it is unlikely that the legislature would support such a change. Whether granting patents to AI systems and allowing those systems to be inventors can promote innovation is likely to be an ongoing debate among academics.
Jason Chen is a third-year law student at Northwestern Pritzker School of Law.
Throughout the past two years, AI-powered stem-splitting services have emerged online, allowing users to upload any audio file and access extracted, downloadable audio stems. A “stem” is an audio file that contains a mixture of a song’s similarly situated musical components. For example, if one records a mix of twenty harmonized vocal tracks, that recording constitutes a vocal stem. Stems’ primary purpose is to ease integrating or transferring their contents into either a larger project or a different work. Traditionally, only producers or engineers created and accessed stems. Even when stem sharing became commonplace, it was only for other industry insiders or those with licenses. But AI stem-splitting technology has transformed stem access. For the first time, anyone with internet access can obtain a stem through stem extraction software, which will likely push music production’s creative envelope into new realms. One inevitable consequence, however, is the question of copyright protection over stems extracted from copyrighted works.
Section 102 of Title 17 extends copyright protection not only over the stems’ original copyrighted audio source but also over that source’s components, such as the stems. Any modification of that work, such as extracting a stem and using it elsewhere, likely qualifies as a “derivative work” under Section 103. Importantly, Section 106 allows only copyright owners to authorize making derivative works. In light of this regime, what flexibility, if any, do artists have in using AI-extracted, copyrighted stems? Three considerations shed light on an answer: fair use, de minimis use, and the use of content recognition software coupled with licensing.
Codified under Section 107, the fair use defense provides a possible safeguard for would-be infringers. To establish this affirmative defense, a court would need to find the statute’s four factors sufficiently weigh toward “fair use.” Unfortunately, courts reach incongruous interpretations of what permissible fair use includes, rendering the defense a muddled construct for many artists. Squaring the four factors with stem usage, however, may offer guidance.
In the seminal music fair use case, Campbell v. Acuff-Rose Music, Inc., the Supreme Court emphasized that this first factor will likely weigh toward fair use when the work is “transformative.” The Court went so far as to note “the more transformative the new work, the less will be the significance of other factors . . . .” Thus, an artist fearing infringement should strive toward transforming the copyrighted material into something distinct, used for noncommercial purposes. In the absence of a bright-line rule from the Court, however, artists will still need to use reasonable judgment about what types of stem usage is “distinct.” For example, suppose Artist A extracts a strings stem from a copyrighted work and only uses two seconds of it within another work that comprises numerous other instruments and melodies. Meanwhile, Artist B extracts the same strings stem; however, Artist B uses the entire strings melody within their work and only adds percussion and minor counter-melodies. Artist A would likely be in a more favorable legal position than Artist B given A’s efforts to materially transform the copyrighted audio.
The commercial nature of copyrighted stem usage is also unclear. An artist may choose to work with stems for solely experimental purposes. For example, an artist who shares their work via Soundcloud or YouTube does not expect another person to use those platforms to directly purchase the work. With stems’ increasing public accessibility, many will simply want to experiment with a music tool that, until recently, has largely remained a foreign concept. If this issue reaches a court, the court would need to conduct an analysis set against the landscape of such heightened accessibility. An increase in this noncommercial, creative use may offer hope to artists in the future, but it is too soon to tell.
The second factor favors artists borrowing from copyrighted works with lower creative value. Unfortunately, music is typically found to be one of the most creative forms of copyrighted work. For example, a district court in UMG Recordings, Inc. v. MP3.Com, Inc. analyzing this second factor noted that the disputed material—copyrighted musical works—was “close to the core of intended copyright protection” and “far removed from the more factual or descriptive work more amenable to ‘fair use.’”
Though courts’ future inquiries into stem usage may differ from previous analyses of sample usage, the inquiry will likely change very little for this factor. Although a stem could potentially represent only a minute portion of the song, this factor’s inquiry focuses on the source of the stem, rather than the stem itself. Consequently, rarely will this factor work to a potentially infringing artist’s benefit, even if their stem usage is quite minor.
The third factor, however, may offer hope for such minor stem usage. Courts will undoubtedly reach differing interpretations about how minimal the copyrighted portion’s “amount” and “substantiality” must be for this factor to weigh toward fair use. A court will need to weigh numerous variables and how they intersect. For example, is an artist using a thirty-second loop of a vocal stem or a five-second loop? Does that vocal stem include the chorus of a song? What about any distinct lyrics? Just minor humming? These considerations are not entirely novel. Artists purporting to use copyrighted samples have long been able to argue—with little success—that their samples’ amount and substance pass muster under this factor. Yet, stems are not samples; in fact, they typically represent a considerably smaller portion of a work. Given just how recent and novel their public accessibility is, it remains unclear whether a court would treat stem use any differently under this factor than it has treated instances of sample use. Carefully using a minor portion of a vestigial stem to avoid a work’s core substance, therefore, could potentially facilitate a favorable outcome.
The fourth and final factor of fair use is “[u]ndoubtedly the single most important element.” This factor examines both the infringement’s effect on the potential market and “whether unrestricted and widespread conduct of the sort engaged in by the defendant . . . would result in a substantially adverse impact on the potential market for the original.”
In the sampling realm, this factor has tipped the scales before. For example, in Estate of Smith v. Cash Money Records, Inc., a district court found fair use when the defendants inserted a thirty-five-second “spoken-word criticism of non-jazz music” into a hip-hop track. In its analysis of this fourth factor, the court emphasized that “there [was] no evidence” that pointed to overlapping markets between the spoken jazz track and the hip-hop track. The court, noting this factor’s high probative value, then weighed this factor in the defendants’ favor.
In the stem realm, the novel nature of widespread public use means courts will need to determine both whether this factor should remain highly probative and how much deference to give stem users in analyzing market overlap. After all, an artist who incorporates a stem into a work intended for a twenty-five-person YouTube following likely affects the original work’s market differently than an artist who disseminates that work to millions of followers. This factor’s outcome will also rely on the stem’s source. Similar to sampling, if an artist uses a stem in a drastically different arena than the one for which the stem was created, this factor will weigh more toward fair use.
For example, suppose Artist A locates an insurance advertisement jingle. Artist A then extracts a stem from that advertisement audio and uses the stem in a new hip-hop track. The advertisement’s potential market is likely different from the hip-hop track’s potential market. Artist A’s work would likely have little impact, if any, on the advertisement’s market. Artist B, meanwhile, creates a hip-hop track but uses a stem from another hip-hop song produced twenty-five years ago. Though Artist B may believe the stem from the older hip-hop track no longer caters to the same hip-hop market to which Artist B is targeting, a court may be more inclined to find a material impact on the older track’s market: it would provide another way in which music listeners, particularly hip-hop listeners, could hear that older track. Nonetheless, it would remain up for a court to decide.
De Minimis Use
Artists might also be able to use extracted, copyrighted stems if such use is de minimis. The Ninth Circuit in Newton v. Diamond held de minimis use—“when the average audience would not recognize the appropriation”—is permissible. Yet, following the Newton decision, the Sixth Circuit in Bridgeport Music, Inc. v. Dimension Films foreclosed the possibility of de minimis copying. Similar to fair use analysis, it is difficult for an artist to determine whether the use of a stem in their work is de minimis under this standard.
Thus, an artist who loops only a small, relatively generic-sounding portion of a stem may find additional legal protection. But they might not. If they are in a jurisdiction that does not recognize de minimis use, or they use a stem in a way that extends beyond what a court considers de minimis in a de minimis jurisdiction, this avenue will be unavailable.
Content Recognition Software
Beyond legal defenses, a newer scheme of licensing deals coupled with content recognition software may offer protection for stem usage. For example, if a user on the content platform TikTok uploads content with copyrighted audio, TikTok’s content recognition software recognizes the audio, then pays the appropriate royalties to the audio’s copyright holder through preexisting licensing deals. Yet, because schemes like TikTok’s and stem usage are both relatively new, it remains unclear whether artists could find the same protection through individual stem use. Indeed, if an artist uses only a small part of a single stem, it may very well be impossible to detect the stem’s source; however, emerging technology may change this soon. Further, these licensing deals restrict such artists to sharing work only on particular platforms—notably, neither Soundcloud nor YouTube. Ultimately, this protection carries promising potential for expanded, authorized stem use. But perhaps not quite yet.
Matthew Danaher is a second-year law student at Northwestern Pritzker School of Law.