The United States Copyright Office recently issued new guidelines regarding copyright applications for works created with artificial intelligence tools. The new rules recognize that work made with both AI input and human creation can be eligible for copyright protection, but any part of it that is entirely made by AI is not eligible. Which is to say, copyright protections can extend only to work that is attributable to human authorship.
“If a work’s traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it,” the report states. “For example, when an AI technology receives solely a prompt from a human and produces complex written, visual, or musical works in response, the ‘traditional elements of authorship’ are determined and executed by the technology—not the human user.” The report hypothesizes that a work that creatively combines AI-generated elements into something new or AI-generated work that an artist then heavily modifies after the fact would indeed be eligible.
The new guidelines could provide clarity for artists using AI tools like SongStarter, a feature of the cloud-based DAW Bandlab that allows users to “generate royalty-free music in seconds”—presumably to arrange and manipulate to create original, and, now, copyrightable, work. Ditto for AI-generated instruments like those from Google’s “NSynth” project, which allows users to create never-before-heard instruments with a neural network trained on more than 300,000 instrument sounds. Per the USCO’s current guidelines, these are tools that humans can use to create copyrightable works.
“Historically, copyright eligibility doctrine has focused on the work as a free-standing object,” says Joseph Fishman, a Vanderbilt law professor whose research focuses on intellectual property and its relationship to creativity and the creative process. “It hasn’t focused much on the process through which that object is generated—until recently it didn’t really need to. But AI-assisted expressive outputs have become common enough that copyright policy is now being forced to confront that issue in earnest.”
Still, the current guidance leaves questions unanswered. For instance, if a song is just a series of existing beats and notes creatively combined by a human artist, then where do they come from in the first place? An AI “learns” by training itself on countless recordings—often copyrighted—the synthesis of which is used to create “new” songs. When Pharrell Williams produces a song, like, say, “Blurred Lines,” his training comprises all the songs to which he’s ever listened, and the new piece is a synthesis of everything to which he’s ever been exposed—like, say Marvin Gaye’s “Got to Give It Up.”
Just like it is easy to infer but difficult to prove which songs Williams knew inherently, it’s difficult to determine whether or not an AI model was trained on a particular copyrighted work. There is currently no legislation or apparatus requiring companies training AI models to disclose the data in their training sets, but Fishman says that if the resulting artwork made using an AI model is “substantially similar” to a copyrighted work, the same test of “how similar is too similar” would apply as it does for non-AI works, such as in the infamous lawsuit over the provenance of “Blurred Lines.”
Some labels have already gotten into the habit of pre-empting lawsuits by sharing songwriting credits with artists whose music shares similar elements, regardless of whether the artist was involved or an actual sample was used. That tactic that could potentially be employed similarly with music created with AI.
The exact amount of human involvement required to obtain a copyright will be decided in the courts, too. “One big threshold question is how much deference courts would give to this policy statement, or any subsequent office publication on this issue, to begin with,” says Fishman. “They might run with the office’s ‘creative control’ standard, or they might come up with something different.”
Erin M. Jacobson, an intellectual property rights attorney who works with musicians, songwriters, publishers, and other music industry professionals, interprets the office’s guidelines of “creatively combined” or “heavily modified” works to refer to situations where a human has control over how AI elements are arranged or used, as opposed to a machine arranging the elements in response to a prompt. “There are many songs that feature the same three to four chords,” Jacobson says, “but those songs are vastly different in sound and feel based on the creative expression of the human authors of those songs.”
FN Meka, the “A.I. powered robot rapper” whom Capitol Records signed to a record deal then quickly dropped after intense public criticism, provides a potential case study for AI-assisted music. FN Meka was voiced by a human, but the lyrics and song structures were generated by AI, created using thousands of data points compiled from video games and social media. Anthony Martini, founder of Factory New—the company behind FN Meka that describes itself as a “first of its kind, next-generation music company, specializing in virtual beings”—told Music Business Worldwide that Factory New “developed a proprietary AI technology that analyzes certain popular songs of a specified genre and generates recommendations for the various elements of song construction: lyrical content, chords, melody, tempo, sounds, etc. We then combine these elements to create the song.”
Jacobson says that output from something like Factory New’s process—as described by Martini—would not be eligible for copyright under current USCO guidelines. “There is a difference between creativity in selecting what works or materials are used to train the AI model, and the creativity used in creating the actual finished work,” she explains. “In this case, while there may have been great care and creativity in the selection of material used to train the AI, the finished songs themselves were created by the AI model, not by a human. The copyright office is clear that to be eligible for copyright registration, the finished work must have been created at the direction of a human, meaning that a human actively selected, used, and combined AI elements in a creative and original way. When the finished work is created whereby the AI model has chosen the selection, arrangement, and combination of elements, that work is directed by the AI model and would not qualify for registration.”
Even if an AI-assisted work is determined to have human-generated—and, therefore, copyright-eligible—elements, questions linger. Presumably, the human authors are entitled to master recording and/or publishing royalties generated by the song, but if an artist writes their own lyrics to an AI-generated instrumental track, does the instrumental version of the song instantly enter the public domain? DJs and crate-digging producers used to have to get creative when sampling or remixing tracks without the stems. Now, tools like Deezer’s Spleeter can use AI to automatically separate stems 100x faster than realtime, and anyone with a late model iOS device or Apple TV can use Apple’s AI-powered “Sing” feature to automatically remove vocals from tens of millions of songs in their library.
With machine learning technology developing at breakneck speed, the lines of human and machine authorship are increasingly fuzzy. No one knows exactly where they’ll be redrawn until the right case hits the courts. “The bottom line is that the application of AI technology in creative works has a plethora of novel applications that have not been examined before,” Jacobson says, “in light of our current laws and history of only dealing with human-authored works.”
Fishman concurs. “The copyright office is getting the first crack at trying to craft an answer,” he says. “But it almost certainly won’t be the last.”
.jpg)
