Should creators of generative AI systems be allowed to own the copyright to their systems’ output?
In August, Federal Judge Beryl Howell[i] of the Washington D.C. District Court sided with the U.S. Copyright Office in denying Stephen Thaler’s bid to copyright an artwork created by an artificial intelligence system he developed called DABUS. Judge Howell and the Copyright Office’s arguments rest on long-standing precedent in American copyright law that for a creative work to be copyrightable, it must be authored by a human being.
I agree with those decisions and have argued elsewhere for the non-copyrightability doctrine of generative AI: The default case for copyright and other forms of intellectual property should be that work generated by AI should not be eligible for intellectual property protection. But, Thaler’s case might merit different consideration, because he is also the sole author of the AI that generated the work.
First, it’s important to note how this case differs from two other related cases. With the comic book “Zarya of the Dawn,” the Copyright Office partially rescinded an initial copyright registration it initially extended to Kristina Kashtanova, when it learned that the images in the book were largely AI-generated. The Copyright Office ultimately extended copyright for the text and layout of the comic book, but not for the imagery. Similarly, Jason M. Allen, whose submission, “Theatre D’opera Spatial,” won a blue-ribbon prize at the Colorado State Fair in 2022, was denied copyright registration for that work. In both cases, the copyright-seekers used Midjourney, the image-creating generative AI. Kashtanova and Allen are both also on the record citing that they spent significant time spent coaxing Midjourney to generate the images they ultimately picked through what is now widely called “prompt engineering” — constructing and tweaking the text prompts used as input into generative AI systems. Kahstanova is also known to have further Photoshopped some of the images, but none of that moved the Copyright Office, which declared that neither prompt engineering nor Photoshopping met its bar of “sufficient amount of original [human] authorship.”
What distinguishes Thaler’s case is that he’s the creator of the AI that generated the artwork. Unlike Kashtanova and Allen, who used Midjourney, a generative AI someone else built, Thaler’s DABUS is his own creation, which Wired described as “an evolving system ‘at least 30 years in the making’.” The question Thaler and DABUS pose is, If a human being created a gen-AI system from scratch, then should they have intellectual property rights over work generated by that system?
Unfortunately, Thaler shot himself in the foot for copyright purposes, because he’s stated clearly that the artwork he wants to have copyrighted included no creative input from him: “My stuff just sits and contemplates and contemplates and comes up with new [artistic] revelations.” In effect, he said that he’s not the artist behind the artwork, so it’s hard to justify copyright ownership for him. (Thaler’s primary interest appears to be to have DABUS recognized as a sentient machine, and in his original copyright filing, he sought to have DABUS be the owner of the copyright. Only when that was rejected, he filed to have the copyright assigned to himself.)
So, Thaler and DABUS are themselves not a great case to reconsider copyrightability, but there are better examples. Decades before the current generation of generative AI, for example, there was David Cope’s incredible EMI, which could learn the style of inputted musical scores and generate original compositions in a similar style. Cope single-handedly created EMI and used it to compose music that, despite its detractors, was good enough to pass a musical version of the Turing test and to compose scores that were recorded by professional (human) musicians.[ii] Unlike Thaler, Cope has no illusions that EMI is sentient, and he is intimately familiar with its inner workings. And, unlike the neural-network-based generative AI underlying the current generation of AI, Cope can explain every line of his program.
And now, the question: Should David Cope be granted copyright of EMI’s oeuvre?
The U.S. Copyright Office would likely not grant it based on its prevailing policy, but there’s a case to be made that Cope deserves it. The case rests on three pieces, all three of which are essential. First of all, EMI is a technology Cope built based on his own understanding of music.[iii] To the degree that EMI can “understand” and reproduce musical styles, it is Cope’s understanding.[iv] Cope himself is on the record saying, “The programs are just extensions of me,” and that is a justifiable view, in a way that isn’t true of Thaler’s DABUS or current, neural-network-based gen-AI. That’s not to say that EMI’s output is 100% specified by Cope — the system automatically identifies patterns in the music it receives as input; random number generators are used when it composes; and the software can generate music at a speed that Cope could never match — but, Cope can, to a great extent, guess why EMI composed as it did. To put it in terms of current AI parlance, EMI is largely interpretable by Cope.
Second, Cope’s intent for EMI was for it to be a tool to advance his own musical goals, and not for the tool to simply spit out endless copyrightable works. In other words, Cope was arguably less interested in autonomous music generation, as much as to have a system that could either inspire or write the music he wanted to compose. In fact, Cope himself appears to have been taken aback by EMI’s infinite productivity: “You pushed the button and out came hundreds and thousands of sonatas or whatever.” He was so dismayed by this aspect of EMI that he ultimately selected 11,000 works to represent it, and then unplugged the system for good. His subsequent projects include EMI’s successor, “Emily Howell,” which is even more tool-like than EMI — it appears to be built for the sake of human-machine collaboration; it operates through what today might be called interactive prompt engineering. As Cope once said, “My two goals are an original style and to create something I love.”
Third, Cope invested considerable time and effort in EMI. This point is relevant only on top of the first two points above, and it serves to reinforce them. Though Kashtanova and Allen can also lay claim to having put in effort for their work, their effort was entirely that of a tool user using a tool they had no part in designing. In contrast, Cope’s efforts — which spanned decades — were about designing and re-designing the tool itself for the purposes of better serving his artistic goals.
So, it’s clear that Cope’s relationship to EMI is completely different from Kashtanova or Allen’s to Midjourney. But, Cope’s relationship to EMI also differs from Thaler’s relationship to DABUS, and it differs from the relationship of, say, ChatGPT to the engineers at OpenAI. Whereas Cope hand-coded the logic behind EMI, neither Thaler nor anyone at OpenAI claims to understand the “thinking” behind their respective creations. Thaler explicitly insists on DABUS’s intellectual autonomy; and modern AI scientists are as in the dark about how gen-AI works as their users. Cope was also not seeking to create an autonomous creative entity (as Thaler was) or trying to “solve” music composition in some generic way (as arguably, OpenAI does for text and image generation). Ultimately, there is something about Cope’s investment of thought and time in the code and his primary interest in advancing his own musical goals, that makes it difficult to dismiss him as the ultimate author not just of EMI, but of EMI’s compositions. That “something” might be worth considering as the basis for allowing some degree of copyrightability for gen-AI output that is produced by generative AI systems built by people for the sake of advancing their personal creative vision.
Notes
[i] Judge Howell is perhaps best known for overseeing grand jury cases involving Donald Trump.
[ii] Cope’s contributions to computer music are incredible in themselves, but because his programs were arguably the first form of generative AI to truly rival human creators in any creative endeavor, he was decades ahead in terms of thinking about generative AI, as well. For example, he was probably the first to question with any seriousness, his own authorship of creative work created by his own computer systems.
[iii] My sense, based on Cope’s prolific writing about EMI and computer music, is that it is based on Cope’s understanding of Western music. But, I’d guess that much of what he programmed would work reasonably well for much, though not all, of non-Western music, too.
[iv] David Cope has written a lot about his experiments with computer music. In his book, Virtual Music: Computer Synthesis of Musical Style, he has computer scientist Douglas Hofstadter describe how EMI works, based on conversations the two have had. This is very different from today’s gazillion-parameter neural-network AI systems which even their creators are mystified by.
