[Guest post] Why the US Copyright Office’s AI Copyrightability Report is dysfunctional

The IPKat has received and is pleased to host the following guest contribution by Katfriend Zachary Cooper (VU-Amsterdam Law School), on the US Copyright Office’s Copyright and AI Report Part 2: Copyrightability. Here is what Zachary writes:

Why the US Copyright Office’s AI Copyrightability Report is dysfunctional

by Zachary Cooper

On January 29, the US Copyright Office (USCO) released its long-awaited Report on the Copyrightability of Generative AI (GenAI) outputs after setting an early example as a jurisdiction that would not grant copyright over AI-generated materials [IPKat here].

Last year, I wrote an article challenging the functionality of the USCO’s approach, arguing that it was unauditable and at odds with the reality of contemporary creative landscapes.

Now, after receiving over 10,000 comments from stakeholders, the USCO has updated its position to… essentially the same position, fraught with essentially the same issues.

New report, old problems

A Kat generating new artworks on Photoshop

The new position is partly reworded, now dichotomising rights between material where “AI tools assist rather than stand in for human creativity” (which receives copyright) and “purely AI-generated material, or material where there is insufficient human control over the expressive elements” (which does not). Given the vagueness of these criteria and lest anyone mistake them for a major shift in position, the Report is then quick to assert that all current-state outputs of prompting are the latter and therefore uncopyrightable. Similarly, the Report now specifies that for outputs prompted by “expressive inputs” (i.e., prompting a tool with pictures or songs), copyright would only cover those expressive elements already copyrightable in the input but “would not extend to the AI-generated elements standing alone.” If the author then chooses to modify something they have generated with GenAI, they can receive copyright just to their non-GenAI modifications.

While this approach might look clean on paper (copyright for the non-GenAI / no copyright for the GenAI), it is far from clean in practice.

The first question – and a very large one – is: how are we supposed to know what is AI-generated and what isn’t? GenAI tools currently run the gamut from those that negligibly affect the final work (“make my voice prettier”) to those that develop the final work with negligible effort (“make me a song”). The USCO is relying on AI disclosure to ascertain who receives copyright over what. Let’s leave out the very large issue of bad faith actors who do not disclose AI use to achieve full rights to their work. (It feels as if these accusations have become commonplace).

Even assuming good faith actors that want to do the right thing… GenAI tools are increasingly mere buttons to be pressed within creative software. Are we now expecting artists to keep a log of every time they press one of them? I have previously called this “the Old Hard Drive Debacle”. As it stands, if an artist finds an old hard drive filled with songs and pictures that they made and they did not keep a granular record of every single time they re-generated a few seconds of a drum line or inpainted a murky corner with GenAI, how are they supposed to disclose AI use for any of these works? Why would any artist go to the effort of keeping a record of every time they used a tool in their process when they have never had to do this for any other creative tool and it is entirely in service of them losing rights to their work?

Without a means of differentiating when material is developed with GenAI, without GenAI, initially with GenAI but then partly modified without GenAI, or initially without GenAI but then partly modified with GenAI, separating out the copyrightability of every element is unworkable and will surely lead to authors underplaying GenAI usage to receive copyright over the entirety of their work if it cannot be proven otherwise.

Creativity tracking

For this reason, there is much talk of AI-tracking technological solutions. The problem is – none of them are currently reliable in determining whether GenAI has been used at all, let alone at a granular enough level for the USCO’s element-by-element qualitative analysis. Indeed, for the USCO to have this kind of metadata would require click-by-click tracking of the entire artistic process, recording every time a GenAI tool was used. Naturally, this would incentivise the development of black market non-watermarking tools for artists who do not want to be tracked while they work.

Even if this intrusive tracking was ever functional, it would still incentivise artists who had used GenAI tools to just recreate AI-generated portions without GenAI. Here, the technician who is best at copying is best protected.

Strangely, the USCO Report almost seems to support this, specifically identifying GenAI’s core utility as a “brainstorming tool” for song ideation. The Report directly states that “using AI in this way should not affect the copyrightability of the resulting human-authored work”, provided the user is solely “referencing, but not incorporating, the output” [p.12]. It is difficult to fathom how exactly a reference to a song that only the author has heard might manifest itself other than through incorporating its elements. Are we really expecting authors not to use the ideas they are generating? If so, how are we tracking whether they have “referenced” or “incorporated” them?

The registration of a number of AI-generated works since the Report (with copyright limited to the non-GenAI arrangement of the elements) is revealing here. The first of these [IPKat here] - was initially denied copyright after AI use was disclosed, yet has now had copyright for its non-GenAI elements registered after the submission of a sped-up timelapse video of its production, setting an absurd precedent certain to give creators pause in disclosing any Gen-AI-use if they have not thought to film their process top-to-bottom.

A questionable foundation for a questionable policy

As the registrations continue to roll in despite delineation of specific GenAI-details becoming harder and harder, one might wonder why the USCO doesn’t just bite the bullet and cover the entire work. While there are many challenges posed to copyright frameworks if we allow copyright over miniscule-cost GenAI works, these are not the justifications relied upon in the Report.

Rather, the Report repeatedly justifies the USCO’s approach on the ideological basis that copyright is only for “human-authored works” and not “material generated wholly by AI” (emphasis added). The reasoning here is circular, as the USCO believes that the works have been created by creative tools (not with them), yet since tools have no personhood, the works cannot therefore be authored by them. Thus, the USCO determines that the human authors registering their authorship have not authored these works because authorship is for humans. In a display of poetic irony, immediately after the release of the Report the USCO granted copyright to the first “human-author” of a GenAI image - the AI creation company, Invoke.

This is illustrative, as even the circumstances that appear the most detached from human authorship – namely, a corporate AI model bulk content-generating – do not actually have an authorship problem. They have a scale problem – one which exists for both artists and AI-content companies, and one that is not necessarily best dealt with by trying to deny authorship to a portion of the works via a criterion, “GenAI”, that is as qualitatively meaningless as “software” and not possible to audit anyway.

Concluding remarks

The USCO’s approach already feels antiquated – mired in a “man vs machine” narrative that is not reflective of the deeply heterogenous uses by creatives both building and utilising an array of GenAI tools in a litany of spectrums and natures which will be difficult to meaningfully surveil and police. Rather, the USCO relies on a reductive outline of the current-state of the tools and a reductive expression of purpose for its dichotomy.

As the USCO followed the Report with a string of copyright registrations of part-GenAI works, it may fast tire of evaluating every GenAI vs non-GenAI element in each work’s process. If this is the case, the USCO should already consider the profound ramifications of not being able to separate, rather than try to build a system reliant on auditing the details of every use of ubiquitous tools in order for it to function.



Image credit: Zachary Cooper (don’t ask how!)