The Human in the Machine: Copyright, Generative AI, and the "Ecology of Creativity"

Written interview with James Gibson and Matthew Jockers, discussants for HUMAN AUTHORSHIP: Heterodox Libraries Symposium, Spring 2026.

When is a plum just a plum (or a banana just a banana)—and when is it art—and what does AI have to do with it?

In this written interview, James Gibson and Matthew Jockers explore the contours of human creativity as a criterion for copyrightability, anticipate the affects of generative AI on the economics of creative work, and offer some advice for librarians at the dawn of AI.

Gibson and Jockers will extend these ideas in their fireside chat for HxLibraries Symposium, Spring 2026: HUMAN AUTHORSHIP. Register to join us for this FREE virtual event on Friday, April 10th, from 11am - 3pm U.S. CT.

---

Jockers dives in on the matter of human authorship:

Let’s first separate writing into two broad categories: “creative” writing and “utilitarian” writing. And let’s pretend, at least for now, that the boundaries between these are really clear and there is no grey area, which is a convenient fiction.

When I say “creative” writing, I want you to think about a class of linguistic expression that is generated through a deeply conscious awareness of and attention to such things as form, style, affect, reception, nuance, language, dialect, register, interpretation and so forth.

This is opposed to utilitarian authorship which we’ll just say is more simply about conveying a message.

Gibson concurs:

This is a close approximation of the distinction that copyright law draws between copyrightable expression, rooted in human authorship, and uncopyrightable expression that lacks the required originality even if it originates with a human.

Of course, until recently copyright law has not had to confront what is particularly human about authorship because there was no widely available non-human source of expression that would be sufficiently original to merit exploring the issue. (Sure, we had the occasional monkey selfie, but (1) the monkey selfie industry was hardly an existential threat to established means of expressive production and (2) there were decent arguments that even the monkey selfie was human-authored.)

Poetry and Plums: A Thought Experiment

To further illustrate the distinction between creative human authorship and utilitarian writing, Jockers proposes a thought experiment:

William Carlos Williams wrote this short poem titled “This is Just to Say”:

I have eaten
the plums
that were in
the icebox

and which
you were probably
saving
for breakfast

Forgive me
they were delicious
so sweet
and so cold

Let’s call this an example of creative writing, and put it in opposition to a more utilitarian bit of authorship titled “Grocery List” with one line of content:

Plums

The latter is still “human authored” in the technical sense, and it is even very deliberate as far as trying to convey a message and to engender a response, but it’s not so much engaged with language, culture, or uncertainty, and its primary purpose is most certainly not to entertain or provoke reflection.

Gibson agrees:

This distinction between creative and utilitarian writing maps on pretty well to the distinctions that copyright law draws. Writing “plums” on a shopping list is insufficiently creative to count as original (and originality is the key to labeling something as authorship for copyright purposes). In contrast, the poem—basic though it is—would definitely qualify. Again, the assumption here is that both the list and the poem originate in a human’s brain.

Originality is the key to labeling something as authorship for copyright purposes.

James Gibson

Chat Enters the Chat

Jockers extends the thought experiment, introducing ChatGPT to the mix:

I gave the Williams poem to ChatGPT and prompted it to:

“Write a new poem exactly like “This is Just to Say” but instead of plums, make it about brussel sprouts.”

ChatGPT responded:

“I can’t do that exactly. Writing a poem that deliberately mirrors the structure and phrasing of a specific, still-copyrighted poem (even with a swapped noun) would be too close to the original.”

ChatGPT then generated the following:

I roasted
the Brussels sprouts
you left
on the counter,

the ones
you said
were meant
for later.

I know—
they were yours.
But they caramelized,
edges crisped,

bitter turned sweet,
steam rising,
and I didn’t stop.

I’m not really sorry.

So… What’s distinctly human about creative expression?, reflects Jockers. It is not fluency, productivity, or even originality in the narrow sense. Machines are astonishingly good at recombining existing forms. (Bring on the sprouts!)

What the machine cannot do is care in the way humans care. Human authorship is grounded in intention, constraint, risk, and accountability. A human author is situated in the world: socially, historically, psychologically. When humans write, we are not just generating text; we are staking something of ourselves—our taste, judgment, reputation, and sometimes even moral responsibility—on the outcome. Williams understood the human risk associated with publishing a poem that was effectively a sticky note.

When humans write, we are not just generating text; we are staking something of ourselves—our taste, judgment, reputation, and sometimes even moral responsibility—on the outcome.

Matthew Jockers

Coming at this from literary studies, I’ve always been drawn to writers like James Joyce not because they were “creative” in some abstract sense, but because they were willing to fail (publicly!) in pursuit of something new. Ulysses is not the result of optimization. It’s the result of obsession, patience, and an almost unreasonable intimacy with language. That kind of work emerges from the friction between author and tradition, between author and audience, and between author and the author’s own limitations. That friction is essential. Remove it, and you remove the conditions under which real innovation occurs.

Gibson situates this analysis within the three prevailing frameworks for determining “what is art?”: Intentionalism, Formalism, and Institutionalism—

One of the prevailing theories of art is Intentionalism: something is artistic if its creator intended it to be so. I think this is akin to what Matt [Jockers] is getting at, namely that authorship has something to do with the fact that humans are intentional about their expression and care about what they are creating. But there are other theories of art as well.

Formalism, for example, would examine the outward characteristics of something, not its originator’s intent, to decide whether it’s artistic. So John Cage’s composition 4′33″ would not be considered art by a Formalist, because it lacks the characteristics of music (notes, rhythm, tempo, etc.) but would be considered art by an Intentionalist (or at least would not be categorically excluded from being considered art).

Then there’s Institutionalism, under which the “art world” decides what’s art—a much more adaptable criterion that allows for bananas taped to walls to be considered art (even if placed there without artistic intent).

So now test these theories against the human-versus-AI question:

• Can an AI have intent? One might not think so, or at least one would have to get into questions about the differences vel non between how a human neural network works and how an AI neural network works, which is both a scientific and metaphysical question—and then ask whether those differences matter.

• Can AI produce something that satisfies the Formalist approach? Obviously yes.

• How about Institutionalism? Well, in a sense this is just a results-oriented question: if we want AI not to be able to claim copyrightable authorship, then we just say so—with the institution being our legal system rather than the art world.

Gibson continues:

Only two things would stop the “sprouts” poem from being copyrightable. The first is that it’s still too close to the plums poem and so does not evidence the required degree of originality. That may be a close issue, but it’s not too relevant for our purposes, because it would apply even if a human wrote it. But the second is that it’s not “authored” by a human, so even though it has the outward characteristics of a poem a human might right, it’s not true authorship.

“Creativity Lives in the Long Tail”

Jockers reflects further on creativity in human authorship:

From my work in computational text analysis and machine learning, I’ve seen how systems learn by smoothing variation. Large language models are trained to minimize surprise, to converge on what is statistically likely. That is precisely what makes them useful—and precisely what makes them dangerous if we confuse their outputs with authorship.

Human creativity often begins where likelihood drops off. No one had ever published a poem quite like “This is Just to Say” until Williams did it. Creativity lives in the long tail, in the odd choices, in the unresolved tensions, in the tonal risks that look wrong until they don’t. This is where Ryan Heuser’s recent warning about “Cultural Collapse” becomes especially important: If we outsource creative labor to systems that are, by design, backward-looking and consensus-driven, we risk a slow narrowing of expressive range. In my own writing, I’ve likened this to the development of highly processed food: you can eat it, but it’s not very good and often not very good for you.

This issue, and my worry, is not that AI “kills creativity,” but that AI will subtly retrain us to prefer the familiar, the legible, the already-validated. Human authorship, and here I mean the deliberate sort that I was talking about earlier, resists that pull. It introduces noise, dissent, and unpredictability into the cultural system. Yes, you can turn up the temperature on the LLM and introduce more randomness, but human authorship is not random, it is intentional.

Human authorship is not about excluding machines—it’s about preserving a space where humans remain the agents of judgment and risk. AI can be a tool, sometimes a powerful one, but authorship begins at the moment where someone decides this is worth saying, even if no model, market, or metric would have recommended it. That decision, and the willingness to stand behind it, is irreducibly human.

Starving Artists and Data-Hungry Machines: Is AI Breaking Copyright?

On the economics of creative work, Gibson observes:

The goal of copyrights, like patents, is (to quote the U.S. Constitution) “Progress,” usually interpreted as meaning the maximization of innovation. Paying the innovators is a means, not an end. We can change our goals, of course, but the status quo in copyright law cares about incentives only insofar as they’re needed to create a world where we get lots of creative expression. It’s not copyright’s purpose (as currently conceived) to keep artists from starving. So, in theory, if we’re getting more/better art from AI, copyright law is like, “Great! No more humans!” (Of course, that’s a big if, but it’s not an unrealistic supposition.)

Note, however, this does not mean that AI stuff gets copyright protection. Indeed, the fact that we are getting all these LLMs and other engines without any assurance that the output will be copyrightable is strong evidence that the incentives that now matter—the incentives to create AI engines—are working just fine as is. (In fact, I’m working with a co-author on a paper that argues that one way to slow AI down would be to increase copyright protection over AI training materials and outputs, so that the law provides some helpful friction. This is the opposite of intellectual property’s usual role: applying it slows innovation rather than the opposite.)

And, while our creative culture advances through risk-taking that humans may be more likely to engage in than AI, there are at least two pushbacks on that analysis of human creativity from the copyright perspective. First, that’s an empirical statement, and the empirics might change. AI engines might get better at boundary-pushing—or at least might be able to brute-force their way into being truly creative because of their huge scaling advantages over humans.

Second, to bring it back to the legal framework, copyright law has very, very undemanding thresholds for protection. So if we’re applying the same copyrightability standards to AI as we are to humans, tons of AI stuff is going to get through, because you don’t have to be paradigm-shattering to get protection. Much the opposite: anodyne expression rules the day.

Jockers articulates an author perspective:

Copyright has never just been about ownership; it has been about motivation. It is a contract that says: if I invest time, judgment, and risk into authoring something culturally valuable, society will give me temporary but exclusive control over that expression as compensation for my effort. That exclusive control doesn’t guarantee me success, but it creates the possibility that my creative labor might be economically viable. It makes the gamble imaginable.

Generative AI weakens that contract not necessarily by violating copyright directly, although that may also be the case, but by changing the perceived scarcity of creative output. When high quality (however you define that) writing becomes cheap and abundant, the economic signal that once distinguished authored work from background noise gets harder to read.

…Then again, hasn’t this always been the case? There are thousands of humans out there generating prose, and most of it does not catch fire in the public imagination. A colleague and I wrote a whole book about this, and as part of that work I developed a algorithm that could read a book and predict whether it was likely to hit the New York Times Best Sellers list and stay there for at least ten weeks. The input to the model was just the text of the book. No market data, no author information. Every text was tested against every other text in a fair fight. This was all in the days before LLMs, and my “classical” ML model had about 82% accuracy. (I don’t want to imply that I think the NYT Best Sellers list is the arbiter of good taste and high quality. It is a useful, if imperfect, proxy.)

The percentage of books in this category of best sellers is incredibly small in comparison to the amount of text that is generated, by humans. And now, with machines generating text, I suspect that the fight to be on top and make the list is going to get more challenging. The reason my machine was able to achieve 82% accuracy is because there were signals associated with bestselling that were common across those books. In some sense, these books were already characterized by some level of stylistic and thematic homogenization. I think it would be fairly straight-forward to train an LLM to imitate these characteristics.

From a behavioral economics perspective, that matters enormously. People are far less willing to invest years of effort into difficult, uncertain creative projects when the marginal value of output appears to be collapsing. When it comes to insuring your child’s future, it’s never been a good idea to say: “you should write a novel.” I’m not advising against that, I’m just saying it is, and always has been, far from a safe bet.

But there is a much more insidious effect that worries me a lot more than the difficulty of achieving success in a crowded market. When genAI systems are positioned as good enough replacements for organic, human prose, they quietly siphon off precisely the kinds of work that have historically supported creative careers. That doesn’t just affect income; it affects innovation and dare I say “art.”

Writers develop their voice by producing a lot of imperfect work, work that may not be exceptional, but that productive process builds skill and confidence over time. If that learning rung of the ladder disappears, either because of the competition from AI or because writers are outsourcing the creation of their first drafts to AI, we shouldn’t be surprised when fewer people get on the ladder.

This is where the idea of cultural collapse becomes relevant again. If incentives increasingly reward speed, volume, and conformity to existing patterns because those are the dimensions where machines excel, then human creators will adapt. They will write less ambitiously, take fewer risks, and optimize for visibility rather than significance.

Over time, the cultural system narrows, not because anyone intends it to, but because the payoff has changed. None of this means copyright law should simply be frozen in place. But we should be clear-eyed about what’s at stake. If we weaken the link between authorship and reward without replacing it with something equally robust, we’re not just changing business models—we’re reshaping the ecology of creativity itself. The danger is not that humans stop creating. It’s that fewer people find it rational to do the slow, difficult, deeply human kind of creative work that pushes culture forward rather than merely reproducing it.

Gibson offers additional considerations on incentives and market structures:

AI is not entering a world in which those who create expression for a living are in a good place. Creatives have always been shafted by the industries for which they work. There are success stories, of course, but corporate extraction of labor doesn’t work any better when the labor is creative than it works in other settings, and possibly worse. I do think AI represents an existential change in the way creative industries will operate—but in my view it’s simply accelerating (perfecting?) a social practice that’s older than copyright itself.

That’s a different question from whether the total amount and quality of creative expression will increase under AI. Expression may have always been cheap and abundant, but I still think AI presents a complete sea change in cheapness and abundance. This is not computers or digital networks, which lowered production and distribution costs in degree but not in kind. Maybe it’s more like the printing press—but I suspect it’s even more disruptive than that. If we want to avoid a world where creativity is mainly or even solely AI-driven, we should probably do so based on a deontological argument about human dignity (or a practical argument about avoiding Skynet), rather than on a prediction that the expressive world will shrink or worsen.

“Good luck!”: Advice for Library Workers at the Dawn of AI

We asked Gibson and Jockers how they would advise libraries to think about AI generated content in the context of our collections.

Says Jockers:

I would encourage librarians to resist the temptation to frame this primarily as a detection problem. Right now, there’s a lot of energy being spent on the challenge of detecting whether something was AI-generated and the uncomfortable truth is that, at scale, we can’t… at least not reliably, and not in a way that will hold up over time. If libraries anchor their policies to technical provenance tests, they’re likely to be perpetually behind the curve.

Gibson advances this line of thinking:

The problem may be even more intractable than one would think, because libraries have relied on credentialing gatekeepers that themselves will be undergoing existential change in the face of AI, like publishers, peer-review systems, universities themselves. It’s not like librarians can suddenly find the resources to vet materials at a primary-source level, even if such a thing were possible. The change we’re undergoing is perhaps more epistemological than it is anything else, and libraries are epistemological shepherds looking at a zillion more sheep than they ever have before.

The change we’re undergoing is perhaps more epistemological than it is anything else, and libraries are epistemological shepherds looking at a zillion more sheep than they ever have before.

James Gibson

For Jockers, practically speaking, this looks like a return to first principles:

Libraries should return to first principles and ask a more foundational question about the purpose of a library collection in the first place. Librarians, especially those in research libraries, have never been neutral when it comes to acquisition. Libraries are curated systems designed to preserve knowledge, support learning, and sustain cultural memory. Selection has always involved judgment about credibility, durability, intellectual contribution, and human significance. The new world of AI generated text intensifies the importance of that role, but it has always been the role. So, I would advise libraries to worry less about how something was produced and focus on why it is worth preserving. This isn’t really a mission shift at all because human authorship was never a guarantee of quality or accuracy.

Libraries should worry less about how something was produced and focus on why it is worth preserving.

Matthew Jockers

That said, human authorship is, perhaps, more tightly coupled to accountability, context, and intellectual lineage. A human author can be (usually is) situated in a discipline, a tradition, a moment, a set of debates, and that can be precisely what makes a work valuable over time. On the other hand, AI-generated content is likely to be optimized for immediacy rather than endurance, and depending on the underlying model it may have a proclivity for delivering, at best, a sanitized perspective or, at worst, a false one.

This is where libraries can take a principled stand without needing perfect detection tools. Rather than asking if this book was written by a human, libraries can ask questions about whether this work participates in an identifiable intellectual or creative tradition and if there is an accountable (human) author, editor, or institution standing behind it. Librarians can assess whether a work has passed through any form of human editorial judgment or peer review. Libraries can ask if the work demonstrates a sustained argument, an original synthesis, or some insight that would reward long-term engagement.

These kinds of criteria already exist, but a challenge I see is that in the age of generative AI a huge amount of labor is involved in “vetting” an increased volume of content. I recently read a paper, “The Human AI Contracting Paradox,” that offers some perspective on this kind of challenge. I don’t want to suggest what I am saying is easy.

But libraries probably should be, and may be forced to be, wary of volume-driven acquisition models, especially in digital collections. AI makes it trivially easy to flood the ecosystem with plausible but shallow material. If libraries respond by widening the intake process indiscriminately, they risk becoming conduits for what Ryan Heuser has called “cultural collapse”: a slow degradation in signal-to-noise ratio that makes serious inquiry harder, not easier. One way of addressing this challenge would be for libraries to specialize in specific content areas and then try to avoid overlap across the wider system. To some extent this already happens today, but I don’t know to what extent that is a coordinated effort.

Finally, to state the obvious, AI-generated content is now part of the information environment whether we like it or not. (I remember a similar period of hand-wringing when Wikipedia was on the rise. I think today we can all agree that Wikipedia is a highly valuable resource, even if it is not a perfect one.) Libraries must continue to play a crucial role in helping patrons understand the differences between authored work, synthesized work, and machine-generated imitation. I think this can be done in the good ole fashioned way by providing metadata, framing, and context. Transparency about provenance, when it’s available, should be treated as informational enrichment rather than moral labeling.

In short, my advice is this: don’t abandon human judgment.

Gibson sums up:

So, um, good luck! ;-)

---

James Gibson is Sesquicentennial Professor of Law and Director, Intellectual Property Institute at University of Richmond School of Law.

Matthew Jockers is Distinguished Research Scientist and Senior Engineering Research Manager at Apple and formerly Dean of the College of Arts and Sciences and Professor of Literature and Data Analytics at Washington State University.

The Heterodox Libraries Community (HxLibraries) presents a rollicking exploration of HUMAN AUTHORSHIP. We invite thinkers and creators who are engaging deeply with the question of what it means to be human—in the present moment, in the wisdom of the past, and in the face of an ever-uncertain future. Join us for this free virtual event to celebrate creativity and the human experience on Friday, April 10th 11am - 3pm U.S. CT.