Can Artificial Intelligence (AI) be an author? A common reason cited against AI authorship is that AI is not a person. This article argues that the lack of personhood, in itself, does not disqualify AI as an author. Why not? Because we already detach personhood from authorship for many reasons, such as with pseudonyms, heteronyms, forgeries, or where a lack of historical documentation makes it difficult to ascertain authorship, or where a single individual is responsible for a radically diverse set of works, and in forms of legal authorship that are held by corporations. Personhood may be required for authorship, but not always. It depends on the context in which authorship is being evaluated.
The many names of Fernando Pessoa
I took the opportunity during a recent visit to Lisbon to learn about the poet Fernando Pessoa. I highly recommend Sofia Saldanha’s audio tour, which you can listen to anywhere, but is ideally experienced while walking through the streets of Lisbon.
Pessoa wrote not only under his own name, but also under a number of other names he called “heteronyms”. Far beyond simple pseudonyms, to Pessoa, his heteronyms were fundamentally other entities, different people with their own lives and stories. The most famous of these: Alberto Caeiro, Ricardo Reis, and Álvaro de Campos, are so distinct and each represents such a large body of work, that Pessoa has been described as “the four greatest poets of Portugal.”
Pessoa’s heteronyms were part of his lifelong philosophical quest to understand and inhabit the multiple potentialities of life. When reading about his life, his prolific work, his attraction towards the occult and the mystical, one senses to some extent a frustration with the limited planes of time and space to which life coalesces, and inhibits a universal existence, a pluralism of consciousnesses. It is the infinite futility of exploring the innumerable riches of Borges’ Garden of Forking Paths, the constant, compounding, irreversible regret of Frost’s Road Not Taken.
Pessoa complicates how we understand “an author”. Here is a single individual, a single body with a single brain, giving rise to multiple distinct entities, each of which fulfills the most important aspects of poetic authorship: a name, and a body of work with some thematic and stylistic coherence. Is Pessoa then, a single author, or many? Or do his heteronyms in some sense not “count” as separate authors, since they emanated from Pessoa?
Authorship does not necessarily need personhood
In a recent paper, I argue that personhood is an unnecessarily restrictive and crude bar for authorship. This is pertinent to artificial intelligence tools, which now produce language at or exceeding human performance in many areas. In many fields, from poetry and literary writing, to education, to scientific scholarship, the question is being asked: can AI be an author? Can ChatGPT, or Microsoft Copilot, or Bing Chat, or Google Bard, or Anthropic’s Claude, be the author of the texts it writes?
This question can be approached from many interesting angles. We could discuss where the training data has come from, and who the authors of those are. We could discuss the creative input from the human who has “prompted” the AI to produce a text. We could discuss the system developers, whose heuristics and biases are wrapped around the “core” models to mediate and moderate the text as it flows through the system. These are valid and important concerns, but let us set them aside for the moment.
Here we are simply concerned about personhood. Clearly, these systems are not “persons” in the sense that we understand the word colloquially. They are not human individuals with human bodies and human identities. The question is, does this lack of personhood, ipso facto, entail that AI cannot be an author? I argue it does not. What follows in italics are passages quoted verbatim from the paper with minor edits:
[…] one might ask: to what extent is authorial intent relevant to AI? We do not want to attribute undue personhood or agency to the AI system itself (Dennett’s intentional stance). Perhaps it is easier to argue that the human-AI-data complex responsible for a particular AI output is the author. However, it is also possible to argue that AI itself can be considered an author, without granting it personhood or agency.
Foucault considers the question of what an author entity, or “author function” actually is, especially when used as an interpretive resource to help establish the meaning of a text:
“Modern criticism, in its desire to ‘recover’ the author from a work [… is] reminiscent of Christian exegesis [… in determining the authors of a text when it is unclear who wrote it.] According to Saint Jerome, there are four criteria: the texts that must be eliminated from the list of works attributed to a single author are those inferior to the others (thus, the author is defined as a standard level of quality); those whose ideas conflict with the doctrine expressed in the others (here the author is defined as a certain field of conceptual or theoretical coherence); those written in a different style and containing words and phrases not ordinarily found in the other works (the author is seen as a stylistic uniformity); and those referring to events or historical figures subsequent to the death of the author (the author is thus a definite historical figure in which a series of events converge)”.
These criteria: a level of quality, conceptual coherence, stylistic uniformity, and historicity, can be applied to AI to give it an author function without also granting it agency or personhood. These ways of producing an author entity may seem imprecise in our modern world, where it is far easier to definitively establish the author of a text, and often possible to ask them directly: “did you write this?” and “what did you mean by this?” Even with these powers, unavailable to Saint Jerome, we still conceive of authors in Jeromian terms. For instance, Wittgenstein’s later work contrasts so heavily with his earlier work that references often qualify him as “early Wittgenstein” or “late Wittgenstein”, as though they were two separate people (this is also sometimes done with Marx to distinguish “young Marx”, presumably from “old Marx” but the latter is rarely used). Artists’ lives are commonly periodised (e.g., Picasso’s “blue period”). This schizophrenic segmentation shows that we are willing to detach a singular person from the authorial entity or entities they gave rise to over the course of their life. Legal conceptions of authorship have similarly long been divorced of personhood, where “investments of capital and administrative organisation” can constitute authorship.
In January 2023, several academic publishers, such as the Science family of journals, updated their editorial policies to ban the naming of ChatGPT as author on papers, some going as far as to prohibit the use of AI technologies to generate content for papers altogether (though it is unclear how this can be enforced). Ostensibly, the practice of naming ChatGPT as an author stems from a desire to attribute its “participation” in the creative process. Yet this was not a safe or acceptable solution. A reason cited by publishers for this ban is that authors are accountable for the content in their papers, and ChatGPT is not an entity that can be held accountable to anything by anyone. Publishers also interpreted the use of ChatGPT output in a paper, as the (human) co-authors plagiarising text from ChatGPT, even when named as an author. Again, we witness a community response to the challenge of selecting one imperfect notion of authorship (and therefore creativity, or plagiarism) among a number of imperfect notions. The inclusion of ChatGPT as an author in this scenario was culturally unacceptable because of one particular interpretation of the author function.
[…] With the view that creativity comes from authorial intent and discourse, when considering whether a particular episode of AI-assisted production is creative or not, we must not look only at the output itself. The output may be even be “found” or “readymade” from the training data, without attribution. A discursive assessment of creativity will consider, in what context is this output being used or presented? What is the authorial intent? What discourse is it facilitating? Who or what are we considering to be the author entity?
[…] In several historical circumstances, attribution has been deliberately avoided. Victorian women often published under masculine pen-names (e.g., Mary Anne Evans took the pen name George Eliot; the Brontë sisters and Louisa May Alcott also took masculine names) to increase the commercial acceptability of their work, and to avoid negative societal stereotypes attached to women writers. J.K. Rowling published a detective series under the pseudonym Robert Galbraith, to spare the series from the inevitable and unfair comparison to Harry Potter. Other writers published controversial or contentious ideas under pen names to avoid persecution or censorship (e.g., Voltaire – already a pen name – was known to deny the authorship of his controversial writings, preferring to attribute them to imaginary or sometimes real people). Eric Blair took the name George Orwell when publishing his memoir of growing up in poverty, to protect his family from embarrassment. Some men take on women’s names when writing for an audience comprised of primarily women, in some cases multiple authors collaborate under a single pen name. The point is that there are many legitimate reasons for subverting “proper” attribution which have no bearing on deemed creativity.
Another case where the assumed relationship between creativity, attribution, and plagiarism is challenged is in the forgery. The 18th century poet Thomas Chatterton forged numerous poems in a medieval style, attributed to a 15th century priest, which significantly influenced English, French, and German literature. Similar episodes of forgery include James MacPherson’s “Ossian” and Edward Williams’ “Iolo Morganwg”. There are many examples in the world of visual art, as well. Initially, these forgeries are considered creative on account of their content and their presumed authorship. Later, when the forgery is exposed, the creativity associated with these works does not disappear but is transformed. Today we celebrate the ingenuity of Chatterton and Williams, and appreciate the creative effort and talent required to masterfully reproduce the form of another artist, with the same conflicted fascination that we watch the protagonist in Frank Abagnale Jr.’s Catch Me If You Can. It is not without reason that they are called con-artists.
It is impossible to determine every influence on our work and attribute them, so we settle on a culturally constructed compromise on the limits to what must be attributed. The result is that the overwhelming majority of influences that feed and enable any creative work remain unattributed. Thus the re-use and non-attributive natures of AI are not universal grounds to deny creativity, or accuse it of plagiarism. We must first ask: are we applying the standards of attribution appropriate to the context? What types of reuse are we (in)visibilising? Do the authors of sources desire attribution?
Conclusion
The relationship between a text and its authors is complex. Fernando Pessoa’s heteronyms, postmodern conceptions of the author function, and historical cases where attribution has been obscured or denied for various reasons, show us why personhood is not a universal prerequisite for authorship.
Rather than adhering strictly to a notion of the author as person, we must consider authorship as a culturally and contextually contingent concept. While accountability remains important, especially in domains like scientific research, our concept of authorship must evolve to encompass emergent and collaborative forms of creation. In evaluating specific cases of AI-assisted production, the focus should be on discerning the nuanced intents and discourses which surround that production.
Artificial intelligence invites us to hold conceptions of authorship with greater subtlety and pluralism. In so doing, we can thoughtfully navigate new frontiers of expression and meaning-making that technologies like AI may afford, without preemptively closing doors through rigid standards of individual personhood.
And now, a summary poem: As consciousness shifts mutable as flame, Now mathe-manual language-making ways,
Who gives the pen its power to freely speak,
When personas and perspectives may entwine?
As webs of influence ideas bestow,
The provenance of pen, shaded, remains.
The true birthplace of words stays veiled from view.
Produce such prose that all, or none, envisioned.
The muse and machina in union raise,
Works none could claim though all be christened.
References
This post draws on the following paper:
Advait Sarkar. 2023. Exploring Perspectives on the Impact of Artificial Intelligence on the Creativity of Knowledge Work: Beyond Mechanised Plagiarism and Stochastic Parrots. In Proceedings of the 2nd Annual Meeting of the Symposium on Human-Computer Interaction for Work (CHIWORK ’23). Association for Computing Machinery, New York, NY, USA, Article 13, 1–17. https://doi.org/10.1145/3596671.3597650
More details: Download PDF • BibTeX • ACM Digital Library • DOI: 10.1145/3596671.3597650
Disclosure
I am affiliated with Microsoft, which has interests in AI technology. All views in this post and website are mine alone and are written in an independent capacity. This post and website do not reflect the views of any individuals at Microsoft. This post and website do not reflect the views of Microsoft Corporation or Microsoft Research Ltd.