The Authorship of Theseus
By Jimmy Alfonso Licon
Over the last couple of years, with the advent and widespread availability of large language models (LLMs), a typo in an essay or a stray comma in the wrong place was almost entirely a signal of haste and sloppiness. The convention was that serious writers should carefully and painstakingly proofread their work before releasing it to readers. That world has not entirely vanished (yet), but it has quietly acquired a rival. In the new age of large language models, a typo can also be read—sometimes quite reasonably—as a costly signal that a human being actually wrote the piece you’re reading. This is because writing occurs in an environment where it is often easier for writers to paste a prompt into an LLM and receive clean, grammatically correct prose than it is to write a careful, flawed paragraph yourself, at least with regard to first drafts, since LLMs still struggle with writing that moves beyond the generic. Due to this shift, the opportunity cost has flipped. That reversal of costliness alters the social meaning of imperfection.
The framework of costly signaling helps clarify this shift in social meaning. To clarify, a costly signal is a behavior that is reliable precisely because it is expensive or risky to fake. The classic biological examples involve peacocks and gazelles: the extravagant tail or the energy-wasting stotting run are honest displays for the simple reason that weak or injured animals would struggle to perform such physical feats. In social life, costly signals abound from, for example, college degrees from demanding institutions or time —consuming apprenticeships acting as costly signals of conscientiousness and reliability. What is interesting about typos in the LLM era is that they now occupy a space that looks, oddly, like this. If it is cheap and simple to run your text through an AI system that scrubs away most surface errors, then choosing not to do so involves the potential to take a reputational hit (‘that guy couldn’t even be bothered to run his work through AI; how lazy can someone be!’). Here the writer accepts looking less polished and maybe even less professional to signal that at least part of the text was written by a human without the aid of LLMs.
The second cost is subtler but no less real. Readers who are aware of these tools may wonder, when they see typos, why you did not at least ask the machine to do it. In a world where everyone knows that such systems are widely available, and where even free LLMs are good enough to fix typos, neglecting to use them begins to look less like mere carelessness and more like a deliberate refusal of an obvious opportunity to improve appearances. You could have looked sleeker, more technically competent, more in step with the cutting edge. The costliness of that choice is what allows it to serve, in the right context, as a signal of human fallibility and thus human origin. (This signal may change as LLMs become better at imitating human fallibility in the text, too; but for the moment, the signal stands).
Of course, not every typo magically counts as a badge of authenticity, and it is vital to remember that context matters. Not to mention, there is a threshold beyond which the errors in the text are no longer charming and begin to look like sloppy human writing rather than something handcrafted. What interests me is the way the baseline has moved. Before cheap, powerful LLMs, polished text was a straightforward proxy for care and competence, while errors nearly always detracted from credibility. Excessively smooth prose and oddly generic phrasing trigger a suspicion that no one sat with these sentences long enough to embed a distinctive voice in them but just copied and pasted the output of the LLMs and called it their own. Here a bit of roughness—an untrimmed sentence, the occasional error—can reassure that there is someone on the other side of the screen who added their own stylistic flavor to the LLM output.
This signaling story dovetails nicely with the metaphysics of writing where the edit history of a document begins to resemble the Ship of Theseus. The Ship of Theseus is an ancient Greek thought experiment from ancient Greek philosophers that asks us to imagine a ship with wooden planks that gradually rot and are replaced, one by one, over many years of voyaging. Eventually, every single plank has been swapped out for a new one. The question philosophers have debated for millennia is this: Is it still the same ship? If the identity of the ship depends on its physical components, then it seems like a completely different vessel. But if identity lies in continuity—in the fact that the ship was never dismantled all at once, that each replacement preserved the ongoing function and form—then it remains Theseus’s ship despite sharing no original materials with the vessel that first left port. The puzzle becomes even thornier if someone collects all the discarded planks and reassembles them into a ship somewhere else.
A similar puzzle applies to a writer who uses LLMs to generate drafts that they then revise over and over again. Imagine a writer who struggles mightily with getting started, who sits, like so many of us have, staring at a blank screen while the task expands in their mind into something impossible to begin. For that person, the main value of LLMs is their lowering the psychological barrier to entry by at least providing a rough draft, even if not a single word of the draft will survive the editing, polishing, and reshaping that writers engage in.
In that workflow, the initial AI-generated draft is the scaffolding to help the writer get started who would not otherwise be able to begin the process, or who would struggle all the more to begin. Imagine a writer who would otherwise struggle more to begin the writing having a LLM begin the process for her by spitting out a mediocre draft. And then, sentence by sentence, paragraph by paragraph, the human writer intervenes, deleting trite formulations, inserting their own examples, rearranging the structure to match the argument they actually want to make, and transforming the piece into something that is entirely their own. After several passes, it is possible that nothing of the original wording remains. Every phrase has been rewritten, every sentence deleted or radically restructured. And yet the influence of the machine remains in the very fact that the writer did not begin from an empty page. The identity of the text, like the identity of the ship whose planks have all been replaced, is preserved nowhere else except, perhaps, in the causal history of the piece (e.g., it would appear in the edit history).
This raises uncomfortable questions about how we should describe such a piece of writing and what, if anything, it signals. If the final text results from human judgment at every point that matters, then it is strange to say that it is inauthentic work merely because the earliest ancestor draft originated from LLMs and not the hand of the writer herself. At the same time, it would be disingenuous to pretend that the tool played no role. The jumpstart it gave may have been exactly what allowed an anxious or blocked writer to move from inchoate thoughts to articulated prose. The time savings may not show up in the number of minutes clocked, but in the reduction of friction at the outset, in the way it turned a psychologically daunting task into one that felt manageable enough to begin.
When we place this Ship of Theseus model of AI-assisted drafting next to the earlier point about typos as costly signals, an interesting tension appears. On the one hand, writers may increasingly want to mark their work as authentically theirs, tempting them in the drafting process to leave in a bit of imperfection, to resist the urge to feed every paragraph through an LLM for polishing. On the other hand, those same writers may quietly rely on LLMs at the very beginning of their process exactly because those tools make it easier to begin the writing process—like jump-starting a car, only applied to the writing process instead.
Typos have become costly signals of authenticity because of a new, cheaper way of producing polished language that has additionally brought into the foreground puzzles about the metaphysics of drafting and writing.
Jimmy Alfonso Licon is a philosophy professor at Arizona State University, where he has won over twenty teaching awards. He previously taught at the University of Maryland, Georgetown University, and Towson University. He works on issues in political economy, ethics, AI, and God.
As I read your thought-provoking article, I was reminded of an e-newsletter on AI in business that I co-published with a colleague. Sometimes just to get it started, Ed (who was used to writing on deadline), would get me a first, what we called "dirty draft" even if it was very rough and quick. Since I was the better wordsmith, I would then polish it and make my own rewrite/contribution. On some of those articles, I felt like an old-school, deeply involved editor - that if we weren't sharing authorship for the entire newsletter, he would have been first or sole author.
In your example, when the AI creates the first draft, even if all the points are subsequently revised by the human writer, it felt like the AI I should get at least co-authorship credit, or perhaps even first authorship. It's interesting how my opinion of who would be author was influenced by whether the first draft was created by a human or an LLM.
Thanks for a fascinating read that brought me to think about new (and old) questions!
Incredible essay. Cheers for sharing!