What does it mean to speak well and what connection—if any—is there between speaking well and speaking the truth? In his 2024 book, Seeing Double, Raymond Geuss identifies five senses in which ancient rhetoricians believed someone could speak well. A person spoke well if they could persuade other people. They spoke well if their speech were particularly beautiful or well formulated. A person spoke well if they gave good advice and spoke appropriately and in turn. Finally, someone spoke well if they told the truth. To these senses, Geuss adds another, inchoate initially in the classical period but starting to come into its own: someone could be said to speak well if they spoke correctly.1 All of these ways of “speaking well” were, of course, predicated on the assumption that it was a human being who was speaking. The whole model of classical eloquence depended on there being an essential connection between a given person, their means of expression (voice or writing), and a character (or personal style) that those means of expression conveyed.
What does it mean when, due to the widespread availability of large language models, the connection between a given piece of text and the person who claims authorship of it can no longer be guaranteed? Can a text generated by an LLM (that is, a fundamentally authorless text) speak well or speak the truth?
The digital humanist Leah Henrickson argues that the reason we care so much about distinguishing human writing from computer-generated text is because we remain wedded to a naïve way of interpreting the reading experience that she terms the “hermeneutic contract.” The contract encourages readers to believe that reading forges a semi-mystical connection between themselves and what they imagine the author’s mind and intentions to be, a set of idealised projections that are all really in the reader’s head. AI authorship breaks this contract.2 Because the author in the reading process is simply the reader’s imagined or projected author, it does not matter whether the author is a machine or a human, just so long as the reader “finds meaning” in the text.3
Illustrating what she sees as the increasing pointlessness of distinguishing human from machine authorship, Henrickson observes that most people cannot reliably detect when they are reading AI-generated content.4 But the question of origins cannot be evaded that easily. Mistaken subjective beliefs have no bearing on how something was actually produced. A magic trick is not “magic” just because no-one in the audience knows how the sleight of hand was achieved, or if people find meaning in the belief that they have witnessed a miracle. It matters whether one lives in a world where magic is deemed possible or not. The real mystification, the actual fetish, lies in suggesting that the distinction is meaningless—that it is uncouth (or a sign one has not yet got with the programme) to point to the technique underlying the illusion.
There remain strong ethical and intellectual reasons for attempting (at least) to maintain a distinction between human and machine authorship. When one Australian academic confided to a Guardian journalist that he had students who, when asked, “could not tell [him] a single thing about the entire semester, yet” had each submitted “a beautiful essay,” the seeming “beauty” or otherwise of those essays is irrelevant.5 Without a clear-cut connection to a student-author, essays like this become simulacra of student achievement rather than evidence that anyone has actually met the learning outcomes. Without authors, in other words, they are also without function (and thus meaningless). The problem is what some AI researchers have come to call “inseparability of content.”6 An essay that took a student several weeks to write based on wide reading and a thorough understanding of all course concepts has become indistinguishable from instantaneous GPT-6 output generated by pasting the essay question into the prompt box. Even worse, an essay generated in this way using the latest models will likely read as more accomplished than anything an undergraduate student—someone still in the process of learning the discursive ropes—will be able to produce. “It is,” as Geuss puts it, “a serious error to confuse an image with the object represented in the image.”7 And yet to give essays like this passing grades—as the Guardian’s academic whistleblower claims university administrators forced him to do—is to make precisely this mistake.
Indeed, the simulation analogy for generative-AI could be extended further. Identifying Renaissance stucco decorations as the prime example of the “counterfeit” objects representing the first order of simulacra, Jean Baudrillard described stucco as “a single new substance” that “embraces all forms” and “imitates all materials.”8 Stucco (like cast concrete) could be shaped and modelled to resemble almost any form, while simultaneously consisting of material that is valuable on its own terms only for its ductile qualities. Generative-AI output is arguably similar. Ingested as pre-training data, the human cultural legacy is now endlessly extrudable, imbued with the fluid potential of stucco, concrete, or the three-dimensional plastic “ink” of the 3-D printer.9 And yet while such output can be prompted into assuming a superficially convincing and impressive textual shape, it arguably remains on the level of an image or a simulacrum. In being like a human-produced cultural artefact, it only serves to represent what human culture is not.10
Despite the temporal gulf that separates the world of authorless chatbot statements from that of classical rhetoric, something like the traditional framework for judging whether one can “speak well” continues to govern our responses to generative-AI. For Rafael Alvarado, large language models are incapable of speaking the truth because, lacking an author, they have no access to the means by which human beings “seek or validate truth claims.” “Even when LLMs produce ostensibly true sentences,” he writes, “they do so accidentally—these sentences are all cases of what epistemology calls unjustified true belief.”11
When he was accused earlier this year of using ChatGPT to write his Commonwealth Prize-winning short story, “The Serpent in the Grove,” the Trinidadian writer Jamir Nazir fell back on “correctness” as his defence. When critics ridiculed his simile “The girl smiled like sunrise over a sink,” Nazir hit back by claiming that the image came straight from a childhood memory of reflected sunlight streaming into the family kitchen. “It’s from my lived experience,” he insisted.12 “I lived this.”13 One sense in which someone can be said to “speak correctly,” as Geuss notes, is to be seen to speak in a particularly “pure,” authentic, or unmediated way.14 Nazir’s claims to ownership over his story essentially bypass questions of linguistic mediation (that is, language and style) completely. Both prose and fictionality disappear into the immediacies of pure experience, which is beyond question (or more to the point, perhaps, Pangram analysis).
Claims like this may suggest why those interested in maintaining a distinction between human and machine authorship keep coming back to the question of style. Literary criticism (in its destructive mode) remains how the battle of what counts as literature is fought. When the AI researcher Tuhin Chakrabarty and his students looked into the authorship of “The Serpent in the Grove,” questions of style, mode, and genre underpinned their analyses, even though (as computer scientists) they may not have seen it in quite that way. Interpreting Nazir’s short story as LLM output with a surprising degree of reliance on fragments of online fan fiction evidently memorised from pre-training data, Chakrabarty wrote that large language models “stitch together verbatim expressions from different sources,” regardless of genre or position on the high/low taste continuum.15 It is this obliviousness to formal and generic boundaries, he suggested, that perhaps explains the characteristically jarring and “off” nature of much AI-generated figurative prose. What can seem at first glance like an original and imaginative use of language—such as Nazir’s line “the air sweet with cane and forgetting,” the latter part of which one of Chakrabarty’s PhD students traced to an online review of a Better Call Saul episode—is best read as a kind of intentionless trompe-l’œil effect, one that simulates surface coherence while actually stemming from a machine-generated confusion of wildly different sources and semantic elements.
2
“As each individual has his own appearance, his own voice, his own character and disposition, so each has his own style of writing,” Erasmus wrote in his 1516 edition of Saint Jerome.16 The idea that one’s personal style was an intrinsic part of one’s own self had become widespread in Britain by the eighteenth century, where the world’s first copyright law (the 1709 Statute of Queen Anne) made explicit the idea that the products of one’s pen were forms of personal property. In 1778, James Boswell recorded a conversation with Samuel Johnson and Thomas Percy where the three men discussed the personal styles of painters and writers in detail:
We talked of the styles of different painters, and how certainly a connoisseur could distinguish them. I asked, if there was as clear a difference of styles in language as in painting … so that the composition of every individual may be distinguished? JOHNSON. ‘Yes. Those who have a style of eminent excellence, such as Dryden or Milton, can always be distinguished’. I had no doubt of this; but what I wanted to know was, whether there was really a peculiar style to every man whatever, as there is certainly a peculiar hand-writing, a peculiar countenance … always enough to be distinctive …. JOHNSON. ‘Why, Sir, I think every man whatever has a peculiar style, which may be discovered by nice examination and comparison with others.17
It is just these sorts of longstanding relationships between personal style, the act of writing, and the essential self that AI-generated prose radically undermines.
To a large extent, of course, the advent of generative artificial intelligence is simply accelerating a crisis of meaning that was already well advanced when the first public-facing LLMs went online in 2022.18 As Nils Gilman puts it, gen-AI is “acting as a catalytic solvent, titrating out the incoherence that was always there.”19 The student term paper was already a kind of travesty of the essay form. Wordcounts, rigid structural templates (such as the five-paragraph essay form), and the coercive effects of assessment itself had already rendered what Montaigne invented as a means for open-ended self-exploration into the kind of highly directed and purposive activity that the essay was once explicitly opposed to.20 The fact that Claude can now extrude a more plausible-looking end of module assignment than most undergraduates are capable of writing underlines the fact that the essay had long ago become a formula, and such things are now easily modelled, mirrored, and automated.21 By invoking in us the longstanding suspicion and unease associated with images and simulacra, LLMs encourage us to ask what value we continue to find in the unmediated, the un-ingested, the live and non-formulaic, the imperfect or not yet fully formed, and that which might yet survive intact its proliferation into innumerable fragments and copies.
Note: this is an expanded version of some opening remarks made during the “Authorship” plenary panel at the “Telling the truth? Authorship, audiences and authenticity across discourse, texts and narratives” conference, The Open University, 5 November 2026.