Provenance Anxiety: Death of the Author in the Age of Large Language Models
David M. Berry
Generated on Google Gemini Dec 2025: Prompt: "A high-quality, realistic, and colorful 3D infographic diagram illustrating the 'Jagged Frontier' of AI capabilities. The image shows a progression from left to right using distinct shapes".
"In 1950, it may have been possible to choose other paths. In the third decade of the new millennium, however, our reliance on cognitive assemblages and computational media has progressed so far that there is no going back"
N. Katherine Hayles, 2025.
"Though the sway of the Author remains powerful... it goes without saying that certain writers have long since attempted to loosen it"
Roland Barthes, 1977.
"What matter who's speaking, someone said, what matter who's speaking?"
Michel Foucault quoting Samuel Beckett, 1977.
Earp et al describe a scenario in their paper in which a researcher prompts an LLM with fragments of an argument they are trying to make and receives a coherent paragraph in response. This fictional writer asks the LLM to make "light edits", with it creating a text that the fictional writer is unaware closely mirrors the substance and flow of "Smith (1975)". Earp et al treat this as a problem of failed attribution ("epistemic fairness").[2] However, it is notable that Earp et al do not distinguish between automation and augmentation in their discussion of LLM use in academic writing – and human creativity can be magnified when humans, augmented, become "centaurs" (half human, half machine) (see Berry 2014: 8-9).[3]
At the end of their piece they reveal in the acknowledgements that they used LLMs in writing the article themselves. This follows from an earlier paper suggesting an ethical requirement to declare LLM use (Porsdam Mann et al 2024). This raises a curious question as to why they single out this particular computational tool for disclosure? LLMs are software, yet we do not require scholars to declare their use of word processors (which increasingly incorporate LLM features in obscure ways), spell-checkers and grammar-checkers (which have long suggested stylistic changes), digital databases (which algorithmically retrieve and rank sources), search engines (which shape what literature scholars encounter), or bibliographic management software like Zotero (which automates citation formatting and suggests related works). The demand to declare LLM use whilst treating other computational tools as transparent instruments seems to misunderstand the extent to which scholarly work has been softwarised since the 1990s. Indeed, in Benjaminian terms, it is an attempt to restore the cult value of "aura" to the manuscript (2008). Every stage of contemporary academic writing, from literature discovery through algorithmic search, to note-taking in computational environments, to drafting in software that autocorrects and suggests, to citation management through automated tools, is already mediated by computational processes that shape intellectual labour in ways that remain largely unexamined and undeclared. The selective anxiety about LLMs thus appears less like a principled ethical stance and more like a reaction to a threshold where the computational mediation of thought becomes uncomfortably visible.[4]
This selective anxiety points to a deeper question of whether this provenance problem is actually a new crisis or is it the materialisation and acceleration of something that was always true. That is, that academic writing has always been a kind of bricolage, authorship has always been a construction, and that the fantasy of individual academic creation has never corresponded to the actual practices of scholarly production. This would mean that LLMs do not introduce a radical break in how academic knowledge is produced, rather, they reveal and intensify processes that were always operative but remained partially obscured by romantic ideologies of authorship and creativity.[5]
I would argue that the anxiety surrounding LLMs is best understood as a crisis in the mythology of academic authorship rather than in its actual practices. Their belief that LLMs encourage "'cryptomnesia', in which [their fictional academics] reproduce ideas that they have encountered previously but mistakenly believe to be original" seems to me to be problematic (Earp et al 2025). In contrast, I suggest that the LLM makes visible what was always the case, that is that scholarly writing proceeds largely through assemblage and recombination, complete attribution is largely impossible (e.g. the bibliography for every article and book one had ever read or consulted would be larger than the article itself). Thus a scholar's "own" ideas were already assemblages of half-remembered readings, classroom discussions, conference conversations, theoretical frameworks absorbed and internalised. The notion of originality, I suggest, serves only to police boundaries and distribute academic capital rather than to describe the processes of knowledge production. As Foucault (1977: 138) argues, "the author – or what I have called the 'author-function' – is undoubtedly only one of the possible specifications of the subject and, considering past historical transformations, it appears that the form, the complexity, and even the existence of this function are far from immutable. We can easily imagine a culture where discourse would circulate without any need for an author".
In fact, we can see in recent attempts to use so-called "AI detectors" to shame authors, students and even casual writers that the "author-function" is being reinforced through technical means, even if these systems are themselves prone to error and hallucination, and suffer from the same black boxing that hinders our understanding of LLMs (see deBoer 2026; Jancer 2025; Klee 2026; Requarth 2026; Taranto 2026; Titcomb 2026; Vara 2026). These "detectors" are created by companies like Copyleaks and Pangram, private companies that seeks to profit from this provenance anxiety (Pangram describes its tool as a way to “protect editorial integrity”). Indeed, they contribute to "build[ing] a culture of accusation... and you get something not only brittle but perhaps even falsely reassuring: a system that comforts readers and writers that the A.I. problem has been solved while harder questions go unasked" (Requarth 2026). As Kirshner (2026) recounts, this is starting to have perverse effects on the practice of writing itself,
A.I. writing is now in my water supply, whether I can identify each new gulp or not. It feels inevitable that I’ve started to internalize its tics, even the most infamous ones, like incessant em-dashing and overreliance on the “It’s not x but y” structure... I have started to self-police against this outcome. A few times lately, I’ve caught myself modifying my own work to sound “less like A.I.” In the process, I’ve probably picked words I wouldn’t have settled on otherwise. It feels dystopian—and I doubt I’ve succeeded... A time or two over the past few weeks, I’ve caught myself in a cursed loop: Draft a story with my human brain and human fingers, then put it into Pangram before it appears anywhere on the internet. Usually, the model concurs that I did it myself. In a few instances, it determined that single-digit percentages of my writing were either A.I.-generated or A.I.-assisted Kirshner (2026).
This is ironic, because when pressed on the accuracy of their models by the Wall Street Journal, Taranto (2026) concluded,
This raises more questions than it answers. What exactly is meant by “the underlying assessments“? How can they be “similar” when the results for the articles the study spotlighted are materially different? If various “iterations” of Pangram produce such divergent results, why should anyone have confidence in the tool’s reliability? I posed these questions in an email to Ms. Russell [research scientist at Pangram] and Mr. Spero [CEO of Pangram Labs]. Both declined to answer. They didn’t reply when I asked for an explanation of the discrepancies between the paper and the online database (Taranto 2026).
I think this relates interestingly to what I have elsewhere termed the Inversion (Umschlag) (Berry 2025). This I describe as a critical threshold where machine-generated content becomes somewhat indistinguishable from human writing but also actively reshapes our understanding of it. The provenance problem is therefore a controversy that lets us see the clash between traditional modes of attribution (predicated on stable authorial origins) and what I call diffusionisation, a process through which knowledge and cultural production become subject to probabilistic dissolution and reconstitution via computational processes.
What becomes visible under these conditions of diffusionisation? Italo Calvino's essay Cybernetics and Ghosts (1986) offers a suggestive metaphor where he explores literature as a combinatorial system. Calvino's reading of the prisoner Edmond Dantès in The Count of Monte Cristo, from the book t zero (1969), provides an interesting analogy with our situation under the Inversion. Calvino describes Dantès attempts to imagine the perfect prison from which escape is impossible, reasoning that if he succeeds, either he has perfectly modelled his actual prison (and must accept his fate), or he has imagined a prison even more secure than his actual one, suggesting his real prison has a flaw enabling his escape. Just as Dantès uses his model to identify points of difference from reality, we might use our encounters with LLMs to identify what cannot be captured by statistical recombination, that is, what resists diffusionisation. The provenance problem becomes an epistemological question as the very untraceability of "Smith (1975)" in the LLM's output reveals something about the nature of scholarly production that academic citation practices obscure.
Calvino speculates about "a writing machine that would bring to the page all those things that we are accustomed to consider as the most jealously guarded attributes of our psychological life" (Calvino 1986: 10). LLMs represent such machines, automating existing cultural processes and dissolving the boundary between human and machine authorship through probabilistic media. What Calvino understood was that the challenge facing us is technological and epistemological. We need to examine how to maintain critical reflexivity when "the more enlightened our houses are, the more their walls ooze ghosts" (Calvino 1986: 25).
What Is an Author?
What Barthes identifies as the historical emergence of the "Author" is helpful for understanding this argument (see also Craig and Kerr 2025; Hayles 2025: 147). He argues that the Author "is a modern figure, produced no doubt by our society insofar as, at the end of the middle ages, with English empiricism, French rationalism and the personal faith of the Reformation, it discovered the prestige of the individual" (Barthes 1977: 142-3). The author thus coincides with the rise of capitalist social relations and the rise of individual ownership and property. He states that "succeeding the Author, the writer no longer contains within himself passions, humors, sentiments, impressions, but that enormous dictionary, from which he derives a writing which can know no end or halt" (Barthes 1977: 146). The LLM is in this sense a dictionary made computational.
Similarly, as mentioned above, Michel Foucault (1979) argues that the author is rather a particular "function" that emerged historically to perform specific work within discourse rather than a naturally occurring thing. He calls it the author-function, and argues that it serves to limit the proliferation of meanings, to classify and group discourses, to establish ownership, and crucially, to distribute responsibility and credit within systems of knowledge production. In effect the author-function is a bureaucratic filter that limits the complexity of an order of discourse.
When academics cite material, they are acknowledging influence but also participating in a complex economy of intellectual credit that determines hiring, promotion, funding, and scholarly status. The citation serves to contain the dissemination of ideas within disciplinary channels, to ensure that knowledge can be attributed to identifiable subjects who can be held accountable and rewarded. Academic citation practices can, therefore, be understood as an apparatus for maintaining and policing these author-functions.
When an LLM generates text, it does so by interpolating across huge numbers of articles, books and documents that have been turned into a model in its training (often ignoring or transgressing academic disciplinary norms). As Earp et al note, "the contribution of any specific text is often untraceable" because "LLM training distributes the influence of its source across billions of interdependent parameters". The provenance problem, they suggest, threatens the academic citation system by making the author-function irrelevant to generative writing processes.
What is key is that this breakdown in attribution may reveal something interesting about textuality itself rather than a failure of the technology or a justification for academic policing. LLMs show in their operation that ideas circulate, combine, and recombine in ways that exceed the capacity of any attribution system to capture. The fantasy that we could maintain complete chains of intellectual attribution, that every idea could be traced back to its "rightful" creator, was always a myth serving particular institutional and economic functions within the university.
Information Loses Its Body
The LLM is a computational instantiation of this authorial reality. When it generates text, it treats all of its training data as informational patterns, stripped of the material and historical contexts that originally gave those texts meaning. "Smith (1975)" (the example given by Earp et al 2025) becomes a pattern distributed across billions of parameters, available for recombination with countless other patterns rather than a specific argument made by a particular person at a particular time.
This process of what I am calling diffusionisation operates through mathematical abstractions that allow LLMs to blend, morph, and generate new textual forms through probability distributions rather than deterministic rules or simple reproduction (this applies to both autoregressive and diffusion forms of generative AI). This marks a shift from discretisation and encoding towards the generation of synthetic variations that might have no original drawn from human writing.[6]
This is what makes the provenance problem intractable. We cannot trace the influence of "Smith (1975)" in LLM-generated text because the text does not contain Smith's ideas in any recoverable form. It contains patterns that, when recombined with other patterns under certain probabilistic constraints, produce outputs that may resemble Smith's arguments.
This author-function is, for Foucault, about control. He writes, "The author is... the ideological figure by which one marks the manner in which we fear the proliferation of meaning" (Foucault 1979: 159). The LLM represents such a proliferation, a generation of meaning that cannot be contained within existing regimes of authorship and attribution. I argue that the Inversion marks the moment when this proliferation becomes systematic and when the default assumption shifts from human authorship to computational generation. We see this already in practice with people proclaiming on social media their rightful ownership of the em-dash or the word "delve", seeking to defend their human originality against the onslaught of generative writing (see Abebe 2025; Beetlespace 2025; Dominus 2025). But as time goes by, these technologies will undoubtably improve till a moment when being AI-generated is considered more "real" than being human-created – then the AIs will, perhaps, really own the em-dash. I would argue, therefore, that the anxiety this produces is political not just ethical as it threatens the mechanisms through which academics and universities distribute resources and recognition.[7]
We might note that academic citation practices extend this logic into the realm of ideas as they function as a kind of intellectual property regime, ensuring that concepts can be owned, that their circulation can be tracked and controlled, and that credit can be accumulated as a form of symbolic or social capital. For Barthes, writing is "a space of many dimensions, in which are wedded and contested various kinds of writing, no one of which is original" (Barthes 1977: 146) rather than necessarily the expression of a singular authorial subject. The text, he argues, "is a tissue of citations, resulting from the thousand sources of culture" rather than a line of words releasing a single meaning emanating from a kind of Author-God (Barthes 1977: 146).
What Barthes described metaphorically as the condition of all textuality, the LLM generates computationally. When we prompt an LLM to help draft a section of scholarly prose, what we receive back is precisely this assemblage, from fragments and patterns extracted from thousands of texts, recombined according to statistical probabilities into something that appears novel and coherent. However, this computational generation is itself uneven, revealing important limits to statistical recombination.
 |
| Figure 1: Comparison of human (blue) vs. AI (red)"jagged" intelligence profiles. Source: https://substack.com/@tomaspueyo/note/c-182052822 |
Karpathy uses the term the "jagged intelligence" to describe LLM capabilities (see Dell’Acqua et al 2026 who introduce the notion of "jagged frontier"), which I think is helpful in showing that these systems are both "a genius polymath and a confused and cognitively challenged grade schooler" (Karpathy 2025). They excel in domains where patterns are densely represented in training data whilst failing at tasks that seem trivial to humans (see figure 1 and 2). This jaggedness is not a temporary limitation to be overcome through better training. It reveals something about how AIs currently operate. The unevenness shows us that aspects of scholarly writing might be reduced to statistical pattern-matching (i.e. far more than we might wish to admit) but that other elements will resist this reduction.[8]
 |
| Figure 2: Comparison of human (blue) vs. AI intelligence (red) profiles, showing jagged capabilities across different domains. (Karpathy 2025) |
Nonetheless, LLMs are disrupting this symbolic economy by making the very notion of discrete intellectual property incoherent. When ideas have been distributed across the billions of parameters that make up the LLM model and the text is generated through statistical recombination rather than individual creation, the question of "who owns this idea?" becomes difficult to answer (although cases are currently making their way through the courts trying to settle this question). This is what causes anxiety for academics and problems ahead for the institutional structures that they depend on.
Flickering Signifiers
We can look at Hayles's concept of "flickering signifiers", which extends Barthes's work, to help to connect these issues to the specific materiality of LLM-generated text. Where traditional literary theory, following Derrida, focused on "floating signifiers" within a dialectic of presence and absence, Hayles argues that digital textuality operates according to a different logic, arguing that "Foregrounding pattern and randomness, information technologies operate within a realm in which the signifier is opened to a rich internal play of difference" (Hayles 1999: 31). Each time we prompt the model, we receive a different configuration of patterns, a different assemblage of a thousand sources of culture. The text, as it were, flickers between states, never settling into the kind of material fixity that traditional citations rely on. The LLM produces what Hayles calls "transient patterns" that "evoke and embody" previous texts without ever being identical to them (Hayles 1999: 47).
This flickering quality is essential to how LLMs operate. The model does not contain "Smith (1975)" as a stable text. Instead, it contains patterns distilled from Smith's work that, when activated under certain conditions, produce outputs that may resemble Smith's arguments. But these patterns are perpetually in flux, recombining with other patterns, producing endless variations.
I argue that the process of diffusionisation operates through this shift from presence/absence to pattern/randomness. Through vector representation and latent space manipulation, diffusionisation dissolves the stable relationships between texts and their origins that grammatisation maintained. Knowledge and cultural production become subject to probabilistic dissolution where Smith's arguments are encoded and reproduced but also transformed into mathematical abstractions that can generate infinite synthetic variations (Berry 2025).
The provenance problem arises because we continue to treat this output as though it should conform to a model of authorship that Barthes declared dead half a century ago and that the Inversion is rendering confused and confusing. We still expect to be able to trace lines of influence, to identify origins, to assign credit to authorial subjects. But the LLM operates according to a different logic entirely.
Hayles's analysis of cognition helps us to understand this point. The posthuman subject, she argues, is characterised by "a distributed cognition located in disparate parts that may be in only tenuous communication with one another" (Hayles 1999: 3-4). When we use an LLM to draft text, we are enacting this kind of distributed cognition. The boundaries between our own thinking and the model's output become porous and unstable. We might not, therefore, be able to distinguish which ideas originated with us and which were suggested by the model, precisely because both are drawing on overlapping bodies of textual patterns.
More recently, Hayles has described distributed cognition in terms of cognitive assemblages which are "collectivities comprising humans, computational media, and electromechanical systems, through which information, interpretations, and meanings circulate" (Hayles 2025: 6). We might say that when a scholar uses an LLM to draft a paper, they participate in a cognitive assemblage where agency is distributed across multiple actors, not all of them human. The text that emerges is created by the assemblage.
The anxiety around LLMs might then be understood as resistance to what Barthes calls "the removal of the Author" which "utterly transforms the modern text" (Barthes 1977: 146). If we accept that the text is always already a tissue of quotations, that complete attribution is impossible, and that authorship is a construct rather than a reality, then the provenance problem vanishes. What remains is a question of recognition that asks how do we reorganise systems of academic credit once the mythology of individual authorship can no longer be sustained and the Inversion has made synthetic generation the default condition?
Barthes offers a radical solution, arguing, that "Once the Author is gone, the claim to 'decipher' a text becomes quite futile" (Barthes 1977: 147). The entire apparatus of academic criticism, which seeks to discover the Author beneath the work, to explain the text through recourse to its creator's intentions, biography, or psychology, becomes obsolete. The LLM produces text that has no Author, a text that cannot be deciphered by reference to an intentional subject. However, I would argue that this opens it to different modes of engagement and does not make such text meaningless. As Barthes writes, "In the multiplicity of writing, everything is to be disentangled, nothing deciphered" (Barthes 1977: 147). The question becomes "what can be done with this text?" rather than "what did the Author mean?"
Hayles terms this moment technosymbiosis, "the deep symbiosis with computational media" (Hayles 2025: 14). She emphasises that "cognitive assemblages operate in contemporary society" in ways that "refuse the assumption that humans are primary in these arrangements" (Hayles 2025: 39). Agency is distributed throughout these assemblages, and the assumption of autonomous human authorship becomes increasingly difficult to maintain. She argues that this transformation is irreversible and "our reliance on cognitive assemblages and computational media has progressed so far that there is no going back. The only feasible options are to go forward from where we are now" (Hayles 2025: 8). Indeed, treating provenance anxiety as a temporary aberration to be corrected through better tools or stricter policies misdiagnoses our situation. We are facing a permanent transformation in how knowledge is produced under conditions of technosymbiosis rather than a momentary disruption of normal scholarly practices. LLMs simply make visible what was always the case, that thought is distributed, that writing is collaborative, that the boundaries between self and other, human and machine, are permeable and constantly renegotiated.
Conclusion
The provenance problem identified in relation to LLM use in scholarly writing is real, but it is not primarily a problem of ethics or attribution. Rather, it represents a crisis in the mythology of academic authorship, a controversy generated by the clash of romantic ideologies of individual creativity with the material realities of how knowledge has always been produced. I have argued above that LLMs reveal and accelerate processes that were always operative and are not necessarily a break in scholarly practices. Indeed, the citational nature of all writing, the construction of authorship, the assemblage-like character of intellectual production, and the ways in which information has lost its body to become patterns available for endless recombination.
The anxiety surrounding LLMs is therefore understandable but misplaced. We are facing the exposure of attribution's limits and functions rather than the corruption of practices of academic attribution. The elaborate system of academic citation has never, in reality, provided complete chains of intellectual genealogy. What the LLMs show, is that these claims can no longer be sustained in their current form. Indeed, we might follow Foucault (1977: 138) in requesting "No longer the tiresome repetitions:"
Consequently, Foucault argues, "New questions will be heard":"Who is the real author?"
"Have we proof of his authenticity and originality?"
"What has he revealed of his most profound self in his language?"
Indeed, he concludes, "behind all these questions we would hear little more than the murmur of indifference":"What are the modes of existence of this discourse?"
"Where does it come from; how is it circulated; who controls it?"
"What placements are determined for possible subjects?"
"Who can fulfil these diverse functions of the subject?"
"What matter who's speaking?"[9]
Barthes shows that we might move from asking "who originated this idea?" to asking "what can be done with this text?", from obsessing over "ownership" of ideas to exploring destinations, from treating texts as stable presences to understanding them as flickering signifiers in continual circulation and recombination. However, these changes raise important questions. The university depends on mechanisms for distributing recognition and resources that presuppose identifiable individual authors. The labour market for academics, the structures of peer review, and the systems of research assessment, are predicated on the author-function that the LLM renders difficult, if not impossible. Moving beyond this calls for a rethinking of how knowledge is valued rather than a set of prescriptive ethical guidelines for LLM disclosure (or even moral censure!)[10] Instead we should ask how intellectual labour is recognised, and how academic careers are built. Rather than a static idea of provenance (e.g. the idea of a world where information had a body and authors had stable identities) we might argue for a move to living provenance, where the processual elements of writing can be explored through a versioning-type system (e.g. Git or a scholarly ledger) which tracks the evolution of a text through its cognitive assemblage. We might not be ready for this, but as Barthes argued, to give writing its future, "the birth of the reader must be ransomed by the death of the Author" (Barthes 1977: 148).
What seems clear is that treating the provenance problem as a technical issue to be solved through better attribution tools or stricter disclosure requirements misunderstands what is at stake. The LLM has made an old problem impossible to ignore. The question is what kind of postdigital scholarship we will build in the world where both have irrevocably flickered into pattern, and where diffusionisation has become the dominant mode of textual production or mediation through probabilistic media (Berry 2025). We will need to think critically about how cultural forms emerge through statistical possibility, acknowledging that each scholarly text produced with or by LLMs exists as one materialisation from a vast field of potential configurations. Provenance anxiety thus reveals itself as a transformation to be understood and engaged with rather than a moral panic.
____
First published 20 December 2025. Revised 22 August 2026 to take account of recent advances in AI detection software and the discussions in the media etc.
____
** Headline image generated using Google Gemini Pro. December 2025. The prompt used was: "Create a high-quality, realistic, and colorful 3D infographic diagram illustrating the "Jagged Frontier" of AI capabilities, featuring a left-to-right progression where "Tasks of a human job" are depicted as irregular, jagged blue polygons to represent human variability, while "Tasks an AI can do" are rendered as vibrant coral pink shapes. The visual narrative should evolve from a small pink form inside a blue polygon (labeled "The AI is a fun toy") to a growing overlapping shape ("The AI is helping me"), culminating in a large, highly irregular, and spiky "jagged frontier" that drastically overlaps the human shape with distinct peaks and valleys ("We are here"), and finally expanding into a massive form nearly swallowing the blue shape ("AGI"), all presented in a modern, professional style with depth, soft drop shadows, and clean typography" Due to the probabilistic way in which these images are generated, future images generated using this prompt are unlikely to be the same as this version.
Notes
[1] An interesting analogy can be made with Wendell Berry's environmental critique, which became influential on contemporary environmentalism. Berry is an American farmer and essayist who argued in The Pleasures of Eating that a kind of industrial amnesia disconnects consumers from the origins of their food. Ignorance of food's provenance thereby makes people complicit in destructive industrial systems (Berry 1990). Berry argues that responsibility presupposes proximity and knowledge of origins. In a globalised food system, "knowing your farmer", however, represents an elite privilege unavailable to those for whom access and affordability are more important than origin. But the parallel with textual provenance anxiety is really fascinating (see also Berry 1990b). Both claims show a kind of "origin-verification" that acts as an ideological apparatus that acknowledges and disavows the industrial systems mediating production. The fantasy of transparent provenance, whether in food or text, doesn't really describe actual chains of production and in fact often serves to distribute responsibility and police boundaries within systems whose complexity makes such complete attribution impossible. Under conditions of industrial and computational mediation, provenance is, perhaps, a fiction.
[2] Ironically, the proposal Earp et al (2025) give for a public declaration of LLM use in the acknowledgements (at the end of their paper) doesn't solve this provenance anxiety (see also Porsdam Mann et al 2024). As they state, "During the drafting of this paper, GPT-5 and Claude Sonnet 4.5 were used to help edit and shorten a longer draft written by the authors. The authors then further edited and refined the text by hand. Such use and this acknowledgement adhere to proposed ethical guidelines for generative AI use and acknowledgement in academic research. Each author has made a substantial contribution to the work, which has been thoroughly vetted for accuracy, and assumes responsibility for the integrity of their contributions". Curiously, they advocate the use of LLMs (or other AI-powered tools like Scite or Elicit) to "check whether AI-generated passages bear substantial similarity to existing scholarship" (Earp et al 2025). Using AI to verify AI demonstrates the possible infinite regression at the heart of the provenance problem as it seems to be LLMs all the way down, a rather counterintuitive way to address their provenance anxiety.
[3] Automation implies the wholesale generation of text by LLMs, whilst augmentation suggests that LLMs are used as tools that enhance human capabilities. Each has implications for authorship, agency, and intellectual labour. More significantly, Earp et al (2025) focus on LLMs for text generation overlooks what may be the most transformative aspect of AI in academia in the use of AI methods for conducting research itself. AI techniques for pattern recognition, large-scale text analysis, data processing, and hypothesis generation represent methodological interventions that could reshape how knowledge is produced, not merely how it is written up. By framing the problem solely in terms of writing and attribution, Earp et al may miss the more profound epistemological shifts occurring when AI becomes key to research methods rather than simply a writing assistant.
[4] To be fair to Earp et al (2025) in the conclusion they do note, "Perhaps we should be moving toward a view of scholarship that is more collaborative and diffuse by default, involving complex assemblages of humans and machines". But this conclusion sits uneasily with their paper's main focus on disclosure and attribution, suggesting a tension in their argument.
[5] A pragmatic response to these issues might be to develop a threshold-based declaration on publication. Rather than requiring disclosure of every computational tool (creating unnecessary writing bureaucracy) or treating all LLM use identically, we might distinguish significant from trivial computational mediation through thresholds. A 30% threshold could mark tools requiring a note (notable augmentation), whilst a 60% threshold would require a declaration (major augmentation). Tools below 30% (word processors' autocorrect, citation managers, brief LLM queries or brainstorming) would be part of standard computational infrastructure requiring no individual declaration. This might be implemented with a nested attribution that also locates individual authors within departmental and institutions, for example:
AUTHOR(S): Sarah Sein, Marcus Weber,
↳ DEPARTMENT: Digital Studies, University of Sussex
↳ COMPUTATIONAL TOOLS:
↳ NOTABLE (>30%):
↳ Python 3.11: NLP analysis;
↳ Gemini Pro: initial brainstorming;
↳ MAJOR (>60%):
↳ N/A;
This helps with making visible the infrastructural conditions enabling scholarship whilst pragmatically preserving the author-function, where needed, for career progression and intellectual accountability. So, for example, a PhD student can continue to claim "original contribution" through linking it to an effective orchestration of a set of tools of inquiry. Their "originality" includes the ability to steer these tools toward a specific, novel research question that a computer could not have formulated on its own. I look to explore this idea further in a forthcoming article.
[6] As Hayles (2025: 148) notes, "In addition, as Rita Raley (2022) has pointed out, the productions of GPT-3 are unrepeatable and hence unverifiable. If the same prompt is repeated for GPT-3, it will generate a response different from the one it did the first time. Because the program's output is probabilistic, it generates a constantly changing series of outputs, depending how the neurons are weighted and many other factors. Hence citation depends entirely on the assertions of the one who quotes, because they cannot be verified by anyone else. The resulting uncertainties destabilize the whole enterprise of literary criticism, which traditionally has treated exact quotation and citation as the sine qua non for acceptable work".
[7] As Foucault writes, "Although, since the eighteenth century, the author has played the role of the regulator of the fictive, a role quite characteristic of our era of industrial and bourgeois society, of individualism and private property, still, given the historical modifications that are taking place, it does not seem necessary that the author-function remain constant in form, complexity, and even in existence. I think that, as our society changes, at the very moment when it is in the process of changing, the author- function will disappear, and in such a manner that fiction and its polysemic texts will once again function according to another mode, but still with a system of constraint—one which will no longer be the author, but which will have to be determined or, perhaps, experienced" (Foucault 1979: 159-160).
[8] We might also see a radical shift in writing practice in a transition from prompt engineering to context engineering. This we might call context writing. Early use of LLMs focused on crafting the perfect prompt by using the right question, framing, and constraints within a single written interaction or request of the LLM. Increasingly, however, sophisticated use involves engineering the context itself. This means providing an LLM with background materials, drafts, stylistic examples, and research or examples that shape its probabilistic space before generation begins. This is a technical development but, I would argue, also a step-change in the writer-LLM relationship. Context engineering reorders the boundary between "writing" and "prompting," creating what Hayles would recognise as a cognitive assemblage operating through continuous environmental manipulation rather than prompts. The author might no longer simply ask the LLM to write but constructs an informational milieu within which writing takes place. This represents a further challenge to the author-function, and moves the question from "who wrote this?" to "who designed the context from which this emerged?". Under the conditions of the Inversion, I think it is highly likely that this form of contextual authorship may become the primary mode of writing. Its implications on scholarly writing will be equally profound. Writing may increasingly shift toward the idiosyncratic, the counter-intuitive, and what we might call a "constellational" style, that is, writing that resists normalisation. This might be through anti-probabilistic writing, intentional stylistic, or argumentative choices designed to be unlearnable by LLMs. Perhaps representing a shift from context engineering to signal engineering, where the writer isn't just managing informational context but producing non-statistical noise that generates new concepts and ideas. Under these conditions, the jagged frontier becomes a space to inhabit strategically. This turns the jagged frontier from a threat into a map.
[9] Ironically, Foucault's own text, "What is an Author?" has an unstable provenance, appears to exist in at least two authorial versions and three translations (the canonical form [Foucault 1979] circulates under a description of its origin that is not wholly accurate). As Elden notes, "the 1969 Paris lecture was published in the Société’s Bulletin and was partly translated in Language, Counter-Memory, Practice in 1977, while the Buffalo lecture first appeared in Textual Strategies and is reprinted in The Foucault Reader, Essential Works volume 2 and The Essential Foucault. Textual Strategies was edited by Josué V. Harari, who translated Foucault’s lecture for this collection. In Dits et écrits there is a composite text (#69) which compares the 1969 French publication of the Paris lecture to the 1979 publication of the Buffalo lecture. (The Buffalo version is given text number #258, but not printed separately.) While there are other reprints, and an earlier translation of the Paris text by James Venit, there is no English translation of the Dits et écrits composite text, and only a partial translation of the Paris discussion" (Elden 2021). He concludes, "In summary, I think the Textual Strategies version is not really the Buffalo text. I certainly think it’s not a translation of a faithful transcription of a recording. In the “Preface” to Textual Strategies, 13, Harari indicates it is a “revised version” of the Buffalo text, and that Foucault gave him “a free hand to edit” the text “with an American readership in mind”. To achieve this, Harari translated the Paris text, and edited it to better match things Foucault said in Buffalo – cutting the Introduction, adding the final discussion of ideology. For the latter work did he have access to a recording, and might that still exist somewhere? [update: yes – the recording has been found and is being edited for publication.]" (Elden 2021). It seems that the scholarship on the instability of Foucault's text is itself unstable.
[10] I wonder if we should call attacks on people who use AI or rejecting AI material as an AI hominem? Perhaps we could describe it as: AI hominem: 1. rejecting a claim on the grounds of its AI origin rather than its content, the genetic fallacy, what we might call origin-based AI hominem - AI hominem ex origine. 2. Attributing a personality to the machine in order to evaluate its output, an anthropomorphism or personification-based AI hominem - AI hominem ex persona.
Bibliography
Abebe, N. (2025) ‘With the Em Dash, A.I. Embraces a Fading Tradition’, The New York Times, 18 September. Available at: https://www.nytimes.com/2025/09/18/magazine/chatgpt-dash-hyphen-writing-communication.html
Barthes, R. (1977) The Death of the Author, in Image-Music-Text. Essays selected and translated by S. Heath. London: Fontana, pp. 142–148.
Beetlespace (2025) A Proclamation Regarding the Restoration of the Dash, Beetlespace. Available at: https://blog.nawaz.org/posts/2025/Dec/a-proclamation-regarding-the-restoration-of-the-dash/.
Benjamin, W. (2008) The Work of Art in the Age of Mechanical Reproduction. 1st edition. Penguin.
Berry, W. (1990) ‘The Pleasures of Eating’, in What Are People For? San Francisco: North Point Press.
Berry, W. (1990b) ‘Why I Am not Going To Buy A Computer’, in What Are People For? San Francisco: North Point Press.
Berry, D. M. (2014) Critical Theory and the Digital, Bloomsbury.
Berry, D.M. (2025) Synthetic Media and Computational Capitalism: Towards a Critical Theory of Artificial Intelligence, AI & Society, 40, pp. 5257-5269.
Calvino, I. (1986) Cybernetics and Ghosts, in The Uses of Literature. Harcourt Brace, pp. 3-27.
Craig, C. and Kerr, I. (2025) ‘The death of the AI author’, in Robot Law: Volume II. Edward Elgar Publishing, pp. 250–285. Available at: https://www.elgaronline.com/edcollchap/book/9781800887305/chapter9.xml
deBoer, F. (2026) ‘I Wouldn’t Say Pangram is Broken, But I Would Say That It’s Brittle’, Freddie deBoer, 19 July. Available at: https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-is-broken-but (Accessed: 22 August 2026).
Dell’Acqua, Fabrizio, Edward McFowland, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani. (2026) ‘Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality’, Organization Science, 37(2), pp. 403–423. Available at: https://doi.org/10.1287/orsc.2025.21838.
Dominus, M. (2025) ‘People have been saying lately that em-dashes are a sign that text has been generated by an LLM.I would just like to mention that although…’, Mastodon. Available at: https://mathstodon.xyz/@mjd/114730157688607856
Earp, B.D. Yuan, H., Koplin, J. and Porsdam Mann, S. (2025) ‘LLM use in scholarly writing poses a provenance problem’, Nature Machine Intelligence, 7(12), pp. 1889-1890. https://doi.org/10.1038/s42256-025-01159-8.
Foucault, M. (1977) What is an Author?, in Language, Counter-Memory, Practice: Selected Essays and Interviews. Edited by D.F. Bouchard. Ithaca, Cornell University Press, p. 113–138.
Foucault, M. (1979) What is an Author?, in Harari, J. (ed.) Textual Strategies: Perspectives in Post-Structuralist Criticism, Cornell University Press, 1979, pp. 141-160.
Hayles, N.K. (1999) How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. University of Chicago Press.
Hayles, N.K. (2025) Bacteria to AI: Human Futures with Our Nonhuman Symbionts. University of Chicago Press.
Jancer, M. (2025) ‘You’re Not Imagining It. People Actually Are Starting To Talk Like ChatGPT.’, VICE, 29 June. Available at: https://www.vice.com/en/article/youre-not-imagining-it-people-actually-are-starting-to-talk-like-chatgpt/ (Accessed: 22 August 2026).
Karpathy, A. (2025) Year in Review 2025, bearblog, https://karpathy.bearblog.dev/year-in-review-2025/
Kirshner, A. (2026) ‘A.I. Detectors Are Supposed to Make Our Writing Better. It Might Be Doing the Opposite.’, Slate, 18 August. Available at: https://slate.com/technology/2026/08/ai-detector-pangram-writing-process.html.
Klee, M. (2026) ‘The Pope’s Warnings About AI Were AI-Generated, a Detection Tool Claims’, Wired, 22 April. Available at: https://www.wired.com/story/pope-tweets-ai-generated-pangram-chrome-extension/ (Accessed: 22 August 2026).
Porsdam Mann, S., Vazirani, A.A., Aboy, M., Earp, B.D., Minssen, T., Cohen, I.G., Savulescu, J., (2024) Guidelines for ethical use and acknowledgement of large language models in academic writing. Nat Mach Intell 6, 1272–1274. https://doi.org/10.1038/s42256-024-00922-7
Comments
Post a Comment