Assisted authorship from antiquity to generative AI

Saint Paul Writing His Epistles, oil on canvas, c. 1618–20, attributed to French artist Valentin de Boulogne.

By Rod Benson

The advent and rapid embrace of generative artificial intelligence has revived questions about literary assistance, intellectual originality and the meaning of authorship. This article has its origins in my thoughts on the common practice of professional ghostwriting and the apparent use of an amanuensis by St Paul in the New Testament. It examines the writing practice of ancient secretaries, Athenian speechwriters, imperial speechmaking, Homeric formulaic composition, rhetorical imitation, biblical correspondence and the late-antique cento. I use these examples to challenge the equation of authorship with the unaided production of texts, but I acknowledge that this does not by itself establish a functional equivalence between human collaborators and generative systems.

I then situate the classical evidence in conversation with Edwin Hutchins’s model of distributed cognition and Andy Clark and David Chalmers’s extended-mind hypothesis, while addressing objections to the analogy and to the extension of cognition beyond individuals. I explore the key differences between credit for a contribution, the entitlement to claim authorship, and accountability for the finished work, using the International Committee of Medical Journal Editors’ guidance as an illustration. Finally, I suggest that ancient ghostwriting offers not a simple endorsement of AI-assisted writing but a diagnostic question: at what point does assistance replace the intellectual achievement that attribution to an author reasonably implies? A context-sensitive test of provenance, understanding, critical control, integrity and disclosure offers a more defensible answer than either prohibiting assistance or equating approval of a finished text with genuine authorship.

Generative AI can help to formulate questions, compose prose, identify objections, translate, edit and summarise. Its versatility makes the familiar distinction between author and instrument difficult to apply. Yet anxiety about ‘machine authorship’ often relies upon an implausibly solitary picture of literary production. Greek and Roman works were created within practices of dictation, performance, imitation, editorial assistance and delegated speechwriting. Such histories establish that writing need not be physically or linguistically unaided to count as someone’s work. They do not show, however, that all delegations of composition are legitimate, or that a language model has the capacities, intentions or obligations of a human collaborator. The historical cases are analogies for evaluating different activities, not technological ancestors of generative AI.

An adequate argument must separate three questions often hidden under the term ‘authorship’. One concerns causal contribution: who or what helped produce this text? A second concerns credit: whose intellectual achievement does publication under a particular name represent? A third concerns responsibility: who must answer for the accuracy, sourcing, implications and integrity of what is published? These answers may diverge. A speechwriter can compose a speech delivered under another person’s authority; an editor can substantially improve prose without originating its claims; and a researcher can approve a generated argument without understanding its evidence. The present essay argues that assistance becomes ethically objectionable substitution when the central intellectual activity that a claim of authorship implies has in fact been delegated and is not adequately recovered through critical, informed human engagement. Because those expectations differ by genre, the boundary is principled rather than mechanically measurable.

Roman authors frequently employed secretaries to record dictation, prepare copies and manage correspondence. Cicero’s associate Marcus Tullius Tiro, enslaved before his manumission, became indispensable to Cicero’s literary and administrative work. Later tradition associated him with shorthand, but neither the full history of that system nor Tiro’s precise involvement in particular Ciceronian compositions can be confidently reconstructed. Pliny the Younger provides a clearer example of a limited function: in Letters 9.36 he describes composing and revising thoughts before summoning a shorthand writer. The recorded words could thus express intellectual work largely undertaken before inscription. These sources make it unreasonable to equate physical transcription with the origination of an argument, while also warning against assuming that every ancient secretary was merely a mechanical recorder (Cicero, Letters; Pliny, Ep. 9.36; Richards 1991).

Modern dictation and routine language correction are comparable in function, though not identical in mechanism, to such secretarial activities. Yet the comparison loses force if an AI system supplies the thesis, premises and interpretation rather than helping to record or clarify them. Nor should ancient secretarial labour be idealised: enslaved workers often lacked control over the recognition accorded to their expertise. The history of literary assistance is also a history of unequal power and obscured contribution. That lesson complicates contemporary claims that AI writing is a frictionless achievement of one user, since modern systems likewise depend on distributed labour, infrastructure and prior human expression, although the forms of dependence are historically and morally distinct.

Classical Athenian litigants frequently delivered their own courtroom speeches but could also engage professional speechwriters (logographoi). Lysias is the best-known example: speeches attributed to him were crafted for the first-person delivery of litigants. The extent of consultation and control exercised by clients is disputed; Kenneth Dover’s interpretation of the speechwriter as consultant has itself been contested. It would be inaccurate to claim that the speeches were wholly authored by Lysias independently of clients. Composition, performance, factual testimony and legal responsibility could belong to different participants (Dover 1968; Todd 1993; Dickinson College Commentaries, ‘Lysias’ Life and Works’).

Tacitus provides a different illustration in Annals 13.3: the funeral address delivered by Nero for Claudius was, according to Tacitus, composed by Seneca. Nero’s office and delivery gave the address its public authority; Seneca furnished its literary form. Yet the fact that Tacitus identifies the composer makes this an example of delegated composition, not proof of concealed deception. Both cases demonstrate that authorship conventions vary. A lawyer-assisted or speechwriter-assisted public address may constitute an authorised communication even though the speaker did not personally frame every sentence. A scholarly article submitted as evidence of independent research makes a different claim: that the named scholar performed or responsibly directed the inquiry and can justify the argument.

This distinction also answers an obvious objection: if human ghostwriting has long been accepted, why is the analogous use of generative AI assistance so widely criticised? The answer is that ghostwriting has never possessed a single ethical status. Its legitimacy varies with expectations attached to the text constructed by the ghostwriter. An authorised ceremonial speech, a signed memoir, a student’s examination and a peer-reviewed research paper make different implicit representations. The crucial problem is not borrowed wording but misleading attribution of the relevant achievement. Conversely, the use of AI to generate a substantial preliminary draft does not automatically extinguish authorship if the scholar subsequently conducts the actual research, independently checks the evidence and reconstructs a defensible argument. Merely reading and approving plausible paragraphs, however, is not the same thing as performing that creative work.

The Homeric epics employ recurring formulas and inherited narrative resources. Parry and Lord demonstrate the importance of formulaic composition in oral epic traditions, although the relationship between those traditions and the surviving written epics remains debated (Parry 1930; Lord 1960). The comparison with language models should remain modest: human performers belong to socially sustained traditions and make situated artistic decisions, whereas machine generation is a different process. However, an original work need not consist of expressions without antecedents. An artist may display originality by arranging, interpreting and transforming inherited material.

Roman rhetorical instruction made this point explicitly. Quintilian regarded imitation as indispensable but warned that ‘imitation alone is not sufficient’ (Institutio oratoria 10.2.4). Virgil’s Aeneid reworks Homeric patterns for a recognisably different historical and moral project. A still sharper example is the fourth-century poet usually identified as Faltonia Betitia Proba, whose Cento Vergilianus de laudibus Christi reconstructs biblical narratives from fragments of Virgil. The attribution of this cento has been debated at the margins, though the conventional identification remains predominant; unlike the typical outputs of generative AI, its borrowed lines are intentionally traceable. Its originality lies in selection, juxtaposition and Christian reinterpretation rather than linguistic independence (McGill 2005; Cambridge History of Early Christian Literature; Classical Quarterly, 2024).

One objection follows immediately: is not the claim that AI merely ‘recombines’ prior texts technically misleading? Yes, if taken as a description of all model outputs. Modern language models are not ordinarily performing a cento-like cut-and-paste operation, nor can the history of any output normally be reduced to identifiable source fragments. The defensible analogy concerns the evaluative question of transformation, not an asserted identity of mechanisms. A text may be verbally novel yet analytically derivative, or linguistically dependent yet intellectually original. Scholarship must determine what interpretive work was performed, and by whom.

In the New Testament, Romans 16:22 contains a greeting from Tertius, who identifies himself as the writer of the letter. The statement warrants identifying a second participant in the text’s preparation but does not by itself reveal whether Tertius took verbatim dictation, shaped individual expressions or exercised broader compositional discretion. The surviving evidence does not justify treating his contribution as either negligible or demonstrably coequal with Paul’s theological authorship (Romans 16:22; Richards 1991, 2004). Similarly, an AI-use acknowledgment identifies a contribution to production but does not by itself establish that the named human author performed the intellectual work claimed. Transparency and substantive authorship are related requirements, not substitutes for each other.

In Cognition in the Wild (1995), Edwin Hutchins studied ship navigation as an activity accomplished through coordinated people, instruments, procedures and representations. The relevant cognitive system was not adequately described by examining a navigator in isolation. Classical literary production can likewise be understood as embedded in networks of memory, notation, libraries, assistants and rhetorical conventions. Contemporary scholars use databases, colleagues, search tools and software in processes that may reorganise how questions are posed and conclusions reached. The claim that cognition is distributed, however, describes how tasks are performed; it does not allocate individual credit or transfer moral responsibility.

Andy Clark and David Chalmers advance a stronger possibility in ‘The Extended Mind’ (1998). Their comparison of Otto’s notebook with Inga’s biological memory challenges the assumption that a cognitive process must occur within the skull. Under sufficiently integrated conditions, an external resource might constitute rather than merely aid cognition. Clark’s later Supersizing the Mind (2008) develops related approaches to embodied and technologically supported thought. These accounts help explain how writing itself can be a mode of thinking and how AI responses may provoke genuine intellectual insight. But it does not follow that any convenient chatbot is part of its user’s mind, or that novel outputs are the user’s achievement simply because they appear within a continuing interaction (Clark and Chalmers 1998; Clark 2008).

Critics including Fred Adams and Kenneth Aizawa question whether causal coupling is being confused with cognitive constitution. They argue that dependence on a tool, however intimate, does not alone establish ownership of all the operations it performs. Indeed, a scholar’s cognitive activity may benefit from a system whose outputs the scholar does not understand. The appropriate conclusion is modest but important: intellectual agency can be exercised through external resources, yet neither the distribution nor the possible extension of cognition automatically validates a claim of individual authorship (Adams and Aizawa 2008). To say that an AI tool contributes to the process of inquiry is therefore different from saying that it deserves a human byline, or that its user has thereby demonstrated the competence attributed to the resulting publication.

In Phaedrus 274c–275b, Socrates tells the story of Theuth’s invention of writing and King Thamus’s objection that written signs may encourage reliance on reminders while producing an appearance of wisdom. Plato is not issuing a straightforward modern ban on writing; the dialogue investigates the relation between discourse, understanding and responsive philosophical inquiry. Its enduring warning is that possession of eloquent expressions does not guarantee mastery of the reasons behind them. Conversational AI differs from an inert written page because it can respond to questions, yet those responses can themselves be incorrect, unsupported or fabricated. The practical test is whether AI use strengthens the researcher’s capacity to evaluate claims or permits convincing language to stand in for knowledge (Plato, Phaedrus 274c–277a).

A legitimate objection is that authors routinely rely on collaborators whose specialised expertise they cannot fully reproduce. This is true, and an untenable account of authorship would require each co-author to know everything. Responsible collective research instead permits differentiated expertise alongside explicit allocation of tasks, sufficient scrutiny, reliable methods and a willingness to investigate problems. The aim is not omniscience; it is informed accountability proportionate to one’s role, accompanied by honest presentation of others’ contributions. Solo authors who use AI cannot appeal to an absent collaborator as a person able to assume responsibility for generated claims.

A sufficiently original scholarly contribution concerns the framing of a worthwhile question, the selection and justification of methods, the critical examination of sources, the interpretation of evidence and the development of conclusions capable of surviving informed challenge. Originality may consist in discovering evidence, correcting an established account, identifying a neglected implication or producing a genuinely illuminating synthesis. It is not adequately measured by the novelty of phrasing or by the percentage of prose written without assistance. A scholar can independently discover and defend an interpretation while relying on extensive language editing; another can type every word of an unoriginal argument. The central issue is whether the work makes a substantive, defensible contribution and whether the named author is entitled to the intellectual credit claimed.

This distinction has special importance in research that depends on historical texts. Tacitus’s report supports a claim about Seneca composing Nero’s funeral address; it does not prove that Roman public speechmaking was always secretly ghostwritten. In the New Testament, Romans 16:22 identifies Tertius as the writer but does not settle what discretion he exercised in performing the task. Similarly, Proba’s cento supports the possibility of original recomposition but not the claim that every machine-generated recombination is creative. Such boundaries between evidence and inference are among the actual intellectual tasks scholarly authorship requires. Language models can help identify possibilities; the author must establish which possibilities are justified.

A precise modern comparison is supplied by the International Committee of Medical Journal Editors (ICMJE). Its recommendations, updated in January 2026, distinguish credit for contributing to a paper from qualification for authorship. The ICMJE proposes four criteria, all of which a person named as an author should meet: a substantial contribution to the conception or design of the study, or to obtaining, analysing or interpreting its data; participation in drafting the work or critically revising its important intellectual content; approval of the final version; and agreement to be accountable for the integrity of the work, including the proper investigation and resolution of concerns. Its guidance states that writing assistance, technical editing, language editing and proofreading do not, on their own, qualify a contributor for authorship. These are standards developed for medical-journal publication—not universal rules for every discipline, genre or literary tradition—and the first criterion is appropriately adapted, not copied literally, when discussing historical or literary research (ICMJE, ‘Defining the Role of Authors and Contributors’).

The ICMJE further advises that authors disclose whether and how AI-assisted technologies were used in producing a submitted manuscript, with writing assistance acknowledged and relevant data or figure-generation uses described in the methods. It says chatbots should not be listed as authors because they cannot assume the responsibilities required of authorship. Human authors must review generated content, ensure proper attribution and remain accountable for accuracy, integrity and originality. These rules do not imply that any human who merely approves an AI-produced manuscript thereby qualifies as an author. Final approval is only one of the ICMJE’s four cumulative conditions. Nor does disclosure automatically legitimise prohibited delegation or inadequate research. The Committee on Publication Ethics likewise treats transparency and human responsibility as essential, while specific journal policies may impose further limits (ICMJE, ‘Use of AI by Authors’; COPE, ‘Authorship and AI Tools’).

An important criticism is that medical publication criteria cannot settle authorship in the humanities. This is correct. Humanities work includes single-author interpretation, critical editions, translations, oral history and collaborative research with different conventions of attribution. The ICMJE is used here as an instructive contrast rather than as an authority with universal jurisdiction. Its broader insight is that contribution, credit, approval and accountability must not be conflated. In humanities scholarship, a researcher must similarly make a substantial interpretive contribution and be capable of defending its evidence, even though the precise institutional rules must come from the relevant discipline and publisher.

Five related points warrant reflection. First, a clear assistance–substitution threshold might seem impossible because ideas emerge through iterative conversation. A scholar may ask an AI system for three interpretations, reject two, develop the third and subsequently discover new archival evidence. Is the resulting argument the machine’s or the scholar’s? The answer cannot be obtained by counting prompts or machine-generated sentences. The decisive questions are whether the scholar supplied or established the evidential basis, independently tested the interpretation, made the governing methodological choices and can explain why the final argument is warranted. Critically developing an externally suggested idea can constitute original research; unexamined adoption of an apparently persuasive idea ordinarily cannot.

Second, a criterion of independent understanding may disadvantage scholars using translation, accessibility aids, dictation or other forms of assistance. This concern is serious. No defensible policy should confuse disability, linguistic background, typing ability or access to conventional writing technologies with intellectual competence. The proposed distinction is functional and context-sensitive: it concerns the research judgement that a publication claims, not the unaided performance of every mechanical or stylistic task. Accessible and collaborative methods may allow scholars to exercise their capacities more fully. Ethical evaluation should examine the actual intellectual contribution and the relevant standards, not enforce an ideal of unassisted performance.

Third, one might argue that modern scholarship has always involved invisible editors, research assistants and collective infrastructure, so singling out AI is inconsistent. The history reviewed here supports the premise but not the conclusion. Existing forms of assistance also generate questions of recognition, exploitation and ghost authorship. AI makes those questions more acute because one system can move rapidly from editing prose to constructing plausible arguments and references, while offering no human participant who can take professional responsibility for its unsupported assertions. Consistency requires examining traditional ghostwriting as critically as AI delegation, not exempting AI because earlier arrangements were imperfect.

Fourth, the proposed emphasis on human responsibility may ignore harms occurring before the final manuscript is written. Training-data provenance, copyright, confidential source material, culturally biased outputs, environmental costs and labour conditions raise independent ethical concerns. A scholar could exercise strong interpretive control over an AI-assisted article while still using a system in ways that breach confidentiality or contractual obligations. Conversely, a text produced entirely without AI may exhibit plagiarism, fabricated evidence or exploitative ghostwriting. Responsible authorship is therefore a necessary but not sufficient condition of responsible research. Technology governance, fair acknowledgement of human labour and compliance with institutional and legal requirements remain separate obligations; specific legal entitlements should not be inferred from ancient literary practice.

Finally, the classical analogies might be accused of anachronism. Ancient categories such as logography, imitation and scribal assistance did not anticipate machine learning or modern scientific publication. This objection is valid against any claim of historical identity. The examples are useful precisely as controlled comparisons showing that textual production, intellectual achievement and public attribution can diverge. They cannot determine AI’s cognitive status or contemporary publishing rules. Historical comparison identifies a durable problem; philosophical analysis and present-day standards must supply the normative response.

The most useful practical test asks five connected questions. First, there is the issue of provenance: which substantial elements of the research question, evidence, analysis and argument came from the author, other humans or AI? Second, control: who made the decisive interpretive and methodological choices, and could the author explain why alternatives were rejected? Third, understanding: can the author defend the claims using checked primary or trustworthy secondary evidence rather than an AI-generated account of its own reasoning? Fourth, integrity: have citations, quotations, data, confidentiality and the rights of others been appropriately handled? Fifth, disclosure: is the extent and function of assistance reported according to the applicable publisher, institutional and disciplinary expectations? These are diagnostic considerations rather than a universal numerical threshold or a claim that all writing genres demand the same independence.

Consider three cases. In the first, a researcher identifies a neglected classical passage, verifies its context, develops an interpretation and uses AI to simplify prose. The central scholarly achievement remains recognisably that of the researcher. In the second, AI proposes a possible reading, but the researcher subsequently undertakes independent source criticism, rejects unsupported claims and develops a defensible new thesis. The initial suggestion need not negate the eventual original contribution, although disclosure may be required. In the third, AI invents an argument and citations; the named author lightly revises the prose and publishes it without checking the sources. Here the appearance of scholarship substitutes for the relevant intellectual achievement. Intermediate cases require judgment and the rules of the intended publication, but the contrast reveals why neither word counts nor final approval are adequate tests.

Ancient amanuenses demonstrate that writing words and originating an argument are different activities. Athenian logographers and Seneca’s imperial speechmaking show that public voice can be supplied by someone other than the speaker. Homeric composition, rhetorical imitation and the cento reveal that creative originality may arise from the purposeful transformation of inherited material. Hutchins, Clark and Chalmers help explain why thinking can be socially and materially supported, while their critics caution against equating every external causal contribution with cognition itself. Plato supplies the further warning that impressive language is not evidence of knowledge. Together, these arguments undermine the myth of wholly self-sufficient authorship without authorising indiscriminate delegation.

The ethical contribution of ancient ghostwriting is thus not the reassurance that people have always used assistants. It is the distinction between delivering a text, endorsing its message and performing the intellectual work for which the named author seeks credit. In academic scholarship that work includes forming worthwhile questions, selecting defensible methods, interpreting evidence and justifying conclusions.

Generative AI can genuinely assist these activities, and at times improve them, but it may also produce a polished substitute for their performance. The appropriate standard is consequently neither an absolutist prohibition on assistance nor the assumption that human approval alone establishes authorship. Authorship remains warranted when a person makes the substantive contribution claimed, exercises informed critical control, reports material assistance honestly and accepts responsibility for the work. The point at which assistance becomes substitution is the point at which these conditions cease to describe the actual process of production.


Rev Dr Rod Benson lives in Sydney and writes about a range of topics from books and local history to food, faith, philosophy, ethics, research skills, and everyday life. Away from the keyboard, he enjoys cooking for friends, walking in the Australian bush, exploring a beach, discovering new restaurants, and reading philosophy, fiction and poetry. He is married to Emma and is minister of North Rocks Community Church, a Uniting Church congregation.

References

Primary classical texts

Cicero. Letters to Atticus and Letters to Friends.

Plato. Phaedrus, 274c–277a.

Pliny the Younger. Letters, 9.36.

Proba. Cento Vergilianus de laudibus Christi.

Quintilian. Institutio oratoria, 10.2.

Tacitus. Annals 13.3.

The New Testament.

Virgil. Aeneid.

Selected scholarship

Adams, Fred, and Kenneth Aizawa. The Bounds of Cognition. Wiley-Blackwell, 2008.

Clark, Andy. Supersizing the Mind: Embodiment, Action, and Cognitive Extension. Oxford University Press, 2008.

Clark, Andy, and David Chalmers. “The Extended Mind.” Analysis 58 (1), 1998, 7–19. https://doi.org/10.1093/analys/58.1.7.

Dover, K. J. Lysias and the Corpus Lysiacum. University of California Press, 1968.

Hutchins, Edwin. Cognition in the Wild. MIT Press, 1995.

Lord, Albert B. The Singer of Tales. Harvard University Press, 1960.

McGill, Scott. Virgil Recomposed: The Mythological and Secular Centos in Antiquity. Oxford University Press, 2005.

Parry, Milman. “Studies in the Epic Technique of Oral Verse-Making. I: Homer and Homeric Style.” Harvard Studies in Classical Philology 41 (1930), 73–147.

Richards, E. Randolph. The Secretary in the Letters of Paul. Mohr Siebeck, 1991.

Richards, E. Randolph. Paul and First-Century Letter Writing: Secretaries, Composition and Collection. InterVarsity Press, 2004.

Todd, S. C. The Shape of Athenian Law. Clarendon Press, 1993.

Verified institutional and online references

Committee on Publication Ethics (COPE). “Authorship and AI Tools.” https://doi.org/10.24318/cCVRZBms.

Dickinson College Commentaries. “Lysias’ Life and Works.” https://dcc.dickinson.edu/lysias-24/intro/lysias-life-works.

Harvey, Susan Ashbrook Harvey, “Women and words: texts by and about women,” in Frances Young et al (eds), The Cambridge History of Early Christian Literature. https://www.cambridge.org/core/books/abs/cambridge-history-of-early-christian-literature/women-and-words-texts-by-and-about-women/79FDB379B819BB6BA5D982FD12AC87B1.

ICMJE. “Defining the Role of Authors and Contributors” (January 2026). https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html.

ICMJE. “Use of AI by Authors” (January 2026). https://www.icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html.

MIT Press. Hutchins, Cognition in the Wild. https://mitpress.mit.edu/9780262581462/cognition-in-the-wild/.

Quintilian, Institutio oratoria X.2 (Loeb translation). https://penelope.uchicago.edu/Thayer/E/Roman/Texts/Quintilian/Institutio_Oratoria/10B%2A.html.

Tacitus, Annals XIII (Loeb translation). https://penelope.uchicago.edu/Thayer/E/Roman/Texts/Tacitus/Annals/13A%2A.html.

Image source: Psephizo

Leave a Reply