Zum Symposium

Too Polished to Be Human?

Orwell and the Suspicion of AI in Academic Writing

15.11.2026

In the opening chapter of 1984, Winston Smith adjusts his expression before the telescreen. He cannot know exactly when he is being watched, so he must behave as though observation is continuous. That anticipation is what makes Orwell useful for thinking about academic responses to generative AI. 

Academic evaluation has purposes and safeguards quite different from those of an Orwellian state. Still, the comparison concerns the adjustment of expression in anticipation of scrutiny. For scholarship, the question is what happens when the prospect of a suspicious reader begins to influence how an argument is written.

This post argues that a culture of suspicion can affect scholarship before any accusation is made, encouraging authors to organise their writing around the need to appear untouched by technological assistance. It first distinguishes this suspicion from established academic scrutiny, then examines the standard of human writing it presupposes. It concludes by asking what institutions must reconsider about scholarly labour and assessment if integrity is to mean more than a declaration of compliance.

Writing for an Imagined Accuser

Consider an author revising a paragraph that accurately states the argument and reflects the sources consulted. On rereading it, the author worries a phrase looks too polished or the transitions too regular. The author changes both to avoid the possibility that they might be taken as signs of AI. Removing formulaic language can, of course, improve a text. Here, however, the reason for revision has shifted. The author is responding to an anticipated allegation, and a sentence that already expressed the point adequately must now reassure someone, an imagined accuser, about how it was produced.

Where that expectation takes hold, the imagined accuser becomes an audience within the writing process. The writer has to anticipate which features will be recognised as human, even where no one has clearly specified what those features are. A requirement to disclose a particular use of AI can be understood and followed. An expectation that a paragraph should simply feel human leaves the writer trying to satisfy an ever-moving benchmark derived from certain intuitions.

Such expectations may be exclusionary without imputing wrongdoing. Suspicion arises when features of a particular text become grounds for doubting how it was produced or the contribution of its author. Jean d’Aspremont’s intervention questions this move from stylistic impressions to judgements about scholarship. How, then, anticipating that move influence writing before any allegation is made?

Academic scrutiny already exposes authors to rejection and reputational damage. Challenges to evidence or interpretation concern the merits of a contribution; allegations of plagiarism, ghostwriting or undisclosed assistance concern its production. AI suspicion extends that history. It introduces a question about degrees of technological mediation: a model may revise a sentence or help organise an argument, and stylistic impressions do not establish which occurred. An objection to interpretation challenges expertise; an allegation of concealed delegation can question scholarly honesty. This additional scrutiny may arise even where the assistance was permitted.

The Unequal Burden of Looking Human

In international legal scholarship, English is a shared language for researchers with very different linguistic backgrounds. A phrase familiar to one academic tradition may sound awkward in another. Repetition sometimes reflects uncertainty; it may also preserve a distinction between legal concepts or reflect a particular literary quality. Before treating a stylistic feature as suspicious, a reader needs to consider what it does in the argument.

Research on detection tools illustrates the difficulty. In a study published in 2023, Weixin Liang and colleagues tested seven detectors using human-written TOEFL essays and essays by US school students. The detectors misclassified many TOEFL essays, and changing vocabulary also changed their classifications. These results concern particular tools and samples from 2023; they cannot establish the performance of every current detector or the frequency of false accusations in legal publishing. They do show how judgements based on linguistic features can diverge from the history of a text’s production.

The study also found that AI-assisted vocabulary revision could reduce the misclassification of human-written essays. Technological intervention could, in that setting, make a text more readily classified as human.

These experiments concern automated classifications. Extending the concern to human evaluation requires a further step: If an institution treats impressions of linguistic form as grounds for suspicion, authors may begin anticipating those impressions when they write. That response is the hypothesis examined here; the study does not establish its occurrence or prevalence.

Whose writing, then, supplies the implicit standard of authenticity?  Editors and reviewers give particular conventions practical force through publication decisions, drawing on traditions transmitted through training and supervision. Detector developers add classifications of linguistic form. Odile Ammann’s analysis of language bias examines how gatekeeping in international legal scholarship privileges particular linguistic and academic traditions. When these preferences become evidence of human authorship, a standard of acceptable prose acquires a further disciplinary function.

Generative AI may also disturb this hierarchy. Linguistic assistance can make a valued register more accessible while encouraging conformity to it. Liang’s finding that AI-assisted revision reduced misclassification makes the tension concrete: technological assistance can help satisfy the standard against which human authorship is assessed. Its potential to widen participation therefore coexists with the risk of reinforcing the conventions that make such assistance necessary. Institutions must decide whether greater access to fluent prose prompts broader recognition of scholarly contributions or further demands for proof of individual authorship.

When Suspicion of AI Enters the Argument

Orwell’s account of habitual self-control helps explain why the effects might reach beyond vocabulary. Winston monitors his expression because someone may read it as a sign of disloyalty. A scholar anticipating suspicion faces a different concern: that a sentence might be treated as evidence that its author did not think it.

The call for this discussion asks whether what d’Aspremont calls “inscription” can be separated from “thought”. Writing helps determine the scope of a claim. His description of legal scholarship as a coded discourse helps locate the difficulty. Legal terms carry inherited argumentative commitments: distinguishing an obligation from a recommendation involves more than changing expression. A qualification may emerge only when an author attempts to formulate a proposition precisely. Rearranging a paragraph can expose a missing step in the reasoning. These conventions make legal argument intelligible, even as their regularity makes them available to machine reproduction. Abandoning a useful formulation because it looks suspicious may therefore alter the reasoning that the formulation helps sustain.

Consider, hypothetically, an author who repeats a treaty term throughout a passage because its meaning is under examination. Concerned that the repetition looks machine-generated, the author replaces some occurrences with near-synonyms. The passage gains variety, but it becomes harder to tell whether each expression refers to the same legal concept. The revision has weakened the consistency needed to assess the argument. Anticipating that their craftsmanship may be called into question, legal craftspeople may feel compelled to abandon their craft. 

AI use, compliance with editorial rules and intellectual contribution also raise separate questions. A wholly human text may contain an unsupported argument. A text prepared with permitted, disclosed assistance may contain a contribution its author understands and can defend. Assessing either requires reading it.

Neither quality nor an ability of the author to defend a text settles every question of authorship. A sound argument may still involve prohibited or undisclosed intellectual delegation. Establishing compliance therefore requires identifying the applicable rule and examining the contribution claimed. When these inquiries are merged, a classification of probable origin can begin to stand in for conclusions about intellectual effort, originality and honesty, each of which requires its own grounds.

Integrity Requires Reasons

There are sound reasons to scrutinise the use of generative AI in academic knowledge production. Fabricated references and misleading accounts of the research process warrant investigation, as do breaches of applicable rules. Questioning indiscriminate suspicion should make those inquiries more precise. Different restrictions may also be justified where the purpose is to assess a particular unaided skill. However, the conditions of a student assessment cannot simply be transferred to research publication.

The current guidance on proposals of the German Research Foundation (DFG) permits AI assistance while prioritising scientific quality and good research practice. Applicants must identify models, purposes and extent of use; tools that do not affect content are exempt, and individual passages need not be marked. These choices make disclosure practicable, but leave open how an evaluator should assess the account provided. A statement about what a tool was used for cannot, by itself, establish how far it shaped the argument.

Would a complete prompting history, including supplied data, resolve that difficulty? Such a record could help investigate a specific question, but examining it would add work where editorial labour is unpaid and time is scarce. A short declaration risks becoming a formal assurance; an exhaustive appendix risks demanding attention that the institution has not allocated.

Jenna Burrell’s 2016 analysis of algorithmic opacity distinguishes secrecy, technical illiteracy and opacity arising from the operation and scale of learning systems. Elena Esposito’s communicative approach to explainability argues that useful explanations need not reproduce opaque internal processes. Applied here, their work suggests distinguishing an account of how an author used a system from an explanation of how it generated its output. Disclosure may support the first inquiry without resolving the second.

Institutions should therefore specify what a declaration is intended to establish and what resources exist to assess it. A concern about misconduct still requires an identifiable restriction and evidence of a breach. Authors should have a meaningful opportunity to respond, and the decision should address their explanation. Accountability cannot depend on complete access to either human or machine thought.

Debora Weber-Wulff and colleagues’ 2023 evaluation of detectors documented limitations across fourteen systems. Those results concern the tools and materials tested; they cannot settle the reliability of every later system. An institution must still explain what a classification establishes in the particular case.

These limits reach into the organisation of academic work. Requiring individual declarations while leaving verification unresourced would shift responsibility without creating a workable practice of assessment. Publication expectations, access to language support and the place of reviewing in academic workloads therefore belong in this debate. Journals should involve authors from different linguistic traditions in deciding which conventions deserve preservation and which forms of assistance warrant acknowledgement. Institutions cannot plausibly demand sustained intellectual scrutiny while treating the time it requires as an unlimited resource.

Authors remain responsible for the accuracy of their claims and the contributions they present as their own. Disclosure cannot cure a false citation, and a declaration of human authorship cannot establish that an argument was adequately researched. Keeping these responsibilities in view allows a concern to be investigated and, where the evidence permits, resolved.

Returning to the Page

Even Winston’s unwritten diary is a compromising possession. Suspicion reaches him before he puts a thought on the page. The warning for academic institutions lies in the pressure to anticipate an accusation whose terms remain uncertain. Investigating that pressure would require listening to authors and examining editorial practices, including the experiences of those who publish across languages. An institution may influence the production of knowledge through the expectations it creates, without ever issuing a prohibition.

Generative AI therefore requires a discussion about how academic institutions recognise intellectual contributions and sustain the work of assessing them. Rules on assistance and disclosure belong within that discussion, but cannot settle it. If writers begin spending their attention on proving innocence through the shape of a sentence, scrutiny will have acquired a cost of its own. A practice intended to protect intellectual judgement may begin to make authors distrust the ordinary work of writing.

Autor/in
Gustavo Leite Neves da Luz

Gustavo Leite Neves da Luz holds a PhD in Law from the University of Hamburg. Specialized in international procedural law, international environmental law, and the law of the sea.

Profil anzeigen
Artikel drucken

Schreibe einen Kommentar

Wir freuen uns, wenn Du mit den Beiträgen auf dem Völkerrechtsblog über die Kommentarfunktion interagierst. Dies tust Du jedoch als Gast auf unserer Plattform. Bitte habe Verständnis dafür, dass Kommentare nicht sofort veröffentlicht werden, sondern von unserem Redaktionsteam überprüft werden. Dies dient dazu, dass der Völkerrechtsblog ein sicherer Ort der konstruktiven Diskussion für alle bleibt. Wir erwarten, dass Kommentare sich sachlich mit dem entsprechenden Post auseinandersetzen. Wir behalten uns jederzeit vor, hetzerische, diskriminierende oder diffamierende Kommentare sowie Spam und Kommentare ohne Bezug zu dem konkreten Artikel nicht zu veröffentlichen.

Deinen Beitrag einreichen
Wir begrüßen Beiträge zu allen Themen des Völkerrechts und des Völkerrechtsdenkens. Bitte beachte unsere Hinweise für Autor*innen und/oder Leitlinien für Rezensionen. Du kannst uns Deinen Text zusenden oder Dich mit einer Voranfrage an uns wenden:
Abonniere den Blog
Abonniere den Blog um regelmäßig über neue Beiträge informiert zu werden, indem Du Deine E-Mail-Adresse in das unten stehende Feld einträgst.