Abstract. This essay examines how theological journals distinguish legitimate AI assistance from the substitution of human authorship, especially when argument and prose develop through sustained human–AI dialogue. Beginning with a practical comparison of submission policies, I reconstruct the objections, documentary checks, mistaken readings, and exchanges that transformed the inquiry into a question about scholarly authority. The analysis distinguishes the value of a contribution, the justification of its claims, the allocation of credit, and eligibility under a journal’s rules. It argues that generated wording, manual typing, and stylistic impressions are inadequate proxies for intellectual contribution. A comparison with formal verification clarifies both the force and the limits of independently checkable results, while the interpretive character of theology raises questions about accountable editorial discretion. Policy differences and detection failures warrant scrutiny, but do not themselves establish deliberate gatekeeping or invalidate every restriction. The essay proposes examining concrete collaborative workflows and the reasons given for permitting or excluding them. Its own development supplies a situated example of the process under investigation, without resolving the general problem of authorship.
Keywords: AI-assisted theology; scholarly authorship; iterative writing; publication ethics; epistemic authority; human–AI collaboration; boundary-work.
How a question about journals became a question about knowledge
I began with a practical question about journals. I wanted to know how publications in theology, religion, and the dialogue between science and religion treated the use of artificial intelligence. Could I use AI to investigate an argument, develop an objection, reorganise a manuscript, or draft substantial passages? Would journals such as Philosophy, Theology and the Sciences, Zygon, Theology and Science, or Modern Theology regard these activities as legitimate research assistance? Were any of them willing to encourage the possibilities, rather than merely tolerate a little grammatical correction?
I was asking about encouragement as well as permission. The word I used was “booster”: could AI make theological research more capable, more searching, or more productive? A policy allowing spelling correction did not answer that question. I wanted to know whether journals could recognise intellectual work developed with AI, and whether the differences among their policies might themselves become a research topic. At that stage, the practical search for a suitable venue and the critical examination of the rules were still intertwined.
As the comparison developed, I became increasingly dissatisfied with the question I had started with. Knowing which journal permitted which activity was useful, but it left a more consequential issue untouched. What feature of scholarly authorship was supposed to disappear when AI participated in developing an argument or producing its wording? If a contribution was insightful, well supported, and open to criticism, what exactly justified excluding it? I found myself asking whether journals were evaluating contributions to knowledge or looking for reassuring traces of a familiar production process.
My reaction was sometimes impatient and severe. I described the suspicion surrounding substantive AI use as if it were being treated as an “original sin.” I wondered whether appeals to human dignity and authentic voice sometimes protected professional status more effectively than they protected scholarship. Those reactions belong to the history of this inquiry, but they cannot serve as its evidence. The discussion became more useful when I began separating what a policy actually said, what an editor might reasonably be concerned about, what an AI system could help establish, and what I was inferring about the motives of other people.
The boundary between assistance and intellectual contribution
The Open Library of Humanities policy was an early source of tension. In the December 2025 text I brought into the discussion, OLH accepted several substantive applications of generative AI, including declared research methods involving data analysis, controlled simulations, and synthetic data. It also accommodated translation and accessibility. At the same time, it prohibited the wholesale generation of the main argument or methodology and the generation of the majority of an article’s text. This was already more complicated than a simple opposition between accepting and rejecting AI. The policy recognised that AI could contribute to research while drawing a restrictive boundary around particular forms of intellectual and textual production (Open Library of Humanities, 2025).
The translation provision deserves closer attention than I initially gave it. OLH says that an AI-assisted translation should meet the journal’s written and intellectual standards regardless of the translation method. Imagine an author who writes a developed argument in one language and uses AI to produce most of its English wording. The submitted sentences could be machine-generated while the argument remains recognisably the author’s. There is a plausible reason to distinguish this from asking for an argument to be invented: translation can preserve intellectual work already undertaken. But that reason concerns the relation between the research and its new expression. It cannot be captured by the mere presence or quantity of generated wording (Open Library of Humanities, 2025).
My first challenge was partly about enforcement. How could an editor establish that an argument had been generated “wholesale,” or that most of the text came from AI, when the final document did not reveal its production history and detection tools were unreliable? I initially moved quickly from difficulty of verification to the thought that the policy might be unusable. But I was also beginning to notice a prior difficulty. Before asking whether a rule could be enforced, I needed to understand what conduct it was trying to distinguish. A prohibition might be difficult to verify because the relevant evidence was unavailable; it might also be difficult to apply because its central categories did not adequately describe the research process.
The assistant proposed a research question: “How do theological journals distinguish intellectual assistance from authorship substitution, and how consistently can those distinctions be applied to iterative human–AI research?” I found this useful, but asked for a narrower focus on the iterative method of theological research, especially substantial writing. Searching, summarising, translation, generating objections, analysing data, and reviewing manuscripts all mattered, yet an inventory of activities could leave the central problem untouched. I wanted to understand why the OLH distinction worked more comfortably for a simulation than for a sustained conversation. A model used to generate a dataset could be described in a methods section. But what happened when a model identified a contradiction in my interpretation, I returned to the sources, and the resulting discussion changed the argument? The contribution might be substantial without being separable into a block of material that could simply be labelled “AI output.” The policy’s categories seemed least settled precisely where my interest was strongest.
I also needed to separate a threshold for disclosure from a threshold for prohibition. In the OLH policy, substantial and known AI use can be acceptable when adequately declared; substantiality therefore does not itself mean misconduct. A short model-generated objection might deserve disclosure because it changed the argument, whereas a long translation might preserve an argument already developed. Declaring either use tells an editor something about the process. It does not automatically establish permission, earn authorship, or demonstrate the truth of the result. My difficulty with “substantial” concerned the different decisions the term was being asked to support, and whether the policy made those decisions distinguishable.
Modern Theology made the problem particularly sharp. Its instructions permit AI research assistance and spelling or grammar improvement, but state: “Authors may not use AI to generate any part of the text of the manuscript.” The same instructions require disclosure, documentation, and human responsibility (Modern Theology, n.d.). The prohibition on generated text is clear enough as a submission rule. Its epistemic rationale is a further question. Suppose a researcher has thoroughly investigated an issue, developed the interpretation through critical exchange, and verified the resulting argument. Why should AI’s contribution to the final exposition necessarily invalidate that work?
Wiley’s general guidance complicated the picture. When a sentence about collaborative drafting appeared in our discussion, I did not initially find it on the linked page and asked where it was located. This was a small but consequential interruption: the assistant’s reassuring formulation needed a traceable source. I subsequently brought the relevant passage back into the conversation. Under “How can I use AI effectively for manuscript drafting?”, Wiley explicitly describes AI helping researchers explore, refine, and express ideas through “collaborative and iterative drafting.” Yet the accompanying questions still ask whether the ideas and analysis are the author’s own and whether suggestions are being selectively implemented (Wiley, n.d.). The guidance therefore gives iterative collaboration a recognised place while leaving an important issue open: how should “my own” be understood when the idea develops through interaction?
Finding the passage briefly changed the practical direction of my questions. I asked about Wiley’s theology journals, their standing, and whether a study of AI submission policies might find a place among them. There was now documentary evidence that substantive drafting could be discussed positively within scholarly publishing. That mattered because my growing impression of hostility was in danger of becoming too comprehensive. Yet the opening did not resolve the conceptual problem, and it did not establish that every Wiley journal would accept the same workflow. The difference between publisher guidance and journal instructions had become part of the inquiry itself.
Further source checking brought an especially direct answer to my original search for encouragement. Wiley’s guidance for editors advises them to assess research quality rather than act as “AI police,” and warns against making detectors the primary basis for decisions. Yet it also says that AI should support drafting rather than serve as the main generator of arguments, text, or conclusions (Wiley, n.d., editor FAQs). This combination was more revealing than a simple classification of the publisher as friendly or hostile. It explicitly resists detection-led evaluation while retaining a boundary around substantive generation. The question therefore returns in a more precise form: what makes the model the main generator in a process whose direction, criticism, and exposition develop over many exchanges?
I could not simply use Wiley’s more permissive language to erase a journal’s explicit restriction. Nor could I infer a unified theological attitude from one restrictive title. The instructions for Scottish Journal of Theology, for example, provide a route for declaring generated text and AI-assisted analysis or extraction of insights (Scottish Journal of Theology, n.d.). The preliminary comparison was useful as a map of published rules. It was neither a representative survey of the whole field nor evidence of how editors actually decide difficult cases. Its most productive result was to expose the need for a better account of the activity being regulated.
The argument that takes shape between turns
At first, I pressed the objection in a relatively limited form. Suppose the intellectual contribution had already developed through serious human–AI collaboration, and AI was then used to organise its exposition: why should the generated sentences disqualify it? That was enough to challenge a blanket prohibition on generated prose. But it left the relation between thinking and writing too tidy. Writing can reveal a missing premise, expose an equivocation, or produce a distinction that was unavailable in the initial outline. The relevant interaction is recursive: I formulate something; the model responds; I object; a revised formulation sends me back to the evidence; what I can then say differs from what I originally intended. A rule permitting AI to express an argument that was already mine would still leave open the harder case in which the interaction helped make the argument possible.
A hypothetical theological example makes the difficulty concrete. I propose that two authors understand a particular doctrine in fundamentally different ways. AI suggests that I have conflated a difference in terminology with a difference in conceptual structure. I check the primary texts and find that the objection is partly right, but that it misses a historical qualification. I introduce that qualification, the model identifies a consequence, and a more precise argument develops. AI subsequently helps draft the exposition. At which turn did assistance become substitution? Merely identifying the first machine-generated sentence cannot answer the question.
The word “own” becomes unstable at this point. It might mean that I first originated an idea, that I understand and endorse it, that I directed the investigation, or that I made a substantial contribution to developing it. These meanings can come apart. I can understand an idea I did not discover, and I can initiate an investigation without anticipating what it will reveal. Intellectual control cannot reasonably require foreknowledge of every useful answer. Yet understanding or endorsement alone cannot establish a discovery claim. A workable account of collaborative authorship needs to distinguish these relations instead of allowing one possessive adjective to stand for all of them.
Nor would requiring me to write every final sentence solve it. I could manually compose the whole manuscript while remaining deeply dependent on an AI-supplied argument. Conversely, AI could draft a large proportion of the wording while I supplied the research question, interpretation, evidential checks, and revisions that made the contribution possible. The proportion of generated prose and the significance of intellectual contribution are different quantities. One suggested distinction can change a project; several pages of drafted exposition can express decisions already carefully worked through.
There is also no guarantee that a series of small interventions remains intellectually small. Replacing “may support” with “establishes,” for example, changes the force of a claim. Repeated revisions can alter the relationship between premises and conclusions even when each request is described as editing. A policy that accepts editing but forbids intellectual generation therefore needs to explain how it deals with cumulative change. Naming a software category or counting words does not resolve that problem. The permission to improve grammar raises the same difficulty at a smaller scale: some changes are straightforward corrections, while others alter emphasis, qualification, or inference. A general instruction can be clear about its intention and still leave these transitions unresolved. In theological writing, where an altered qualification can change the interpretation being defended, the boundary cannot simply be assumed to follow the distinction between language and thought.
The AI assistant helped articulate the question that increasingly concerned me: a journal would need to explain “what necessary feature of authorship is missing from that process.” I responded that this seemed to be the core. The formulation mattered because it shifted the burden from identifying machine-produced sentences to explaining what those sentences were supposed to show about the author’s contribution. Authorship already has meanings beyond literal inscription. The ICMJE criteria, although developed for medical publishing and not universally binding in theology, allow substantive critical review instead of drafting, together with substantial research contribution, approval, and accountability (ICMJE, n.d.). They do not make personally originating every sentence the universal test. The useful AI contribution here was a question that made my dissatisfaction more precise; its persuasiveness did not make it a complete theory of authorship.
This does not mean that accepting responsibility automatically earns authorship. A person cannot establish an intellectual contribution simply by placing their name on a result and promising to answer questions about it. But receiving an insight from AI also does not erase the investigation, discrimination, correction, and development that may follow. The difficulty is to describe those contributions honestly without assuming either that the human conceived everything in advance or that the model supplied a finished intellectual product to a passive recipient.
Reading suspicion without turning it into a verdict
The COPE discussion titled When editors suspect AI intensified my concern. Its introductory indicators include overly formal language, grammatical consistency, particular authorship patterns, and deficiencies such as unsupported references and circular reasoning. My immediate reaction was that very different kinds of evidence were being placed near one another. A citation that fails to support a claim is a substantive problem. A formal register is a stylistic feature. An unfamiliar affiliation or a single-author manuscript concerns the author’s position in an academic community. These may prompt different inquiries, but none should quietly become interchangeable with proof of inadequate scholarship (COPE, 2026).
I was particularly irritated by the suggestion that writing might be suspicious because it sounded excessively academic. I had thought a formal register was among the qualities academic writers were expected to acquire. It seemed perverse to make its successful acquisition a reason for suspicion, especially when an author was writing in another language. That initial reaction still needed refinement. Academic prose should be precise and intelligible; formality alone is no achievement. But a text also does not become intellectually defective because its grammar is unusually consistent or its tone unusually neutral. If a passage is generic, an evaluator should identify the missing specificity. If it is repetitive, the repetition can be shown. A diagnosis becomes less accountable when an observable feature is translated into an impression of machine origin and that impression then substitutes for examination of the argument. My objection concerned that substitution, rather than a demand that every polished or formal passage be admired.
The status of the material mattered as well. COPE’s formal authorship position excludes AI tools from the author list because they cannot fulfil its accountability requirements, requires relevant disclosure, and makes humans responsible for submitted content, including AI-produced portions. It does not impose a universal grammar-only limit (COPE Council, 2024). The forum summary expressly says that its comments do not imply formal advice or consensus. Itamar Ashkenazi likewise states that the appended answers express his personal opinion. A reader can criticise those arguments without converting them into an institutional policy that COPE has not adopted.
A fuller reading also qualified my initial sense that the discussion spoke with one hostile voice. Ashkenazi explicitly acknowledged that the presenters’ medical editorial experience might not transfer to humanities or book publishing, anticipated improvements in AI, and treated the admissibility of a faultless AI-generated paper as an unsettled community question (COPE, 2026). Those qualifications do not answer my objection, but they matter to whom it is directed. The disagreement is partly about the purposes of publication even after obvious errors have been removed. It would be inaccurate to portray every reservation as ignorance of changing capabilities when the speaker expressly anticipates those changes.
This distinction also prevented two questions about AI from collapsing into one. Excluding a system from the author list does not establish that its output cannot contribute to a publishable argument. The requirements attached to a byline and the conditions under which an insight becomes useful are different matters. A model might supply a consequential objection without being a person capable of accepting an author’s obligations. An honest account could acknowledge that contribution while explaining what the human researcher investigated, developed, and took responsibility for. The formal authorship rule does not, by itself, require the fiction that every valuable insight first arose unaided in the named author’s mind.
Here my own reading needed correction. Ashkenazi’s note disclosed Grammarly, a thesaurus, and Microsoft Copilot. I initially treated this as an example of someone using AI while claiming not to use it. The assistant pointed out that the passage made no such denial: Copilot was explicitly named. Grammarly also distinguishes its longstanding foundational assistance from optional generative features introduced in 2023 (Grammarly, n.d.). A product name alone could not establish which operations he had used. The sounder objection concerned the rationale for distinguishing permitted language assistance from prohibited intellectual contribution. My accusation of concealment was not supported by the passage. This correction mattered precisely because I was criticising judgments made from ambiguous signals. If I inferred a denial that the speaker had not made, I would reproduce the same weakness in my own argument. The stronger criticism survived, but it had to be directed at the distinction he defended rather than at an unsupported accusation about his conduct.
A related article located during the source checking for this essay makes the disagreement explicit. Ashkenazi and Howard Browman distinguish an approach centred on the value of the published content from one centred on human creativity, and argue that increasing AI contribution can reduce the human creativity and accountability involved (Ashkenazi and Browman, 2025). A subsequent correction clarifies their organisational roles and that the views are their own (ICES Journal of Marine Science, 2026). I question the implied inverse relationship. Collaboration can change the task and enlarge what becomes possible; it need not divide a fixed quantity of creativity between two competitors. That is a conceptual objection to their argument, not evidence about their private motives.
The appended answers contained a stronger challenge than a simple objection to machine prose. Ashkenazi asked why someone who checks an AI-generated paper should receive authorship credit when a reviewer or reader may also check it (COPE, 2026). That question identifies a weakness in treating approval as a sufficient contribution. My example of iterative research concerns something more demanding than receiving a completed paper and finding no errors. It includes choosing questions, investigating objections, revising interpretations, and deciding what the evidence warrants. A checking process might itself develop into substantial research, but its significance would have to be shown. Neither “I reviewed it” nor “AI helped write it” settles that assessment.
Another answer asked what value the LLM had provided that the author could not have achieved alone. Read as a request to explain the contribution of a tool, this is reasonable. Read as a requirement that AI assistance be indispensable, it demands too much. A method can improve the range of objections considered, the time available for investigation, or the clarity of exposition even when a researcher might eventually achieve something similar unaided. The relevant comparison concerns what the assistance enables under actual conditions. Requiring proof of human impossibility would make convenience, accessibility, and increased capacity strangely inadmissible reasons for adopting a research method.
My frustration also concerned the comparison being made. Human fallibility was often treated as a familiar background condition, while an AI error could be presented as evidence against an entire mode of working. I had introduced King and Worthington’s October 2026 study early in the conversation, attracted by its challenge to assumed human superiority. Checking the paper supplied a more bounded picture. Its corpus contained 124 manuscripts and 134 human reviews; three former editors preferred AI reviews in 26 of 30 judgments across ten sampled pairs. Much of the wider evaluation used AI judges, and the human judges did not also review the underlying manuscripts. The study did not establish whether subsequent revisions improved the papers (King and Worthington, 2026). The result made a simple appeal to human excellence harder to sustain. It did not settle every question about reviewing, theology, or authorship.
The standard I wanted was a comparative one: the actual quality of specified practices, under identifiable conditions, with errors assessed symmetrically. A human reviewer should not receive epistemic credit simply for being human, and an AI system should not receive it simply for being impressive. Complaints about nonexistent references remain relevant when a particular output contains them; they become insufficient when used to dismiss an entire evolving method without examining the workflow in question. I had also remarked that AI could at least admit when it was wrong. Yet the assistant’s willingness to apologise could not establish that its revised position was correct. An apology is an observable response; successful correction requires checking the revised claim. My preference for comparative evaluation therefore had to include the systems whose usefulness had prompted my criticism.
The comparison also had to preserve differences between tasks. The quality of AI feedback does not establish that a reviewer may upload someone else’s unpublished manuscript to a service without permission. The COPE discussion raises confidentiality as a distinct concern and distinguishes seeking a reviewer’s judgment from obtaining an automated report (COPE, 2026). Those distinctions can justify different workflows for authors and reviewers. They do not establish that an AI critique is intellectually worthless. Keeping permission to access material separate from the quality of the analysis prevents a legitimate procedural concern from doing the work of an unsupported claim about capability.
What a typing record can establish
A public comment by “David C” on COPE’s authorship page shifted the discussion toward provenance. The commenter criticised AI detectors and proposed that “The only reliable way to validate human authorship is by proving a human typed the text.” The comment promoted WritersProof and referred to cryptographic process evidence and C2PA provenance (David C, 2026). Its starting point overlapped with my concern: a statistical judgment about a finished document should not be mistaken for secure knowledge of its history. But I found the proposed solution startling. Improving the evidence of production would not settle which feature of production mattered. Were we trying to establish who pressed the keys, or whether the publication contributed anything worth knowing? A criticism of unreliable detection had led to a proposal whose central criterion itself needed justification.
The distinction survives even if the technical recording works perfectly. Someone can type an AI-generated argument without having developed or understood it. Someone else can dictate an original argument without touching a keyboard. A researcher can genuinely revise a document while drawing on an AI conversation outside the recorded session. In each case, the record may accurately describe an activity while leaving the relevant intellectual history unresolved. The commenter acknowledged the possibility of transcription. Raising its cost might have practical value, but it would not make manual entry equivalent to scholarly contribution.
The underlying technical documents were more qualified than the sweeping comment. The CPoE draft recognises retyping as a threat and discusses accessibility accommodations and the need for further validation. These are reasons to evaluate particular implementations carefully, rather than assume that they either solve authorship or are trivially defeated (Condrey, 2026). C2PA’s explainer likewise separates provenance information from establishing the truth or accuracy of the content (C2PA, n.d.). Securely binding a record to a document cannot, by itself, establish that an interpretation is insightful or that the person associated with the record deserves a particular kind of credit.
The limits of final-text detection are related but distinct. Liang and colleagues found that the detectors they studied frequently misclassified non-native English writing, while later work has documented vulnerabilities to transformations and unfamiliar generation conditions (Liang et al., 2023); (Dugan et al., 2024). Such studies concern specified tools and datasets. They do not prove that every present or future detector is equally ineffective. Nor do they establish that every suspicious manuscript is sound. What they do challenge is the conversion of a probabilistic signal into a confident verdict about an individual’s conduct.
David C also suggested that fear of false accusations could lead writers to alter otherwise satisfactory prose. I had no evidence establishing how widespread that behaviour was, and the comment could not supply it merely by asserting it. The possibility nevertheless sharpened the institutional question. If an author had to make accurate, clear writing look less machine-like to pass an informal test, the test could work against the quality it was supposed to protect. The relevant concern was then broader than a mistaken accusation against one writer: it included the practices a system of suspicion might encourage.
A more basic limitation can be stated without a detector benchmark. Different production histories can produce the same final text. A method that sees only that text receives the same evidence in both cases and cannot invariably recover which history occurred. Additional records can reduce uncertainty, but their interpretation still requires an account of what counts as contribution. Even an oral defence establishes a person’s current understanding more directly than it establishes the historical origin of the argument. Someone can learn an externally supplied argument deeply; learning, checking, and extending it may then become valuable work in their own right.
The objection I had to make to the AI
At one point, I put the argument in particularly provocative terms. If AI could find ways around these constraints, I suggested, perhaps the concerns themselves were intellectually infantile. To assess that claim, I need to make the possible routes concrete. Some are experimentally demonstrated ways of reducing detector performance. Others show how a person could satisfy a visible requirement without establishing the intellectual contribution that requirement is supposed to certify. The following examples distinguish the procedure from what it actually proves.
-
Retype AI-generated prose instead of pasting it. A person obtains a passage from AI, opens a document with writing-history capture enabled, and manually types the passage. The recorded keystrokes are real, and the person genuinely performs the recorded labour. Nevertheless, the wording and argument can have originated elsewhere.
What this gets around: a test equating physical typing with intellectual authorship. The CPoE draft explicitly recognises this “retype attack” and proposes behavioural defences, so this example does not establish that every implementation would accept the record. It establishes that authentic typing alone cannot settle intellectual origin (Condrey, 2026).
-
Develop the argument before the recorded writing session. A researcher discusses a thesis, its objections, and its structure with AI, then begins a tracked document and writes from the resulting understanding. The session contains genuine hesitation, deletion, revision, and correction. None of those events needs to be fabricated.
What this gets around: the assumption that an authentic document history is a complete research history. A record beginning when the document is opened cannot establish everything that shaped the argument beforehand. Its reliability within that interval does not make its coverage complete.
-
Obtain the intellectual structure from AI, then write all the sentences personally. For example, a researcher asks for a central claim, supporting premises, the strongest objection, and a response. They then compose the manuscript in their own wording without copying any generated sentences.
What this gets around: a rule concerned exclusively with AI-generated prose. The final wording may be human-composed while the intellectual structure remains substantially AI-supplied. A policy covering AI-generated arguments might still prohibit the process; a sentence-origin test cannot establish compliance with that broader rule.
-
Keep the generated wording small while retaining its decisive intellectual contribution. Imagine a 5,000-word paper in which AI supplies a thirty-word distinction that resolves the central difficulty. The researcher writes the remaining exposition, including the explanation and consequences of that distinction.
What this gets around: treating the percentage of generated wording as a measure of intellectual significance. The retained AI wording represents only 0.6 per cent of the text, yet its contribution might be decisive. Conversely, extensive generated exposition could communicate an analysis the researcher had already developed. Even perfect word counting would measure the wrong quantity.
-
Distribute a substantial transformation across small requests. Begin with a human draft, then request changes paragraph by paragraph: clarify this distinction, qualify that claim, reorganise these premises, revise the response to this objection, and improve the transition to the conclusion. Feed the revised version into the next exchange. Each intervention is local, but their cumulative effect can transform the argument.
What this gets around: an assessment that classifies each request separately as minor editing without evaluating the resulting manuscript. Some requests clearly involve intellectual development; calling them editing does not settle their status. The gap appears when a policy regulates isolated operations but leaves their cumulative effect undefined.
-
Perform an equivalent rewrite through a differently labelled interface. In a hypothetical comparison, submit the same paragraph and revision instruction to a generative feature inside a grammar application and to a chatbot. Both might reorganise sentences, supply formulations, and alter emphasis.
What this gets around: a classification based on product names, such as permitting a grammar assistant while rejecting equivalent chatbot assistance. This is a conditional counterexample, not a claim that every product behaves identically. Where the functions are equivalent, the interface label cannot explain the different authorship judgment.
-
Change the generator’s stylistic behaviour before producing the text. Pedrotti and colleagues describe a public research pipeline using paired human and machine texts and Direct Preference Optimization to fine-tune language models toward human writing styles. They then generate fresh material and evaluate it against existing detectors. The resulting stylistic shifts reduced detection performance in the tested settings (Pedrotti et al., 2025).
What this gets around: detectors’ dependence on features associated with familiar generators, such as particular lexical and syntactic patterns. This is a model-training method with published code and data, not a demonstrated universal prompt for making any text undetectable. Its importance is that machine origin remains unchanged while the detector’s evidence changes.
-
Paraphrase generated text repeatedly. The concrete sequence is: generate a passage, pass it through a paraphrasing model, then use that rewritten output as the input to another paraphrasing pass. An illustrative instruction would be: “Rewrite this passage using different syntax and vocabulary while preserving its claims and qualifications.” In a controlled evaluation, compare detector results for the original and transformed versions, while checking whether meaning has survived.
What this gets around: some statistical signals associated with the original generation. Sadasivan and colleagues tested recursive paraphrasing against several detection approaches and reported substantial reductions in detection rates under their experimental conditions. Success varied. The DAMAGE study also reported a detector robust against the humanising transformations and adaptive attack it evaluated, so repeated rewriting is not a guaranteed bypass of every detector (Sadasivan et al., 2025); (Masrour et al., 2025).
-
Learn an externally generated argument well enough to defend it. A person receives an AI-developed argument, studies its premises, checks its sources, works through objections, and becomes able to explain it without assistance. In an oral examination, they may answer demanding questions accurately.
What this gets around: the inference that successful explanation proves personal origination. The examination may correctly establish understanding while leaving the history of discovery unresolved. Furthermore, checking and extending the argument might become a real contribution. The example separates understanding, discovery, and development; it does not imply that understanding acquired through AI is counterfeit.
-
Obtain valid provenance for a document whose intellectual origin remains mixed. A person edits an AI-assisted document and uses a provenance-enabled workflow to sign assertions about the resulting file and recorded changes. A verifier may successfully confirm that the credential is valid, associated with that file, and unaltered.
What this gets around: treating successful provenance verification as proof of independent human authorship or scholarly merit. The cryptography can work exactly as intended. C2PA explicitly distinguishes provenance from establishing whether content is true, accurate, or factual. A valid credential therefore needs to be read for the particular assertions it authenticates; it is not a general certificate of intellectual contribution (C2PA, n.d.).
These examples do not all describe the same kind of bypass. Paraphrasing and stylistic adaptation have been tested against detectors; the other cases expose gaps between a criterion and the conclusion drawn from it. Several procedures could violate a particular journal’s policy, while others could be legitimate and openly disclosed. Their analytical importance is that an authentic signal—real typing, genuine revision, successful explanation, or a valid signature—can coexist with a more complicated intellectual history. Different histories can also produce the same final text. A system observing only that text cannot invariably distinguish those histories, because the evidence available to it is identical.
The assistant replied to my broader criticism that being bypassable did not make a rule meaningless, adding: “A rule against fabricating evidence remains justified even when fabrication is difficult to detect.” That general point was logically sound. But I objected to its application. I was discussing the treatment of a genuine contribution produced through collaboration, including practices whose legitimacy remained disputed. Fabrication introduced a different wrong and made the restrictions look analogous to an already defensible prohibition. The analogy risked taking for granted precisely what I was questioning: that substantial AI collaboration belonged in a category of suspect conduct. The examples sharpen the question I wanted answered: what scholarly value does the restriction protect, and does the evidence used to enforce it actually establish that value?
My objection redirected the inquiry. Suppose a human–AI process really does produce a valuable interpretation, a defensible argument, or a discovery. Why should the journal devote its attention to ambiguous signals of AI involvement instead of examining what has been learned? In a disputed area, an activity cannot be classified as intellectual fraud simply because it does not fit an inherited image of solitary authorship. The legitimacy of the restriction must be argued before its circumvention can do all the moral work assigned to it. This was a correction to the object of the conversation. The assistant’s answer had preserved a valid general proposition while losing the particular question that made the exchange necessary. My rejection did not establish that every restriction was invalid; it required the next answer to address whether the restriction had been justified in this case.
This did not make the history of research irrelevant. My phrase “provided that people are actually gaining new knowledge” carries a considerable burden. Readers need grounds for believing that the contribution is sound. They may need to reconstruct a method, inspect a source, follow a sequence of inferences, or understand why an alternative interpretation was rejected. A reliable account of the process can help them do those things. Yet the process detail that helps establish a conclusion may differ substantially from the detail that shows how much of the final prose someone physically entered.
I began to distinguish three decisions: whether something deserves publication as a contribution, what reasons justify accepting its claims, and how its intellectual history should be credited. A result can have value even when AI supplied its decisive insight. The person who investigates and explains it should describe that role accurately. There is no need to pretend that the human independently discovered everything in order to recognise the value of what has become available to others. Equally, worthwhile content does not automatically settle who deserves the credit attached to it.
The “cat and mouse” problem then became an institutional concern. Imagine two comparable manuscripts produced through extensive AI collaboration. One author provides a candid account; the other makes the collaboration unobtrusive. If visible AI involvement is penalised independently of scholarly quality, the candid author can bear the greater burden. Authors may be encouraged to optimise the appearance of independent composition, and journals may learn less about the processes that actually produce their submissions. This is a plausible consequence of the incentive structure, not evidence that journals already select mainly for skill at concealment.
I wanted an alternative closer to the familiar principle of standing on the shoulders of giants. Scholarship develops through inherited concepts, criticism, translation, instruments, conversation, and the work of others. AI changes the scale and character of those relations, and its outputs need appropriate scrutiny and attribution. But intellectual dependence itself cannot be the disqualifying feature: inquiry would scarcely begin without it. My concern was that a useful new form of dependence might be condemned through a conception of originality that scholarship could not consistently apply to its own history.
The analogy also needed a limit. An AI synthesis is not a transparent substitute for the sources on which a scholarly claim rests. OLH explicitly requires authors to consult the works they cite and to verify generated citation information (Open Library of Humanities, 2025). Finding that a reference exists is only one check; the source must also support the use being made of it. For theological interpretation, a fluent account of a thinker’s position still needs to be tested against the relevant texts and context. Standing on others’ work becomes scholarly dependence through the reasons one can give for relying on it.
Where a proof can force recognition
The comparison with mathematics gave my frustration a sharper form. I had brought into the discussion OpenAI’s 26 October 2026 announcement of claimed advances concerning local smoothing, the Exponential Time Hypothesis, and quantum field theory. The announcement described AI-produced arguments, formalisation, and different arrangements for external examination (OpenAI, 2026). Faced with that scale of claimed achievement, familiar complaints about machine-like style seemed particularly inadequate. I wondered whether formal disciplines were being forced to confront AI’s capabilities while interpretive disciplines could continue to protect themselves through less determinate standards. I expressed the contrast starkly: if AI produced a correct result, mathematicians would have to accept it, potentially even when it threatened work through which they had earned their living. In theology and the humanities, I suspected, less determinate judgments might leave more room to refuse recognition. This was the conjecture I wanted to examine, not something the announcement itself demonstrated.
Those announcements could not be treated as settled results simply because their implications were dramatic. The group organised around the ETH claim, led by Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, explicitly said that its involvement was not certification. It called for independent examination and recognition of the work of checking, explaining, correcting, and extending results (Impagliazzo, Paturi and Williams, 2026). That response did not fit a simple story of experts defending their status by refusing to look. It described a way of making a potentially disruptive contribution publicly examinable.
The announcement’s account of credit also complicated my original contrast. OpenAI distinguished the original AI-produced discoveries it was claiming from the understanding, explanations, and materials that external researchers might subsequently produce, and proposed recognition for that later work (OpenAI, 2026). Whatever became of the particular results, the distinction supplied a possible way of describing contribution without pretending that the human investigators had originated the discovery. It left room for valuable work after generation. That possibility mattered to theology as well: investigating, interpreting, and developing a consequential suggestion can contribute to knowledge even when its initial appearance is honestly attributed elsewhere.
The epistemic asymmetry I had noticed nevertheless survives in a more precise form. Once a proof has been correctly checked against an agreed formal statement and assumptions, the producer’s identity cannot reasonably invalidate that inference. A mathematician’s dislike of AI does not change what follows. Formal checking can therefore constrain a particular kind of discretionary refusal. But it remains necessary to establish that the encoded statement represents the intended claim. Lean’s own documentation distinguishes verification from questions about the meaning of the statement being verified (Lean, n.d.).
Coding requires a similar qualification. A program may pass all the tests we happen to have while failing in an untested situation. A verified implementation may satisfy a specification that poorly describes what users need. Correctness, significance, maintainability, explanatory value, and suitability for use are related but distinct judgments. My initial shorthand—mathematics and coding have clear inputs and outputs—was useful for identifying some cases, but too broad to characterise whole disciplines. The better comparison concerns which claims can be independently checked, by what means, and at what cost.
The current policies also resisted a clean disciplinary division. The London Mathematical Society’s October 2026 policy expressly accommodates AI use in generating ideas and constructing or verifying proofs, subject to disclosure, responsibility, and editorial judgment (London Mathematical Society, 2026). Yet the June 2026 Leiden Declaration also treats human credit, understanding, autonomy, and the organisation of mathematical research as important concerns (Leiden Declaration, 2026). Acceptance of a valid result does not automatically settle how institutions should distribute recognition or support people whose work changes. Theological journals likewise differ among themselves, as the contrast between Modern Theology and Scottish Journal of Theology showed.
Humanities research has publicly discussable constraints even when it lacks a decisive automated checker. A theological interpretation can misquote a source, ignore its historical setting, confuse a technical term, or draw a conclusion that its premises do not support. Two interpretations can also remain defensible because they illuminate different features or make different background commitments. A reader can inspect those commitments without sharing them. Such plurality does not make every interpretation equally good, and forcing a single numerical output may conceal the very question under investigation.
Even within a theological article, the standards may differ from claim to claim. Whether a passage occurs in a particular text, whether a translation preserves a technical distinction, whether a conclusion follows from stated premises, and whether an interpretation is illuminating are different questions. Some permit relatively direct checking; others require comparison among competing explanations. Describing the whole field as subjective would obscure those distinctions and weaken my criticism. The useful demand is that an evaluator identify the kind of claim at issue and the grounds on which it succeeds or fails, rather than invoke the discipline’s complexity as a general exemption from explanation.
The asymmetry therefore concerns the scope for challenge. A formal proof can sometimes settle a narrowly defined dispute with exceptional decisiveness. A judgment about theological depth will usually require more explanation and may remain contestable. That gives the evaluator greater responsibility to articulate reasons. It should not give them a privileged space in which an impression of authenticity becomes sufficient grounds for exclusion.
The reference to livelihoods also exposed a further distinction. The value of people, the organisation of academic employment, and the quality of a particular argument are related questions, but success at one task cannot decide all three. A change in who or what can perform intellectual work may create serious institutional problems. Those problems deserve discussion in their own terms. My resistance to appeals to “human dignity” concerned their use as an unexplained reason to discount a contribution. Recognising that a model performs a task well does not require treating human worth as a competition that a person loses whenever a machine performs better.
When discretion becomes a defence of authority
This was the point at which my criticism became explicitly institutional. Could scholars exploit the interpretive character of their fields to maintain control over what counted as a contribution? Could they move the criterion from accuracy to originality, then from originality to understanding, and finally to an unspecified human quality whenever AI met the previous demand? I found that possibility intellectually troubling. A standard that continually relocates itself to preserve a predetermined exclusion cannot function as an open test of a contribution.
The assistant connected this concern to Thomas Gieryn’s concept of “boundary-work.” Gieryn examined how descriptions of legitimate science can help secure professional authority, resources, and autonomy. His cases concerned scientists, which immediately complicated any temptation to treat gatekeeping as a peculiarly humanistic defect. He also resisted reducing scientific values to cynical inventions (Gieryn, 1983). The connection gave my suspicion a sociological vocabulary. It did not establish that the editors I had been reading were knowingly dishonest.
Three possibilities needed to remain separate. A subject can contain genuine interpretive uncertainty. A profession can develop discretion for managing that uncertainty. Particular actors or institutions can then use that discretion to protect their position. The first two do not prove the third. Equally, the absence of evidence about a person’s intentions does not prevent criticism of a rule’s effects. A policy can disadvantage transparent collaboration or unfamiliar contributors even when its authors sincerely believe they are defending scholarly standards. I had called the suspected behaviour intellectually dishonest. To make that accusation about identifiable people, I would need evidence of knowing misrepresentation or strategically inconsistent judgment. The texts already supported questions about justification and application. They did not reveal their authors’ private intentions, and my frustration could not supply the missing evidence.
There was a concrete institutional mechanism to examine in the COPE material. Its forum summary discusses checking publication histories, considering mentor coauthorship for early-career researchers, and accepting that stricter vigilance may exclude some sound papers (COPE, 2026). Editorial time is finite, and the medical setting helps explain the urgency attached to trust. Yet these responses can shift evidential weight from the manuscript toward a researcher’s existing position in the profession. A first publication, an interdisciplinary change of subject, or legitimate single authorship can then become a disadvantage. The acknowledgment of possible error is not itself an explanation of why that distribution of error is justified. My question becomes how journals distinguish a necessary screening practice from a rule that makes entry depend increasingly on already having been recognised.
There are conceivable ways to examine this problem. If an evaluator judges an argument illuminating under one account of its origin and intellectually empty under another, the change deserves explanation. But the judgment being measured must be specified. A change in an assessment of evidential strength differs from a change in a decision about authorship eligibility. Production information can sometimes be relevant to trust, especially where underlying evidence has not been independently checked. A study would need to control those differences before treating any response to an AI label as prejudice.
Likewise, an appeal to understanding is not inherently a retreat from evidence. It may identify something important that a correct answer alone does not provide. The problem arises when the quality is left undefined, applied asymmetrically, or made impossible for any AI-assisted process to demonstrate by definition. If only human origin can count as evidence of genuine understanding, an evaluator has adopted a substantive commitment about what scholarship should be. That commitment needs to be defended openly. It cannot simultaneously serve as an empirical demonstration that AI-assisted work lacks intellectual value.
I would also apply this demand for justification to AI companies and their advocates. A company’s announcement is not an independent evaluation of its own achievement. Neither technical fluency nor commercial confidence removes the need for scrutiny. My objection to academic authority would become inconsistent if I simply transferred that authority to the producer of the model. The relevant ambition is to make reasons, evidence, limitations, and contributions available for examination, including when they come from institutions whose technologies I find extraordinarily useful.
The research question that emerged was therefore more careful than my initial accusation: does limited independent verifiability give journals greater scope to exclude AI-assisted contributions through criteria whose application is difficult to challenge? This can be investigated through policy analysis and, potentially, studies of evaluative judgments. Neither policy texts alone nor an experiment with attribution labels would establish private motives. They could nevertheless reveal patterns of discretion, inconsistency, and unexplained exclusion that deserve attention.
Following a journal’s rules without making them an epistemology
The conversation eventually returned to a practical difficulty. If the boundaries remain disputed, what should a researcher follow? The immediate answer is that a particular journal’s applicable requirements govern eligibility for submission there. A general publisher statement does not establish permission where a journal explicitly imposes a stricter rule. Where guidance conflicts or fails to address a proposed workflow, a specific written clarification can resolve a local uncertainty. I have not tested these cases through editor correspondence or submitted a manuscript as an experiment.
Earlier, I had asked whether Modern Theology’s restriction was reasonable, lawful, and acceptable. Those words bundled together several questions that the policy comparison could not answer at once. This inquiry had not examined a jurisdiction, a publishing contract, or a particular legal dispute, so it could not deliver a legal conclusion. What it could examine was the gap between an instruction’s institutional force and the reasons offered for it. Knowing which rule a venue applies can resolve a submission decision while leaving the adequacy of the rule open to criticism.
But local permission and epistemic justification answer different questions. A journal can permit a method that produces a weak paper, and it can prohibit a method capable of producing a strong one. Its authority to set conditions does not demonstrate that those conditions identify the boundaries of legitimate knowledge. Conversely, the absence of consensus does not make every interpretation of a rule equally persuasive or every workflow equally responsible. Reasons are still available even when agreement is incomplete.
I also came to distinguish disagreement from vagueness. Two journals may adopt clear but conflicting rules. A single journal may instead use terms such as “substantial,” “own,” or “assistive” whose application remains uncertain. A clear rule can be poorly justified; a rule with a legitimate aim can be difficult to apply. Clarity, justification, consistency, and enforceability are separate features. Treating them as one problem makes it too easy either to condemn every policy as useless or to assume that a published instruction settles the underlying argument. The lack of a common standard therefore did not leave me with a choice between obedience as a complete epistemology and unrestricted individual preference. I could follow an applicable requirement, describe its limitations, and argue for a different one without pretending these were the same act.
A defensible policy should identify the value it protects and explain how its criterion relates to that value. If reliability is the concern, the relevant question is how a workflow affects the quality and checking of claims. If credit is the concern, the journal needs an account of contribution. If a particular publication wishes to preserve a specifically human practice, that is a normative commitment requiring its own explanation. The evidence needed for each purpose differs. More documentation is not automatically better if it records an activity unrelated to the decision being made.
This also explains why the title’s alternatives cannot be treated as entirely exclusive. A publication communicates research and can also become evidence of a person’s competence in hiring, promotion, or assessment. AI exposes tension between these uses when a contribution remains valuable but its production no longer demonstrates unaided performance. An examination designed to assess a student’s independent skill may reasonably impose conditions that would need a different justification in a research journal. If a journal is partly certifying a particular human capability, it should say which capability and why its publication process is an appropriate test. Otherwise, an unstated credentialing purpose can determine what is presented as a judgment about knowledge.
Even a complete conversation history would not allocate intellectual credit automatically. An objection can have several antecedents; a short intervention can redirect a long inquiry; a suggestion can matter because someone recognised, tested, or transformed it. Records make those histories more accessible, but interpreting their significance remains intellectual work. I would favour proportionate disclosure of consequential AI involvement, explicit accounts of contribution, and scrutiny of the resulting scholarship. Restrictions based on the origin of wording should carry an accompanying explanation of what they protect.
For a comparative study, this suggests using concrete scenarios rather than treating “AI use” as a single variable. One could compare AI drafting of an already-developed argument, AI contribution of a decisive objection followed by human investigation, and iterative development in which both argument and exposition change. A further case would reverse the usual visible signal: a researcher writes every sentence manually while relying on an AI-generated conceptual structure. These cases would expose whether a policy regulates wording, intellectual dependence, the exercise of judgment, or some combination of them. Comparing declared reasons with the treatment of the same scenarios could be more informative than ranking journals on a simple scale from permissive to restrictive. Any later study of reviewers would need to distinguish assessments of quality, contribution, and publication eligibility. At present, these are proposals for inquiry, not findings about how a population of editors behaves.
The study would also need to distinguish an incomplete policy from an inconsistent application of it. Independent readers could apply each published policy to the same described workflows, record the clauses they rely on, and explain where they cannot reach a determination. Later clarification from editors could be compared with those readings. Agreement would not prove that a rule was justified, but disagreement could reveal where authors cannot reliably anticipate its operation. This would test the usability of the categories without relying on a detector to reconstruct a hidden writing history. It would remain a study of policies and judgments, rather than a demonstration that the hypothetical papers themselves deserved publication.
The inquiry has become one of its own cases
This essay cannot stand outside the problem it discusses. My questions, frustration, examples, and objections drove the exchange. The AI assistant helped retrieve policies and scholarship, proposed distinctions, identified counterexamples, and drafted formulations. I challenged its fabrication analogy because it changed the object of my criticism. It challenged my reading of Ashkenazi’s disclosure because the text did not support the accusation I had made. The discussion developed through those interruptions. A polished account that assigned all the insight to one participant would misrepresent what the conversation records.
The AI has also drafted and helped revise the present prose. That fact is especially relevant in an essay about substantive AI-assisted writing. Calling its contribution mere proofreading would be inaccurate. Calling the resulting inquiry a finished argument that I simply received would also erase the questions and corrections through which it developed. The examples considered here expose the inadequacy of both descriptions without establishing a universal formula for authorship. My agency appears in the narrowing of the problem, the demand for a source, the rejection of an analogy, and the decision to keep an uncertainty open. Those actions need to remain visible alongside the substantial work performed by the model.
The first draft of this essay then became another input to the inquiry. After reading it, I asked for a second pass that would restore the history of thinking rather than simply add information. That request bears directly on the argument. A polished reconstruction can make a distinction appear to have been mine from the beginning, turn an AI proposal into an apparently settled conclusion, or make a correction look inevitable. Returning to the exchange helps expose those distortions, but the reconstruction still involves choices about what mattered. The archive constrains that judgment; it does not replace it. This revised essay is therefore also an attempt to make the account of collaboration more faithful to the collaboration it describes.
There is a further risk in having AI help reconstruct the inquiry in my voice. The model can turn a distinction it proposed into something I supposedly realised, or convert a request for more detail into agreement with every qualification it supplied. Preserving the exchange means resisting that retrospective transfer of ownership. The distinctions developed here remain arguments to be assessed; the record of my questions does not automatically establish that I endorsed every later formulation. Asking that the complete earlier version be retained is itself part of this effort: recovery should make the intellectual history more inspectable, rather than replace it with a smoother story of certainty.
The source checking does not absolve me of responsibility for the argument, nor does the coherence of the prose certify every inference. What the archive offers is a more inspectable account of development: what I initially suspected, what the evidence supported, where a comparison failed, and what questions remained. It also makes some mistakes visible. That is part of its value. The purpose of preserving this history is not to provide a certificate of untouched human originality. It is to let a reader understand how the inquiry arrived at its present form.
I began by asking which theological journals would allow me to use AI. I now want to know what those journals understand themselves to be recognising when they recognise scholarly work. A contribution to knowledge, the justification of a claim, the allocation of credit, and compliance with a venue’s rules are connected, but they cannot substitute for one another. The question I would carry into further research is this: how should theological journals justify and apply authorship restrictions when intellectual contributions emerge through interaction and cannot be cleanly assigned to a single source?
I remain convinced that substantial human–AI collaboration deserves serious opportunities for examination. The possibility that it changes the distribution of expertise or recognition is a reason to investigate its consequences. It is not sufficient evidence against what the collaboration produces. A scholarly community should be able to explain what it has learned, what it has reason to accept, and why it recognises a contribution. I want the rules of publication to remain answerable to those purposes, especially when a new method makes the inherited boundaries uncomfortable.
Source and method note: this essay reflects a discussion and source review conducted on 29 October 2026. The journal comparison is exploratory. Policy descriptions are dated snapshots, not evidence of actual acceptance or rejection decisions. Where live access was blocked, the discussion used supplied policy text or indexed official material. The COPE forum contributions and public comments cited below are distinguished from formal institutional guidance. AI assisted the research, organisation, argument development, and drafting and revision of this essay.
References
Ashkenazi, I., and Browman, H. I. (2025). “What’s the point of generative artificial intelligence in science and scientific publishing?” ICES Journal of Marine Science, 82(10), fsaf179. https://doi.org/10.1093/icesjms/fsaf179. Read with the 2026 correction below; the authors’ views are not institutional policy.
C2PA (n.d.). “C2PA and Content Credentials Explainer.” Specification version 2.2. Official explainer. Accessed 29 October 2026.
Condrey, D. (2026). “Cryptographic Proof of Effort (CPoE): Architecture and Evidence Format.” Individual Internet-Draft, draft-condrey-cpoe-protocol-00, May. Experimental proposal; work in progress. Draft text. Accessed 29 October 2026.
COPE Council (2024). “Authorship and AI tools.” COPE position statement. The page records a review date of 13 February 2023 and version publication on 20 August 2024. https://doi.org/10.24318/cCVRZBms. Accessed through indexed official text on 29 October 2026.
COPE (2026). “When editors suspect AI.” March forum discussion, summary, and appended personal responses by Itamar Ashkenazi. Discussion page. The supplied page text states that forum comments do not imply formal COPE advice or consensus and that Ashkenazi’s answers express his personal views.
David C (2026). Public comment on “Authorship and AI tools,” 30 July, 12:24 BST. COPE position statement and comments. The comment is cited as an individual contribution, not as COPE policy.
Dugan, L., Hwang, A., Trhlík, F., Zhu, A., Ludan, J. M., Xu, H., Ippolito, D., and Callison-Burch, C. (2024). “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors.” Proceedings of ACL 2024, pp. 12463–12492. https://doi.org/10.18653/v1/2024.acl-long.674.
Gieryn, T. F. (1983). “Boundary-Work and the Demarcation of Science from Non-Science: Strains and Interests in Professional Ideologies of Scientists.” American Sociological Review, 48(6), pp. 781–795. https://doi.org/10.2307/2095325.
Grammarly (n.d.). “How can I turn off AI in the Grammarly product offerings? Does the use of Grammarly trigger AI detectors?” Official support documentation. Accessed 29 October 2026.
ICES Journal of Marine Science (2026). “Correction to: What’s the point of generative artificial intelligence in science and scientific publishing?” 83(3), fsag040, 17 March. Correction notice.
International Committee of Medical Journal Editors (ICMJE) (n.d.). “Defining the Role of Authors and Contributors.” Recommendations. Official recommendations. Accessed 29 October 2026.
Impagliazzo, R., Paturi, R., and Williams, R. (2026). “ETH Refutation: Examination and Discussion.” ETH: Community Examination and Discussion, 26 October. Community statement.
King, G., and Worthington, S. (2026). “AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices.” Manuscript dated 28 October. Paper; Project page.
Lean (n.d.). “Validating a Lean Proof.” The Lean Language Reference. Official documentation. Accessed 29 October 2026; the “latest” documentation is updated over time.
Leiden Declaration (2026). Leiden Declaration on Artificial Intelligence and Mathematics, 2 June. Declaration text; https://doi.org/10.5281/zenodo.20302944. Accessed 29 October 2026.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). “GPT detectors are biased against non-native English writers.” Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779.
London Mathematical Society (2026). “Ethical Policy for Journals.” Updated October, especially sections 2.1 and 2.5. Official policy. Accessed 29 October 2026.
Modern Theology (n.d.). “Policies and Instructions to Contributors,” section 10, “Policy on Disclosing Use of Artificial Intelligence (AI).” Journal instructions. Accessed 29 October 2026.
OpenAI (2026). “Working with academic communities to build scientific understanding.” 26 October. Research announcement. Cited as a report of the company’s claims and proposed examination process, not as independent certification of the results.
Open Library of Humanities (2025). “OLH Artificial Intelligence (AI) Policy.” Last revised December. Official policy page. Analysed from the full policy text supplied in the discussion; live retrieval was blocked during this source review.
Scottish Journal of Theology (n.d.). “Preparing your materials,” section “Use of Artificial Intelligence (AI) Tools.” Cambridge University Press. Journal instructions. Accessed 29 October 2026.
Wiley (n.d.). “Using AI tools in your research,” especially “How can I use AI effectively for manuscript drafting?” Author guidance. Checked against the supplied excerpt and indexed official text on 29 October 2026.
Wiley (n.d.). “Editor FAQs AI guidelines.” Guidance for editors. Accessed through indexed official text on 29 October 2026. Cited in the text as “Wiley, n.d., editor FAQs” to distinguish it from the author guidance.