Executive Summary

In order of publication, the three run like this. On 30 June 2026 came the Research Ethics Guide to Generative AI for University Researchers, from the Ministry of Education, the National Research Foundation of Korea and the Council for University Research Ethics. In August came the Guidance on Generative Artificial Intelligence from the US Office of Research Integrity. In September came the AI Research Ethics Guide for National R&D and its commentary volume, from the Ministry of Science and ICT and the Korea Institute of S&T Evaluation and Planning. Below they are shortened to the University Guide, the ORI guidance and the National R&D Guide.

Line up the page counts and you get 65, 54 and 13. The temptation is to read those numbers straight off as a scale of diligence, and there is one thing to separate out before doing so. The three documents are addressed to different people. The two Korean documents speak to researchers. The University Guide states its purpose as guiding "researchers (faculty, research staff, graduate students and others) to use AI responsibly," and the National R&D Guide says its "primary audience is every researcher using AI in national R&D." The first sentence of the ORI guidance, by contrast, nails its reader down as the institution.

One side answers a researcher asking what to do; the other answers an inquiry committee asking how to handle evidence. Standing the page counts of documents from different genres next to each other is not measuring with the same yardstick in the first place. Count how the pages are allocated and the difference widens by another layer. Of the ORI guidance's thirteen pages, the commentary ORI newly wrote runs from page 4 to page 7 — four pages — and the remaining six reproduce the text of 42 CFR Part 93 as an appendix. So "the US did thirteen pages" inflates it and "the US did four pages" undersells it. Both numbers have to be written down.

The table below stands the three documents up against the same set of items. The five rows under the page-count row are worth reading first.

Page counts were measured on the downloaded PDFs. The National R&D Guide and its commentary are not separate files; they run one after the other inside a single PDF. So the accurate description is not "two volumes were issued" but "a guide and a commentary were laid out in separate parts."

1.1The US Document Is Thin Because It Answers Something Else

A different addressee brings a different question. The four pages ORI newly wrote cover things the two Korean documents do not. Five passages make the character of the document plain.

• A fake reference produced by AI may not be fabrication. The definition of fabrication attaches to "data or results," and references are not generally that. But a condition follows. The original continues: "However, in some cases, such as literature reviews, references are data because they arise from scientific inquiry. In these cases, fabricated citations could be the subject of research misconduct allegations." Carry over the first clause alone and the practical advice inverts.
• Do not treat an AI detector as the sole basis. ORI recommends against relying on the output of such tools as the sole source of evidence, and adds that plagiarism detection tools may have limited ability to catch plagiarized text inside AI-generated scientific content. This is the same point Pebblous covered earlier in the limits of AI text detectors.
• The automatic image processing in a smartphone camera is itself a kind of generative AI, and such cameras might be used to record research data. The burden of proving honest error rests with the respondent.
• ORI does not require institutions to adopt an AI policy. It notes only that such a policy may help committee members adjudicate cases.
• How records are handled is itself evidence. A respondent who claims to hold records and then fails to produce them on request has furnished evidence of misconduct, and the same holds for records destroyed intentionally after notice of an allegation. Failure to produce may also be an aggravating factor in administrative action. Institutions must generally retain the institutional record and all sequestered evidence for seven years.

All five belong to the seat where a committee handles evidence. The two Korean documents answer what a researcher should do; the ORI document answers how a committee should treat evidence. That leaves the last two rows of the table. Only the US document has no AI authorship clause. The usual reading is that it was unprepared. Open the regulation and an entirely different reason appears.

One more thing to set aside. "Three in three months" holds only once the frame is narrowed to AI research ethics. Widen it to the research-ethics output of these agencies as a whole and this is not the first round. The National Research Foundation of Korea issued Recommendations for the Responsible Use of Generative AI Tools in March 2024 and a revised edition in September 2025. The Ministry of Science and ICT and KISTEP established the National R&D Research Ethics Handbook in 2022, revised it several times, and issued this AI guide after the November 2025 revision. On the US side, ORI amended 42 CFR Part 93 in September 2024 and issued this document as interpretive guidance on the amended rule. While people were saying the rules were underdeveloped, the documents were already being revised and stacked.

Finally, a note so that a similarly named document is not mixed in. Alongside the Research Ethics Guide to Generative AI for University Researchers covered here, the Ministry of Education line also has a separate Guideline on the Ethics of AI Use in Universities, which deals with teaching and learning. The National R&D commentary cites both in adjacent passages, which makes them easy to confuse; this article deals with the former.

How many pages a set of rules runs to depends in the end on how many acts that country has placed on its misconduct list. This is not a guess; lay the two texts open and it is immediately visible. Under 42 CFR Part 93, the US federal standard, research misconduct is three things: fabrication, falsification and plagiarism, known by their initials as FFP. And the section defining plagiarism carries a sentence rarely seen elsewhere.

The regulation names two things and carves them out. A fight over who deserves authorship is therefore not misconduct under the US federal standard. Which leaves nowhere to put the line "AI cannot be an author." The absence of an AI authorship clause anywhere in the ORI guidance is not an omission; it is the shape of the regulation the document stands on. The two Korean documents do the opposite and put that clause up front. The first line of the University Guide's Ten Precepts for Responsible Research Conduct is "do not list AI as an author," and the National R&D Guide's basic principles include the statement that "AI cannot bear legal or moral obligations and therefore cannot hold the status or rights of an author or inventor."

The Korean list is longer. Article 11(1) of the Ministry of Education directive, Guidelines for Securing Research Ethics, sets out seven subparagraphs: fabrication, falsification, plagiarism, improper attribution of authorship, improper duplicate publication, obstruction of an investigation, and, seventh, a catch-all.

The two things the US regulation carves out stand in Korea as independent categories of their own, and paragraph 2 leaves the door open for universities to add more on top. More acts to cover means more to write.

2.1Korea Has Two Lines of Its Own, and Their Scope Differs

Here a further split is needed, because lumping the Korean system into one gets it wrong. The University Guide stands on the Academic Promotion Act and the Ministry of Education directive; the National R&D Guide stands on the National R&D Innovation Act and its enforcement decree. The two lines differ from the definition of misconduct onward. Put all three strands in one table and the differences surface.

The statutory text was checked directly against the originals at the Korean Law Information Center (as of 23 September 2026). An amendment raising the Innovation Act's sanctions to 20 years' exclusion and a 30-fold surcharge takes effect in March 2027; the values above are the ones currently in force.

The row most worth studying is self-plagiarism. Three documents define the same word in three ways. The US regulation writes into its text that plagiarism does not include self-plagiarism. Article 56(2), subparagraph 3 of the Innovation Act's enforcement decree pulls self-plagiarism inside plagiarism, defining it as "using, without proper indication of the source, R&D data or R&D results of the researcher's own or another person, other than general knowledge, in the researcher's own R&D data or R&D results." The Ministry of Education directive handles it differently again, splitting it off from plagiarism as a separate category called improper duplicate publication. One act, three forks.

The layers are easy to confuse as well. Both Korean statutes enumerate only four items. Article 15(1), subparagraph 1 of the Academic Promotion Act and Article 31(1), subparagraph 1 of the Innovation Act both stop at fabrication, falsification, plagiarism and improper attribution of authorship. The count rises to seven at the level of the Ministry of Education directive, which is an administrative rule. So "Korean law defines seven types of research misconduct" is not accurate. The statutes enumerate four; the Ministry of Education directive raised it to seven.

2.2The Difference in the Text Shows Up in the Findings

That this is more than a paper distinction can be checked against actual findings. Here is the share held by the two categories the US regulation carves out — improper attribution of authorship and improper duplicate publication — in Korea's confirmed misconduct findings, by year of attribution.

These figures belong to the National Research Foundation of Korea's survey series on research ethics at universities, and this article confirmed them through the numbers as carried in press reports. They cover the Ministry of Education and NRF remit, that is, the university side. No separate findings statistics under the National R&D Innovation Act could be confirmed.

Taken across all four years, the share swings between 39% and 73%. Because it moves so much year to year, picking a single year lets you draw whatever picture you want. The range itself is the point. Apply the US federal standard to Korea as it stands and roughly half of the findings in a given year fall off the misconduct list entirely. That means the two countries are counting different things with the same word, and it has to be known before their rulebooks are compared by thickness.

So the summary "the US is lax" does not hold. 42 CFR Part 93 is a regulation whose breach draws federal administrative action, and the ORI guidance interprets that regulation. Narrow in scope is not the same as loose. The US hung few items and hung them hard; Korea hung many items and hung them wide.

The scope differs this much, and yet on the final test of judgment the two systems overlap. Asked what an act is ultimately held up against to decide whether it crossed the line, both give the same answer: whatever counts as practice in that field.

Start with Korea. After Article 11, which fixes the scope of misconduct, the Ministry of Education directive puts Article 12, which fixes how misconduct is to be judged. All three of its subparagraphs point to the same place.

So does the US text. 42 CFR §93.103(a) makes the first requirement for a finding "a significant departure from accepted practices of the relevant research community," and §93.200 defines those practices in three layers: practices established by 42 CFR Part 93 itself, practices established by Public Health Service funding components, and "commonly accepted professional codes or norms within the overarching community of researchers and institutions that apply for and receive PHS awards." The first two layers name who writes them. Only the third does not. The ORI guidance carries this over to AI matters as follows.

Yet the duty to actually carry out the judgment falls, on both sides, not to the disciplinary community but to the home institution. In Korea, Article 16(1) of the Ministry of Education directive states that "responsibility for verifying research misconduct lies with the institution to which the researcher belonged at the time the research was conducted." In the US, §93.307 and §93.310 assign inquiry and investigation procedures to the institution. Put the two lines in one table and the structure shows.

Read each row across and the subject of the left cell differs from the subject of the right. The body that sets the standard is the disciplinary community; the body that renders the finding is the home institution.

The two guides are aware of this spot. The University Guide divides the problems generative AI can cause at the research-design stage into three, and the third is "ignoring the particular character of the academic discipline" — AI may propose a generic method that fails to reflect the research practice, ethical standards and data characteristics of the field. But the remedy attached to it is review by the supervising professor, co-researchers, a statistician and the IRB officer. It never asks where the practice of the field is written down; it assumes someone who knows it is standing nearby.

3.1The National R&D Line Does Not Even Delegate That

The Korean row above applies only to the Ministry of Education directive line. The National R&D Innovation Act, which the National R&D Guide stands on, contains no delegation to disciplinary practice at all. Read the Act and its decree through and neither "academic discipline" nor "practice" nor "ordinarily tolerated" appears even once. The Innovation Act enumerates, and even the slot at Article 31(1), subparagraph 8, where a catch-all might have gone, is closed by four items the decree lists out: retaliation against a whistle-blower, forgery of documents evidencing research expenditure, breach of the Bioethics Act, and breach of the Laboratory Safety Act.

That is as far as the facts go. That the enabling statute does not delegate to disciplinary practice is settled by reading the text. So what about the document sitting on top of that statute? Across the 54 pages of the guide and commentary, the word for "field" or "discipline" appears six times, and not one of the six carries the sense that the answer varies by field. They sit in an AlphaFold introduction, a glossary entry and an explanation of bias. What the document points to instead is a different window. It says journal policies on AI use differ and that authors must check and comply with the relevant policy before submission, and it reprints the policies of four bodies — Springer Nature, the Institute of Electrical and Electronics Engineers, SAGE and arXiv — in an example box. What follows is interpretation. Placing the final answer about what is permitted outside itself is the same move; only the direction of deferral changes, from the discipline to the publishing venue. And of those four, the IEEE is a scholarly society. The pattern of societies being cited as exemplary norm-writers while never being named as teachers starts here. (The counts are of the Korean source text, based on the PDF text layer, so lettering inside figures may not have been captured.)

3.2This Is Not a Korean Flaw

One layer can be added here. The absence of anyone to write the practice down is not a defect of the Korean system but a vacancy that recurs wherever a practice-deferring definition is used. The US was not unaware of the problem. During the comment process for the amendment of 42 CFR Part 93, several institutions recommended narrowing the definition of "accepted practices" as too broad, but ORI kept the broad definition on the grounds that the universe of research practice keeps expanding. The vagueness is designed, not neglected. And ORI's prescription is to convene, case by case, committee members with expertise in the relevant field, not to have the practice written down in peacetime. The verbatim quoted above says exactly that. The strength of that summons also varies by stage: an inquiry committee may include experts in the relevant research area, whereas an investigation committee must (§93.307(e)(2), §93.310(f)). The discipline enters the process after something has happened, in the handling of that single case.

Field observation points the same way. Degn (2020) observed that heads of department seldom consult their institution's formal policies when judging research ethics and lean instead on informal disciplinary practice, and noted that this mediating layer is a particularly absent perspective in the research integrity literature. Comparative studies of Nordic regulation point out the cost on the other side: leaving everything short of serious misconduct to the self-governance of the research community, in an effort to avoid legalism, produced confusion about what is actually permitted.

US practitioner commentary leaves the same question open. A commentary from the consultancy Ankura noted that the ORI guidance does not resolve a uniform disclosure format, which AI outputs to retain, a reliability threshold for AI detection tools, or how institutions are to assess discipline-specific practice. The last item is precisely the question of this section. No formal response document from US research-integrity practitioner bodies could be confirmed, so what can be said here stops at "US practitioner commentary also flagged questions the guidance left unresolved."

That the vacancy is universal is no comfort. It means nobody has filled it yet. And as the next section shows, over the same stretch the rulebooks kept stacking while the record on the ground did not grow.

The pace at which rules accumulate and the pace at which the ground changes move independently. A study the National R&D commentary itself pulls in, while explaining a cautionary case, puts a number on that distance. A Peking University team analysed the full text of some 164,000 scholarly papers and published the result in PNAS in 2026.

The denominator needs to be read precisely. The 75,172 are the subset of the roughly 164,000-paper sample published since 2023. It is not a share of the whole sample, and it must not be widened into "0.1% of all papers." The way disclosure was identified is automated as well. Regular expressions swept the methods and acknowledgment sections for tool names and action phrases, and the paragraphs caught were then fed to a language model to separate disclosures of AI use for writing from AI use as a research method. And the place disclosures actually get written turned out to be, for the most part, the acknowledgments rather than a dedicated section.

The same paper produces a second indicator alongside the disclosure rate: an estimate, from lexical statistics, of the proportion of AI-assisted writing. And the authors computed the ratio of the two values themselves.

What makes this figure matter is not its size but its provenance. It is not a derived value made by dividing numbers from two different studies; it is a value one paper produced over one sample. Three caveats have to be read with it, though.

• The AI-writing proportion is not a verdict on individual papers. The authors stated flatly that it is a population-level probabilistic estimate rather than a definitive judgment about any individual manuscript. Rendering it as "this many papers were written by AI" gets it wrong.
• There are two denominators. "About 70% of journals have adopted an AI policy" is computed over 5,114 journals; "0.1% of papers disclosed" is computed over 75,172 papers. They are shares of different sets and must not be mixed in one sentence.
• The authors themselves reject the conclusion that policy is useless. The paper states that the failure of policies to stem the spread of AI does not mean policies have no value, and that policies may be moving some researchers from use toward declaration. The disclosure rate itself did rise, from 0.01% in early 2023 to 0.43% in the first quarter of 2025.

A methodological rebuttal ran in the same journal: without accounting for the lag between when a policy takes effect and when a paper is published, policy failure cannot be demonstrated. The original authors also wrote in their limitations that their results do not necessarily mean policies are ineffective. So "the policies failed" is the claim of that paper's title, not the conclusion of this article.

4.1How Much Have Fabricated References Grown

While disclosure stayed flat, the indicator on the other side climbed. It is the Lancet correspondence the National R&D commentary cites in its cautionary case 2. The commentary carries it as follows: an audit of about 2.47 million papers published between 1 January 2023 and 18 February 2026 found the rate of fabricated references rising from 4 per 10,000 papers in 2023 to 57 per 10,000 in 2026, and most papers containing fabricated references (91.3%) had only one or two.

These numbers need careful handling, because secondary reports citing the same correspondence carry values from different series. Coverage outside Korea rendered it as a rise from 1 paper in 2,828 in 2023 to 1 in 277 over the first seven weeks of 2026; converted to a per-10,000 basis, the 2023 figure lines up closely with the commentary's, but the 2026 figure comes to roughly 36 rather than 57. Which of the two is wrong could not be established. The text of the original correspondence could not be checked directly because of the publisher's access restrictions. This article uses the series the commentary cites, in a reporting construction, and records here that a different series was reported alongside it.

On the bibliographic side, several things are settled. The title is Fabricated citations: an audit across 2·5 million biomedical papers, and the document type is correspondence rather than a research article. The commentary states this accurately, as "correspondence." A correction followed in July 2026, fixing the fact that the audited population was not all of PubMed but the PubMed Central open-access subset. This article therefore describes the population as "about 2.47 million PMC open-access papers." One more thing: the authors of the correspondence disclosed, in their conflict-of-interest statement, that they used Claude for code development and grammar and punctuation assistance in preparing the work. A correspondence auditing fabricated references disclosed its own AI use in the audit. Unlike the vacancy in section 3, disclosure worked properly in this instance.

4.2Most Researchers Are Already Using It

On usage, all three surveys report a surge. But the three draw on different samples and ask different questions, so their numbers must not be blended.

The three surveys differ in both sample and wording. Wiley's 84% comes from the broad definition "any aspect of their work," while the 62% comes from the narrow definition of direct use in research and publishing. The by-use percentages in the Elsevier survey (summarising results 61%, literature review 51% and so on) are computed over all 3,234 respondents.

There are estimates from traces rather than self-report. A 2025 study that analysed 15 million biomedical abstracts in PubMed using an excess-vocabulary method estimated that at least 13.5% of 2024 abstracts were written with help from a language model. It differs from the disclosure rate above in both population and definition, so the two cannot be divided into a ratio. That is why the gap in this section is taken as the forty-to-one figure produced inside a single paper, and this estimate sits beside it, confirming the direction.

One thing should be added. The National R&D commentary, in presenting as a cautionary case a paper that carried AI-generated references across without checking the originals, noted that the paper it used for the case had already been retracted. That paper was indeed published in November 2025 and retracted sixteen days later, and the grounds included AI-generated illustrations without context and false provenance for the training and validation datasets. The commentary handled it accurately, in other words. The ground is not failing to move because the documents are poor.

To sum up: the forms have been built. The National R&D Guide comes with an AI use disclosure form, and the University Guide even has a checklist item on preserving prompts and responses. The details of that form and of the sanction clauses are set out in Pebblous's earlier piece on the national R&D AI research ethics disclosure form. What is missing is not the form but whether filling it in is the practice of that community. And the seat that fixes the practice is, as section 3 showed, empty.

Empty, yes — but not untouched. Section 2 of Chapter 5 of the University Guide reprints, clause by clause, the AI research-ethics rules of five Korean scholarly societies, and the headings of those clauses carry their amendment, entry-into-force and enactment dates as printed. Sort those dates and the order inverts.

These are the five the University Guide reprints as examples in Chapter 5. Not a complete list of Korean scholarly societies. Society names are English renderings of the Korean originals.

The earliest, the Korea Distribution Association's, was amended on 3 July 2025. That is about twelve months before the University Guide of 30 June 2026 and about fourteen months before the National R&D Guide. The two gaps must not be collapsed into one number; the value differs depending on whether the national document or the university document is taken as the reference point.

More important than writing first is what was written. The wording of the clauses is discipline-shaped. Article 6 of the Korean Society of Cinema Studies rules reads as follows.

The Korea Distribution Association runs the other way. It fixed not what must be written but what need not be.

The two clauses do different work. The cinema society enumerated the places AI is actually used in its own field; the distribution association drew the exemption line for disclosure itself and then supplied the wording of the declaration. The national documents do neither. Drawing an exemption line requires knowing what is already practice in that field, and no single exemption line fits every field at once. This is how the vacancy seen in section 3 actually gets filled.

5.1Five Is Not All of Them

Beyond the five the guide selected, ten more societies with rules posted on their websites were confirmed while preparing this article: the Korean Finance Association (한국재무학회), the Society of Korean Linguistics (한국어학회), the Korean Society of Management Engineers (한국경영공학회), the Association of Korean Medicine (대한한의학회), the James Joyce Society of Korea (한국제임스조이스학회), the Modern British and American Poetry Society of Korea (한국현대영미시학회), the Korean Association for Immigration Administration (한국이민행정학회), the Korean Association of Venture and Startups (한국벤처창업학회), the Modern English Drama Association of Korea (현대영미드라마학회), and the Korean Association of Regional Geographers (국토지리학회). That makes at least fifteen. The "at least" must not be dropped. The figure comes from Korean-language web search, not a systematic survey of the roughly 4,000 societies registered with the Korea Citation Index, and the sample tilts toward societies that publish their rules.

By field, the distribution is seven in the humanities, four in the social sciences, and one each in arts and physical education, engineering, medicine and pharmacy, and the natural sciences. The humanities take nearly half. The word "society" tends to conjure science and engineering, but this list does not, and its shape matches the fact that the humanities and social sciences hold a majority among registered journal-publishing bodies.

The most valuable thing in this list is not the number but the route. Of the newly found societies, three — the Modern British and American Poetry Society of Korea, the James Joyce Society of Korea and the Modern English Drama Association of Korea — stated themselves that they had referred to or adopted the rules of the English Language and Literature Association of Korea. The drama association recorded that it consulted both those rules and the revised edition of the National Research Foundation's recommendations. And the English Language and Literature Association of Korea is one of the five the University Guide selected. Writing a rule, in other words, is not an isolated decision by one society; it has a route along which it spreads sideways within the same scholarly community.

The newly confirmed clauses are discipline-shaped as well. The Korean Society of Management Engineers designated as research misconduct "uploading a manuscript under review to a generative AI service without the prior permission of the editor-in-chief, thereby breaching confidentiality," and the review rules of the Society of Korean Linguistics state that "a reviewer may not enter all or part of a paper under review into an artificial intelligence tool, or rely on generative AI to decide the content and outcome of the review." The Modern British and American Poetry Society of Korea asks that, where use is unavoidable, "a chat log or a web link evidencing the record of use be indicated in the citation," and the Korean Association for Immigration Administration drew an exemption line as the Korea Distribution Association did, stating that "this shall not apply to use for linguistic correction such as spell-checking and grammar correction." Sanction clauses, however, were confirmed at only two of them. That means confirmation did not reach the others, not that the others have no sanctions.

5.2Yet Neither National Document Calls on Societies to Teach

Counting words across the full text of the two national documents makes the drop sharp. In the 54 pages of the National R&D Guide and commentary, neither the word for "scholarly society" nor the word for "academic organization" appears even once. In the 65 pages of the University Guide, the word for "scholarly society" appears five times, all of them in Chapter 5, the submission-policy citations seen above. And what comes next is more striking. In 65 pages of a guide for university researchers, neither the word for "education" nor the word for "training" appears at all. The document walks an individual researcher through practical procedure; it never takes up who is to teach, or how.

The document that does speak about teaching is the National R&D commentary. And the body named there as the provider of that teaching is, in every instance, one and the same: the R&D institution.

In all three passages the supplier is the R&D institution. The organizations that have already written the practice into clauses are the scholarly societies, and yet a society never once enters the seat of the body that teaches. Two caveats attach to this word count. First, these are counts of the Korean source text, based on the text layer extracted from the PDFs. The University Guide is an image-heavy layout, so lettering inside figures may not have been captured; what can be said stops at "the body text of this document does not contain the word." Second, "zero mentions of scholarly societies" is a value confined to the National R&D documents. The University Guide has five, but only in places that are not about who teaches.

5.3The European Union Document Has the Same Hole

Open the third version of the European Union's Living Guidelines on the Responsible Use of Generative AI in Research, issued in May 2026, and how the seat for societies goes empty is clearer still. The document splits its recommendations into three parts only: researchers, research organisations, and research funding organisations. The phrase "learned societies" appears exactly once in the whole document, and that once is in the introduction's list of stakeholders.

It goes as far as engage in the discussion. There is no item telling them to do anything. The three addressees, by contrast, each get an item about training. Researchers get "continuously learn how to use generative AI tools properly to maximise their benefits, including by undertaking training"; research organisations get "provide and/or facilitate training for all career levels and disciplines"; research funding organisations get "promote and fund training and educational programmes for an ethical and responsible use of AI in scientific research." Two things stand out here. The "all disciplines" in the research-organisation item does not mean handing the work to a different body per discipline; it means the organisation supplies it across disciplines. That is the same structure as the National R&D commentary pushing training onto the R&D institution. And of the three addressees, only the funding organisation gets the verb "fund."

So reading this as "the European Union is entrusting norm production to learned societies" gets it exactly backwards. More precisely, the European Union has not filled this seat either. What is interesting is that the same document cites, as reference material, the generative AI principles issued by the Association for Computing Machinery. Societies get summoned as a source and stay off the list of addressees.

There is one more layer. This document, too, keeps the standard for disclosure outside itself. Recommendation 2 for researchers says that when generative AI meaningfully shapes results, researchers should note its use in the methods section "according to the guidelines of their journal or standards in their discipline." Which means the texts of all three jurisdictions stop in the same place: Korea's directive at "the range ordinarily tolerated in each academic discipline," the US regulation at "the accepted practices of the relevant research community," the European Union at "standards in their discipline." All three hand the final yardstick to the discipline, and all three leave unwritten who makes it.

The United Kingdom, Australia and Canada were checked too, and no system formally delegating norm production or training to scholarly societies could be confirmed. The UK's concordat to support research integrity places academies and societies in a recognition category of "other organisations," and the UK Research Integrity Office, an independent advisory body, runs on subscriptions that include those from societies. Societies are on the paying side, not the delegated side. Australia's code for the responsible conduct of research is co-owned by the National Health and Medical Research Council, the Australian Research Council and the universities' peak body, with the academies listed only as consultees on revisions. Under Canada's tri-agency framework, the duty to educate and raise awareness is addressed to universities. In none of the four jurisdictions is a scholarly society anything more than a consultee or a contributor of funds; none is a delegated implementer.

The question that set this off was whether scholarly societies should take the centre of research ethics training. The reasoning was that research ethics is the ethics of concrete research activity, and that a university, an organization where dozens of disciplines are split among a small staff, is structurally ill-suited to handling discipline-specific issues. The usual first reaction to the proposal is that it would require changing the law. Open the statutes and it does not.

6.1The Law Already Names Academic Organizations

All six provisions below are in force. They were checked directly against the originals at the Korean Law Information Center.

Provisions in force under the Academic Promotion Act and Ministry of Education Directive No. 449 (confirmed 23 September 2026).

Two more provisions can be added. Societies are already inside the verification procedure.

That is where the facts end. The designation mechanism, the basis for entrustment and contributions, the duty to run training, the power to take part in verification — all of it is already in provisions in force. Three things are worth stating carefully. First, all of the provisions above belong to the Academic Promotion Act and Ministry of Education directive line. The Innovation Act line, on which the National R&D Guide stands, has no academic-organization provision; in the Innovation Act and its decree, the word for "scholarly society" appears only as a training topic, in the phrase "ethics concerning scholarly exchange such as journal submission and conference attendance." Second, that Article 6(4) attaches to academic organizations is a reading of the text, and whether societies actually discharge that duty, or whether the Ministry of Education oversees it, could not be confirmed. Third, what the "designated training body" of Article 29 is currently designated to be is not in the text.

What follows is interpretation. Set the provisions side by side like this and the proposal looks less like a legislative one than an implementation one. Every mechanism needed is there, and two things are missing: the designation was never actually made, and there is no budget line attached. Put that together with the word counts of the previous section and one more thing appears. The enabling statutes pin academic organizations down as one pillar of "universities and the like," and the two 2026 AI guides sitting on top of them never once call on a society to teach. The guides are narrower than their own legal foundation.

6.2How the Roles Would Divide

This is not an argument that societies should take on everything. The structure seen in section 3 becomes the dividing line as it is. Procedure, records and verification already sit with the home institution; what is empty is writing the practice of the field down and teaching it through cases. This division is not one the author drew arbitrarily. The Committee on Publication Ethics (COPE), held up as the most successful case of society-led norm production, draws the same line on itself.

Whether societies are capable of writing norms is not in dispute. COPE, the Association for Computing Machinery, the Institute of Electrical and Electronics Engineers, the American Psychological Association, the American Medical Association, the International Committee of Medical Journal Editors and the American Chemical Society have all written the norms of their own fields. Among them, the one that offers a design hint is the American Chemical Society's approach. Its ethics committee does not build content from scratch; it selects and assembles research ethics material produced by the National Science Foundation and ORI into the language of its own field. The burden of a society having to make everything anew does not appear in the actual precedents. ORI, for its part, does not only put out material. The last page of the guidance directs institutional research integrity officers to contact ORI for technical assistance and mentions the RIO Boot Camp that ORI sponsors periodically. Writing the norms and training the people who will apply them are separated, and in the US one pillar of the latter is carried by the regulator itself.

Funding models differ by organization. COPE runs on membership, with the fees paid mainly by publishers. In the International Committee of Medical Journal Editors, participating organizations each bear their own share. The training-funding structures of other societies could not be confirmed, so no figures are given for scale. That the European Union guidelines attach the verb "fund" only to research funding organisations is, however, one reference point for this question. At minimum the funder's share has been put in writing, and the society's share is not yet in any document.

The training actually being supplied today has to be looked at alongside this. The online training platform run by KIRD, the government institute for science and technology human-resource development, carries thirty-six research ethics courses, but no evidence was found of separate tracks by discipline or field. Figures on how many people complete them could not be obtained. Which closes the picture like this. The law made training mandatory and set up a designation mechanism, and the societies have already written discipline-specific rules. Yet there is nowhere for those rules to cross over into a training curriculum, and what is actually supplied is a general course with no disciplinary division.

6.3What This Proposal Has Not Proven

Four things have to be discounted here. If this section carries any weight, it will be because these four were written down.

First, the proposition that teaching by discipline is more effective is not supported by the best available evidence. A meta-analysis testing eleven hypotheses across thirty studies and seventy-five effect sizes from 1990 to 2020 rejected the hypothesis that single-discipline groups outperform mixed groups. If anything, mixed groups may be better for some learning outcomes. What that study supported was not disciplinary homogeneity but individualization, discussion and applied practice — experience-based design. So the strength of societies does not lie in the bare fact that the discipline is shared. It lies in being able to use the actual cases and practices of that field as material for practice. That is to say, they are well placed to supply the factors the meta-analysis did support, and without keeping that distinction this section becomes an unsupported claim. For what it is worth, the weakness of generic online modules is well documented. The US National Academies' 2017 report stated that "the evidence for its effectiveness is limited" and that relying on classroom or online instruction as the primary means of addressing research misconduct may not be sufficient.

Second, making people read a code does not change behaviour. In a controlled experiment with 63 students majoring in the field and 105 professional developers, exposure to the ACM code of ethics made no significant difference to ethical judgment. This must not be stretched into an argument that training is useless. It is a narrow result about reading a code being insufficient, and it points in the same direction as the practice- and case-centred design the meta-analysis above described.

Third, societies are not free of funding pressure either. The American Medical Association's Council on Ethical and Judicial Affairs has itself treated the conflicts of interest in industry-funded continuing education as something to guard against. Moving the teaching body does not make the funding problem disappear. Fourth, conflicts of interest in journal governance remain. A 2026 study auditing 788 papers and 946 author-editor relationships from two top management journals between 2010 and 2024 reported that 49.7% of them were definite conflicts of interest. That study audited elite management journals in general rather than society-owned journals, so narrowing it to "a problem with society journals" mismatches the sample. The problem lies with journal governance broadly, and society-owned journals are not exempt.

Last, there is the matter of precedent. As the previous section confirmed, no design in which a scholarly society is formally delegated and runs research ethics training for a given discipline exists in the European Union, the United Kingdom, Australia or Canada. That can be read as "so the proposal overreaches" or as "so the seat is one nobody has filled yet." The author reads it the second way. That no other country has a precedent and yet Korean law already has the mechanism is where this article arrives. It is not that it cannot be done for want of a mechanism; it is that the mechanism in hand is going unused. This sentence is the author's judgment, not a fact.

And the other question that set this off — whether an expert ought to be able to tell what may and may not be done without a guide — hangs here too. This is not an argument for abolishing the rules. The National R&D Guide states of itself that "this guide is not intended to regulate the use of AI," and the University Guide positions itself as material to "refer to in establishing and operating a generative AI use guide suited to each university's research environment." The US document takes the same position. After writing that technological advances such as generative AI hold great promise for research, ORI adds that "neither the updated regulation nor this guidance stymies that research." All three documents are saying that their own position is not the final standard. Rules reach this far, and the layer above is set by practice. Growing people who can judge at that layer is what training is, and the material for that training sits inside the discipline.

Pebblous holds on to this subject not because it is in the research ethics business. It is because a fork we keep meeting in AI-Ready Data work has the same structure as this article.

7.1Defects Rules Can Catch, Defects Only Practice Can Catch

In data quality diagnosis, schema violations, missing values, duplicates and format errors are caught by rules. But whether this label means anything on this production line, and whether this measurement holds on this site, can be judged only by someone who knows the practice of that domain. The first ends with documents and automated checks; the second does not end however many check items are added. That the research-ethics rules of both countries defer the final ground of a finding to whatever counts as practice in the field is the same kind of limit. Raising the number of check items from twenty to forty does not get past this layer.

The society clauses in section 5 make the contrast concrete. The Korea Distribution Association could draw the line at "AI-assisted editing need not be separately disclosed" because it knows what is already practice in that field. A national document cannot draw that line, because no single exemption line fits every field at once. Drawing the line at "this missing value is normal" in a data quality rule is the same kind of act. That line is set by the domain, not by the rule engine.

7.2What Fills In a Lineage Is Practice

The fields of the AI use disclosure form attached to the National R&D Guide have the same shape as data lineage metadata. Which tool, at which version, over what period, within what scope. And the disclosure rate seen in section 4 was about 0.1%. That number is the distance between building a form and a lineage actually being recorded. The problem Pebblous runs into when handling training data lineage has the same shape. The columns are already there. The hand that fills them in has not yet entered the working day.

7.3The Question That Comes Back to the Organization

Companies and institutions carrying out national R&D now have to disclose their AI use and be able to explain the process. In practical terms that is a question of whether "which model did what, at which stage" can be reconstructed after the fact. If the pipeline keeps no such record, the disclosure form becomes paperwork filled in from memory. So it narrows to a single question. Does our research and development pipeline leave that record automatically, or does a person reconstruct it just before submission?

The ORI guidance seen in section 1 offers one piece of material for this question. Noting that experimental research often produces a forensic evidence trail, it gives as examples purchasing records, intermediate files related to figure processing, and physical access logs for animal facilities. Evidence of this kind is used to distinguish experimental research from synthetic or AI-enabled research, and it may support or weaken a conclusion that the research occurred as described. What is worth noticing is that not one item on that list is something a researcher wrote down on purpose. All of it accumulated by itself as the work ran. For a disclosure form not to be filled in from memory, that kind of by-product has to be left behind on the pipeline side.

The facts in this article come from four strands. The body text and verbatim provisions of the three countries' guides were checked directly against the downloaded source PDFs and the Korean Law Information Center; the disclosure rate and fabricated-reference figures come from the originals of the studies the National R&D commentary cites, or from the wording of the commentary itself; the findings on training effectiveness and society precedents come from the scholarly literature; and the society rules come from the requotations reprinted in Chapter 5 of the University Guide. Wherever something is a requotation, the body text says so.

Primary Normative Documents and Statutes

• 1.Ministry of Education, National Research Foundation of Korea, Council for University Research Ethics. Research Ethics Guide to Generative AI for University Researchers. National Research Foundation of Korea, 30 June 2026, 65 pp. Permanently posted at the research ethics information portal, CRE publications > guides. The verbatim statements of addressee and purpose in section 1, the ten precepts in section 2, the "ignoring the particular character of the academic discipline" item and its remedy in section 3, and the five society rules and the 20-item checklist in section 5 come from here.
• 2.Ministry of Science and ICT, Korea Institute of S&T Evaluation and Planning. AI Research Ethics Guide for National R&D and commentary volume, September 2026, 54 pp. KISTEP notice. The basic principles and seven recommendations, the 10-item self-diagnosis table, the "field" word count and the publisher and journal policy example box in section 3, the disclosure-rate and fabricated-reference citations in section 4, and the verbatim on who supplies training in section 5 come from here.
• 3.Office of Research Integrity, U.S. DHHS. Guidance on Generative Artificial Intelligence, August 2026, 13 pp. The addressee verbatim and page allocation in section 1, the five passages and the research-records provisions in section 1.1, the disciplinary-practice provisions and the funding-agency policy passage in section 3, the staged committee composition in section 3.2, the boot camp and the self-limiting verbatim in section 6.2, and the forensic-trail paragraph in section 7.3 come from here. The posting date is 14 August 2026 per the ORI blog.
• 4.Ministry of Education Directive No. 449, Guidelines for Securing Research Ethics [in force 17 July 2023, wholly amended]. Articles 2, 6, 11, 12, 16, 18 and 29. The article numbers shifted in the wholesale amendment from Directive No. 263 of 2018, so citing Articles 12 and 13 of the old edition as they stood does not line up.
• 5.Academic Promotion Act, Art. 2(4) (definition of an academic organization), Art. 5(2)1 (entrustment and contributions), Art. 15 (securing research ethics). The basis for section 6.
• 6.National R&D Innovation Act, Arts. 31 and 32, and its Enforcement Decree, Arts. 56 and 58. The National R&D column of the table in section 2 comes from here. The strengthened sanctions (20 years' exclusion, a 30-fold surcharge) take effect in March 2027; the values in the body are those currently in force.
• 7.42 CFR Part 93, Public Health Service Policies on Research Misconduct (2024 amendment). §93.103(a), §93.200, §93.211, §93.212, §93.227(b), §93.307, §93.310.
• 8.European Commission, DG RTD. Living guidelines on the responsible use of generative AI in research (Third version), May 2026. ERA Forum Stakeholders' document. The verbatim in section 5.3, the training item for each of the three addressees, and the "guidelines of their journal or standards in their discipline" in researcher recommendation 2 come from here.
• 9.Ministry of Science and ICT, Korea Institute of S&T Evaluation and Planning. National R&D Research Ethics Handbook, established 2022 / revised November 2025. The basis for the publication history in section 1.
• 10.Springer Nature. Editorial Policies. One of the overseas publisher policies cited in Chapter 5 of the University Guide.

Scholarly Literature

• 11.He, Y. & Bu, Y. (2026). Academic journals' AI policies fail to curb the surge in AI-assisted academic writing. PNAS 123(9): e2526734123. The 0.1% and the forty-to-one, the location of disclosures, and the sentence rejecting the policy-is-useless reading in section 4 all come from the original paper. A methodological rebuttal taking issue with the lag ran in the same journal.
• 12.Topaz, M. et al. (2026). Fabricated citations: an audit across 2·5 million biomedical papers. The Lancet 407(10541): 1779–1781 (Correspondence). doi 10.1016/S0140-6736(26)00603-3. A correction of 18 July 2026 (Lancet 408(10551):218) established that the audited population was the PubMed Central open-access subset. The figures in the body follow the series cited by the National R&D commentary; the text of the original correspondence could not be checked directly.
• 13.Resnik, D.B. & Hosseini, M. (2026). Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Accountability in Research, published online 15 March 2026. A scholarly discussion that reached the same logic five months before the ORI guidance. Whether the ORI document cites this paper could not be confirmed.
• 14.Katsarov, J., Andorno, R., Krom, A., van den Hoven, M. (2022). Effective Strategies for Research Integrity Training — a Meta-analysis. Educational Psychology Review 34(2): 935–955. The first caveat in section 6.3 comes from here.
• 15.National Academies of Sciences, Engineering, and Medicine (2017). Fostering Integrity in Research, chapter assessing the effectiveness of research ethics education. The verbatim "the evidence for its effectiveness is limited" comes from here.
• 16.Degn, L. (2020). Integrating Integrity: The Organizational Translation of Policies on Research Integrity. Science and Engineering Ethics 26(6): 3167–3182. The observation about heads of department in section 3.2 comes from here. Confirmed at the level of the abstract only.
• 17.McNamara, A., Smith, J., Murphy-Hill, E. (2018). A controlled experiment testing the effect of exposure to the ACM code of ethics on ethical judgment. The basis for the second caveat in section 6.3.
• 18.Mindel, V. & Ciriello, R. (2026). An audit of author-editor relationships at two top management journals. Journal of Business Ethics. The basis for the fourth caveat in section 6.3; the audit covers elite management journals in general, not society-owned journals.
• 19.Kobak, D. et al. (2025). Science Advances. A study estimating the scale of language-model-assisted writing across 15 million PubMed abstracts using an excess-vocabulary method. The 13.5% in section 4.2 comes from here.
• 20.Ariyaratne, S. et al. (2023). Skeletal Radiology 52(9): 1755–1758. Two specialists assessed five musculoskeletal radiology papers written with ChatGPT and found that all five contained fabricated references. Empirical work that meets the hallucinated-citation provision in section 1.1.
• 21.Committee on Publication Ethics (COPE). The self-positioning document beginning "COPE is not a regulatory body." The verbatim in section 6.2 comes from here.
• 22.Darri L. Scalzo (Ankura Consulting). The Research Integrity Challenge Behind ORI's New AI Guidance, Mondaq, 27 August 2026. The practitioner commentary in section 3.2 comes from here.

Surveys and Statistics

• 23.Wiley (2025). ExplanAItions 2025: The Evolution of AI in Research. The 57%→84% and 45%→62% in the section 4.2 table come from here.
• 24.Elsevier (4 November 2025). Researcher of the Future — a Confidence in Research report. About 3,234 respondents across 113 countries. The 37%→58% and the by-use percentages come from here, and the denominator for the by-use percentages is all respondents.
• 25.National Research Foundation of Korea. Survey Report on Research Ethics at Universities (2022 and 2023 editions) series. The findings counts by year of attribution in section 2.2 are values from this series confirmed through press reports; the tables in the original reports could not be checked directly.
• 26.Korea Citation Index (KCI) statistics on institutions and journals. Used for the comparison of field distribution in section 5.1 (accessed 23 September 2026).

Society Rules

• 27.Korea Distribution Association (한국유통학회) Research Ethics Rules, Art. 13 (amended 3 July 2025); the Research Ethics Rules on the Use of Generative AI, Art. 3, of the Literature and Film Association (문학과영상학회), the English Language and Literature Association of Korea (한국영어영문학회) and the Korean Association for the History of Religions (한국종교학회); and the Korean Society of Cinema Studies (한국영화학회) Film Studies Research Ethics Rules, Art. 6 (amended 28 April 2026). All five are requoted from the clauses reprinted in Chapter 5 of the University Guide.
• 28.AI-related research ethics and review rules of the Korean Finance Association (한국재무학회), the Society of Korean Linguistics (한국어학회), the Korean Society of Management Engineers (한국경영공학회), the Association of Korean Medicine (대한한의학회), the James Joyce Society of Korea (한국제임스조이스학회), the Modern British and American Poetry Society of Korea (한국현대영미시학회), the Korean Association for Immigration Administration (한국이민행정학회), the Korean Association of Venture and Startups (한국벤처창업학회), the Modern English Drama Association of Korea (현대영미드라마학회) and the Korean Association of Regional Geographers (국토지리학회). Confirmed on each society's website; this is the range reached by search, not a systematic survey.