South Korea's science ministry has issued a non-binding research ethics guide that keeps responsibility for AI-assisted national R&D output with the researcher and tells evaluators not to put material under review into external AI

Responsibility stays with the researcher

The guide's purpose statement in Part I reads "본 가이드는 AI 활용을 규제하기 위한 것이 아니며, 연구개발 과정에서 연구자의 윤리적인 AI 활용을 지원하기 위해 작성". Our translation: this guide is not intended to regulate AI use, and was written to support researchers' ethical use of AI in the R&D process. Everything described below is recommendation, not rule.

Within that frame the basic principle is short. The box in Part II reads "국가연구개발에서 AI의 역할은 도구로 제한되며, 최종 결과물에 대한 책임과 권리는 AI를 활용한 연구자에게 귀속됨". Our translation: in national R&D the role of AI is limited to that of a tool, and responsibility and rights for the final output belong to the researcher who used it.

The guide adds, as a principle and not as a statement of law, that AI cannot hold the status or rights of an author or inventor because it cannot bear legal or moral obligations. Whether Korean patent or copyright law already says the same is a question the guide does not answer, and neither do we.

Its stated audience is every researcher using AI in national R&D, with some roles set out for R&D institutions. The recommendations take generative AI and large language models as the baseline, and the guide says they extend to self-built models and agentic AI where applicable.

Recommendation 3 and the disclosure form

Of the seven recommendations, the third is the one most likely to reach a research office's procedures if an institution adopts the guide. That is our reading. The text in Part III is "(투명성 확보) 연구자는 연구 과정 전반에서 AI를 활용한 경우 모델·버전, 목적·범위* 등을 최종 성과물 제출(또는 발표) 시 투명하게 공개해야 함". Our translation: where a researcher has used AI anywhere in the research process, they should transparently disclose the model and version, and the purpose and scope of use, when the final output is submitted or presented. We render 해야 함 as "should" because the guide presents all seven items as recommendations (권고사항).

The asterisk explains purpose and scope by example: information gathering, inference from data, organising research data, grammar and typo checks, and translation. Grammar checking is on that list, so on the guide's terms the recommended disclosure is not reserved for substantive use.

Part IV supplies a sample disclosure form (AI 활용 여부 공개), marked as an example. It asks for the tool and model, a description of use and the period of use, and closes with a signed confirmation that final responsibility sits with the author. The worked example lists "ChatGPT, Claude(Opus 5)" as tools, which shows the level of version detail the drafters had in mind. A note says the form may be modified as needed, and the press release describes it as a form researchers can use when submitting R&D reports. Neither document makes it mandatory.

Evaluators, and the people they evaluate

The evaluation part of the guide covers national R&D project evaluation and paper review, and it is recommendation too. For evaluators, item 1 reads "평가자는 연구계획서, 결과보고서, 논문 등 평가 대상 자료를 외부 AI에 입력하지 않도록 해야 함". Our translation: evaluators should not enter material under evaluation, such as research plans, final reports and papers, into external AI. The reason given is leakage of security-sensitive and personal information.

As guidance, item 2 asks evaluators to assess on the basis of their own expertise and judgement, and says that accepting AI output without their own review undermines the purpose and fairness of expert evaluation. It does not forbid AI as an aid.

Item 3, still recommendation only, turns to the evaluatee. Those being evaluated should not use AI to evade legitimate review criteria or to distort evaluation, with hidden prompts (은닉 프롬프트) given as the example. The next line carries the strongest language in the document: "해당 행위는 평가의 신뢰성을 심각하게 저해하는 연구부정행위에 해당하며 관련 법령에 따라 엄격한 제재처분의 대상이 될 수 있음". Our translation: such conduct constitutes research misconduct that seriously undermines the reliability of evaluation, and may be subject to strict sanctions under the relevant statutes. The wording is conditional: a sanction may follow under other laws, which the guide does not name, and the guide itself imposes no penalty.

What it asks of research institutions

Under recommendation 6, which like the rest is guidance, institutions should provide institutional and technical security environments, and researchers on security-sensitive projects should use tools inside an environment the institution has approved. Recommendation 7 asks researchers to take care not to enter personal data into AI services, and to de-identify it first where its use cannot be avoided.

Part II separately says, as guidance, that institutions should build support systems, naming guidelines, rules and manuals, along with arrangements to prevent and respond to ethical risk. A ten-item self-check sheet is offered for researchers to run before starting work and again before submission. The guide also says that expert agencies such as the National Research Foundation of Korea (한국연구재단) already prohibit evaluators from uploading material to AI. That is the guide's description; we did not check the agencies' own rules.

Legal weight, and the release date

The guide says it is not intended to regulate AI use, and on our reading it carries no legal force of its own. It creates no duty and sets no deadline. Any force in practice would come from elsewhere: an institution writing it into internal rules, a funder adopting the disclosure form, or the existing statutes the misconduct line points to. That is our reading, not the ministry's.

The press release was embargoed to 12:00 KST on 20 September 2026 for 21 September morning papers, and distributed on 18 September. The ministry says the guide was finalised after the operating committee's deliberation and resolution but gives no date for that meeting, and the guide cover shows only 2026. 9. We use 20 September as the release date. It is a release embargo, not a documented approval date.

The ministry says it will send the guide and a commentary (해설서) to R&D institutions, post both online, and keep revising the guide as it collects field feedback.

What we did not verify

What we opened: the MSIT press release and the full guide attached to it (붙임), retrieved from korea.kr as an HWPX file and read as extracted text, including the purpose statement in Part I, the principles in Part II, the seven recommendations and the evaluation items in Part III, and both sample forms in Part IV. We downloaded the attachment again on 24 September 2026 (UTC) and its SHA-256 matched the copy on file. The korea.kr release page carries a publication timestamp of 20 September 2026, 12:00 KST.

What we did not open: the copies of the guide on the msit.go.kr and kistep.re.kr sites, the commentary, the statutes the misconduct line refers to, the National Research Foundation's evaluation rules, and any record of the operating committee's decision. We did not check the Wiley and Elsevier survey figures the guide cites against their sources, so we do not repeat them.

What we refuse to claim: we do not describe AI-use disclosure as mandatory, because the guide calls its items recommendations and the form an example. We do not say anyone has been sanctioned for hidden prompts, or name the statute that would apply. We do not state an approval date. We do not claim this is the first guide of its kind in Korea or anywhere else. Korean quotations are reproduced from the extracted text of the attachment, and every English rendering is our translation.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• University of Crete doctoral rules on AI use and disclosure
• US Education Department accreditation proposal on research misconduct and AI
• South Korea's MSIT and KISA AI red-teaming guide
• Asia-Pacific AI regulation tracker
• AI Regulation News hub