nanalLabs BlogUsing ChatGPT or another generative AI to tidy up a data table, polish observation notes, or summarize a reference is already routine for many researchers. Whether you can then paste that output straight into your research notebook is a separate question — a research notebook only carries legal evidentiary weight because it's presumed to be "a record of what I personally did and observed." On September 20, 2026, Korea's Ministry of Science and ICT gave the first official answer, finalizing a "National R&D AI Research Ethics Guide" through deliberation by the National Science and Technology Advisory Council. This post breaks the guide down into practical rules for research notebook entries.
Why the guide exists — "the AI is a tool, the researcher is accountable"
The guide's core principle fits in one sentence: in national R&D, AI's role is limited to that of a tool, and final responsibility and rights over the output belong to the researcher who used it. Whether AI wrote a sentence or built a table, the moment it lands in a research notebook, the researcher is the one accountable for whether it's true and accurate.
That principle exists because of generative AI's well-known hallucination problem, and courts were dealing with the same issue around the same time. As fake, AI-invented case citations kept showing up in legal filings, Korea's Judicial Research and Training Institute ran a session for judges on September 4, 2026, titled "Responding to Citations of Fabricated Statutes and Case Law Generated by AI." In one case, 39 out of 40 cited precedents in a filing turned out not to exist. Research notebooks face the same risk: a number or citation an AI summarizes plausibly may not actually exist or may be wrong, and copying it in without checking undermines the notebook's own reliability.
The 7 recommendations, applied to notebook entries
The guide lays out seven things to observe when using AI. Mapped onto actual research notebook practice, they come out to this:
| Recommendation | What it means for a notebook entry |
|---|---|
| Verify facts | Don't copy AI-summarized numbers or quotes in without checking them against the original data or source |
| Exercise independent judgment | Review and rework AI output against your own domain knowledge and logic, rather than accepting it as-is |
| Ensure transparency | Note which parts were AI-written and which model/version you used |
| Manage result reliability | Don't treat an AI-generated conclusion as the final verdict on an experiment — keep the underlying data alongside it |
| Follow institutional policy | Check that AI use doesn't conflict with your institution's or project's own research notebook rules (e.g. handwriting requirements, signature rules) |
| Secure sensitive data | Don't feed unpublished experimental data or project details into external AI services |
| Protect personal information | Don't input data containing subjects' or collaborators' personal information |
Evaluation and peer review get an extra rule: reviewers assessing a proposal or paper must not feed the material under review into an external AI, and researchers being evaluated must not use AI to game or distort the review criteria.
How to actually document "AI wrote this"
The guide doesn't stop at principles — it ships two ready-to-use forms, a researcher self-assessment checklist and an AI-use disclosure form. Applied to notebook entries, that becomes a checklist like this:
- [ ] Marked which pages or entries were AI-written
- [ ] Recorded the AI model and version used (e.g. "GPT-5, 2026-09
release")
- [ ] Specified the purpose and scope of the AI use — information
gathering, data inference, grammar/typo checking, and translation carry different levels of trust
- [ ] Kept the verification basis (original data, source documents) for
the AI output alongside it
- [ ] Confirmed the final judgment and signature are the researcher's own
— AI can't substitute for the recorder's signature
What happens if you skip this? The guide itself doesn't attach a penalty, but given that a research notebook is often the first piece of evidence examined in a research misconduct inquiry, the problem shows up elsewhere. If an AI-introduced error surfaces later without having been checked, "I didn't know" tends to land worse than "I didn't disclose when and how much AI I used."
Does this conflict with existing research notebook rules?
Korea's National R&D Project Research Notebook Guidelines have long required researchers to record entries "personally, at the time" (see how to write a research notebook). The AI guide doesn't contradict that — if anything, it reaffirms the original principle for the AI era: AI is a tool, the researcher remains the one observing and judging. On paper, that can mean printing the AI output, pasting it in, and writing "AI-assisted, GPT-5, verified" next to it by hand. In an electronic notebook, it's cleaner to log AI use as metadata on each entry and fix the record's own timestamp in a way a third party can verify independently — that gives you a clear answer later to "when was this entry written, and with what."
The guide's full text is available on the Ministry of Science and ICT and KISTEP (Korea Institute of S&T Evaluation and Planning) websites. If your institution or project already has its own AI-use policy, check with your managing or specialized institution on how it interacts with this new guide.