🖥️ Lecture slides — Session 19 (Fri Nov 13)
Short page, firm rules. How to be honest, on the record, about AI assistance in research — and what the field currently expects.
What publishers and societies require, in spirit¶
Policies differ in wording across journals and societies (AGU, Nature, Science, and most publishers issued or revised theirs between 2023 and 2025), but they converge on three points:
- Disclose the tools and their role. Not a badge (“AI was used”), but a statement of which tools did what: drafting text, writing code, literature search, figure generation. Undisclosed substantial use, when discovered, is treated the way undisclosed image manipulation is treated.
- AI is not an author. Authorship requires the ability to take responsibility for the work and approve the submitted version. Software cannot; therefore it cannot be an author, and “the model made that mistake” is not an available defense.
- Humans own correctness. Whatever produced the text, the code, or the numbers, the listed authors vouch for them. Disclosure explains how the work was done; it never dilutes responsibility for what it says.
Some journals add specifics: AI use in peer review is commonly restricted (confidential manuscripts must not be pasted into third-party services), and AI-generated images face tighter rules than AI-assisted text. Read the policy of the actual venue before you submit; this changes year to year.
The course disclosure format¶
Every submission in this course (homeworks, project milestones, the final report) includes a disclosure with three columns per entry: tool, task, what you verified.
## AI-use disclosure
| Tool | Task | What we verified |
|---|---|---|
| Claude Code | drafted data-cleaning module (src/clean.py) | unit tests written by us; checked gap handling against 3 stations by hand |
| Copilot | autocomplete throughout | all committed code reviewed line by line |
| ChatGPT | first draft of report introduction | rewrote; checked all 6 citations resolve and support the claims |The third column is the one that gets graded. “We verified nothing” is at least honest and will cost you less than a verification claim that collapses under one question. Substantial AI assistance is also disclosed where it happened: in commit messages and in the repository README, per the final project rubric (1.10).
Where this lives in a paper: CRediT¶
Most journals now use the CRediT taxonomy (Conceptualization, Methodology, Software, Validation, Writing, and so on) to state who did what. AI assistance does not get a CRediT line — contributor roles are for people. It goes in the acknowledgements or in a dedicated AI-use statement in the methods, and the human who directed and verified the AI-assisted work holds the corresponding CRediT role. If an agent wrote most of your processing code, the Software credit is yours, and so is the Software blame.
Disclosure beyond the journal¶
Journals are not the only party with a policy. If you work at a national laboratory, an agency, a company, or on sponsored research, three institutional questions come before any submission checklist:
- Check the employer and sponsor policy, not just the venue’s. Many institutions and funders now have their own AI-use rules — what may be used, for what, and what must be reported in progress reports and deliverables. The venue policy governs the paper; the institutional policy governs the work. Check both before using a tool, not while formatting the manuscript. A serviceable sponsor-facing statement is two sentences in the same shape as the course table: “AI tools (named) were used for (tasks); all outputs were verified by (method) and the authors take full responsibility for the results.”
- Classify the data before it touches a hosted model. Export-controlled, proprietary, pre-publication, personally identifiable, or license-restricted material must be screened before it is pasted into a prompt or read by an agent — a hosted API is a transfer to a third party, whatever the convenience feels like. This is the privacy row of the API-vs-open-weight table in 6.1 acting as policy: data that cannot leave your institution points to a self-hosted open-weight model or no model at all.
- Retain the agent transcripts. Git records what changed; an agent session transcript records what was asked, what was proposed, and what was rejected — the decision-level audit trail. For work that may face an audit, a milestone review, or a misconduct inquiry, keep the transcripts of substantial agent sessions alongside the repository, with the same retention period as your lab notebook. “The agent suggested it and we accepted” is a defensible statement only if the session showing the verification still exists.
One subsection is not a compliance program; it is the habit of asking “who else has rules about this?” before the tool runs.
How over-reliance actually fails¶
A group let an agent write their preprocessing and their results section in the last week of a project. The agent’s deduplication step silently dropped every event within 30 seconds of another event, which removed most aftershock sequences from the catalog; the reported detection recall was computed on the easy remainder and looked excellent. Nobody in the group could explain the dedup logic when asked, the number went into a poster, and the correction cost far more reputation than the original result was worth. Nothing in this story required malice — only unverified fluency and a deadline.
You may be examined on it¶
Disclosure has a companion: you must be able to defend, orally, any part of what you submit. This course reserves the right to run a verification interview on your final project — an examiner picks a processing choice, a line of code, or a reported number, and you explain it, without notes, whoever or whatever wrote it. Some research groups run the same check as a standing practice: before a paper is submitted, each author answers an oral quiz on the data-processing and interpretation choices they are listed as owning, and the answers are logged next to the manuscript as evidence that a human is in control of the workflow. (See also the presentation rubric in 1.10.) The standard is simple: if you cannot explain it, you cannot submit it. Plan your verification effort — the third column of the disclosure table — so that you would pass that interview on any line, any day.