Heidelberg AI Curriculum

T15-L01 · AI in research · Level 1 User · 12-minute read, plus 30 minutes of doing

The question on the submission form

You used an AI assistant to tighten the discussion section of your manuscript, and to write the script that draws Figure 3. Now the journal's submission form asks: "Was generative AI used in preparing this manuscript? If so, describe how." You remember roughly what you did. You don't remember which paragraphs, which tool, or what you changed afterwards.

Or you're writing the state-of-research section of a grant proposal. You gave an assistant the abstracts of twenty papers and asked which ones address your question. Three weeks later a co-author asks whether that needs to be declared, and whether the two papers you dropped were dropped by you or by the tool.

In both cases the use itself was probably fine. The problem is that you can no longer describe it, and a vague description ("AI was used for editing") is exactly what journals and funders are starting to reject. This book shows you where AI helps in research, where the lines are, and how to keep a record small enough that you'll actually keep it. You are at Level 1 User: the work and the responsibility are yours alone, which is the point.

After this you can

Use AI to triage a set of papers and tighten a paragraph while keeping every judgement yours.
Keep a use log as you work, instead of reconstructing it at submission.
Find what your journal or funder says about AI and sort your use into allowed, disclose, or don't.
Write a disclosure statement that names the tool, the task and your check.

Before you start

Books: T01-L01 for the four-quadrant method, and T01-L02 for checking output before it leaves your desk.
An assistant you're allowed to use for unpublished work. Many universities and institutes now provide one with a data-protection agreement. Use that. If you have none, a personal Claude or ChatGPT account is fine for published abstracts, but not for your unpublished manuscript or proposal (T12-L01).
PubMed, or the database you normally use, to export abstracts.
The AI policy of your target journal or funder, open in a browser tab.
About 45 minutes: 12 to read, about 30 to do.

Last verified 23 September 2026. Springer Nature's AI policy (which covers Nature Portfolio journals) sorts AI use by risk. Language polishing, structuring and data cleaning are permitted, and disclosure is encouraged. Drafting summaries, comparing results with the literature and extensive writing support are permitted with human oversight, verification and disclosure. Fabricating data or citations, listing AI as an author, delegating peer review to an LLM and creating photorealistic deepfake images are not permitted. Manuscripts, peer-review reports and sensitive data must not be shared with unsecured or public AI systems. The DFG (German Research Foundation, guidance updated 27 May 2026) asks applicants to disclose which generative models they used, for what purpose and to what extent, for example to summarise the state of research. Grammar, style and spelling checks and translation tools need not be disclosed. Since 16 April 2026, DFG reviewers may use AI, but only under the conditions of the DFG's guideline for reviewers; read it before you do. PubMed exports abstracts under Save → Format: Abstract (text) → Create file, up to 10,000 records.

The idea in one page

Research is the one place where how you got a result is part of the result. So the question isn't only "is the output right?" but also "could I describe exactly what the tool did?" Three zones cover most of what you'll meet:

Usually fine: language, code, triage. Tightening your own sentences, drafting a plotting script you then check, sorting a set of abstracts against criteria you wrote. The thinking stays yours, and you can check the output.
Contested: substantive text. Letting the assistant draft an interpretation, summarise the state of research or compare your results with the literature. Publishers allow it with oversight and disclosure, but it's where reviewers look hardest, and where invented claims creep in.
Not allowed: authorship and other people's work. AI is never an author, because it can't take responsibility. Fabricated citations are misconduct, whoever typed them. And a manuscript you're reviewing, or a proposal you're assessing, must never go into a public AI tool; some funders forbid AI in reviewing altogether.

The rule, whatever the venue: every AI use that touched the content is described specifically, in the place the venue asks for, and you checked everything it produced. Venues differ in the details. The DFG doesn't want to hear about spell checks. Springer Nature encourages you to disclose even language polishing. That's why you read the actual policy (the Springer Nature policy (opens in a new tab) is a clear example) rather than guess.

The use log. You can't write a specific statement from memory. Keep one line per use, as you go:

Date | Tool (and model, if shown) | Task | Input | What I kept | How I checked

Five fields and a date. It takes thirty seconds per use and saves an evening of guessing at submission.

Worked example: two research situations, four quadrants

The method comes from T01-L01. Here it is on a grant proposal, then on a manuscript.

Research situation 1: triaging twenty papers for a grant proposal

Structure it. "For the state-of-research section of my DFG proposal I have twenty candidate papers from a PubMed search. I need to know which ones directly address my question (does protein X regulate autophagy in neurons?) so I can read those in full first. A useful result: a reading order I can defend, with nothing dropped silently."

Understand it. She pastes a few titles into the assistant and asks which are relevant. It confidently adds a twenty-first paper that "also shows this". The paper doesn't exist. That's the tool recalling instead of working (T01-L01, fact 1). She also realises that "relevant" was her word, not a criterion. Back to Structure it: "Give it the abstracts, not titles. Define 'addresses my question' as three yes/no criteria. It sorts; I decide."

Choose tools. Going down the ladder: PubMed already exports all twenty abstracts in one text file. The institute's approved assistant sorts them, because this is going into an unpublished proposal. No reference manager plugin or search tool is needed for twenty papers. The final inclusion decisions and the reading stay hers.

Build it. She exports the abstracts from PubMed (Save → Abstract (text)), attaches the file and asks:

Attached: 20 PubMed abstracts, each with a PMID.
Use only these abstracts. Do not add papers or references.

For every PMID, answer three questions with yes / no / unclear, and quote
the sentence from the abstract that supports each answer:
A. Is protein X (or its gene) studied directly?
B. Is autophagy measured, not just mentioned?
C. Are the experiments in neurons or neuronal tissue?

Return a table with all 20 PMIDs, including those with "no" or "unclear".
Then list the PMIDs with three "yes" answers.

The table comes back with all twenty PMIDs. Seven have three yes answers. She checks the quoted sentences for the seven and for every "no": two quotes don't support the answer given, and one paper marked "no" for neurons actually used primary cortical cultures. She corrects those three by hand. One paper moves onto her list, so she reads eight in full. Her log line: "23 Sep · Institute assistant (model not shown) · Triage 20 abstracts against criteria A–C · PubMed export, PMIDs listed · Reading list of 8, my decisions · Checked every quote; corrected 3."

Her disclosure in the proposal: "We used [institute assistant] to pre-sort 20 abstracts against three stated criteria. All classifications were checked against the abstracts by the applicant, and all included papers were read in full." What it taught her: the tool saved an afternoon of scanning, but the checking found a paper she'd have dropped.

Research situation 2: a paragraph and a plotting script for a manuscript

Structure it. "My discussion section is 30 per cent too long, and Figure 3 needs redrawing from the raw data. The target is a Springer Nature journal. A useful result: a shorter paragraph that says exactly what the old one said, a correct figure, and a disclosure I can submit."

Understand it. He asks the approved assistant to shorten the key paragraph. The result is crisp, and it has turned "is associated with" into "drives". That's a new causal claim he never made. The plotting script it drafts runs, but on closer look it averages the three replicates before plotting, so the individual points he wanted are gone. Back to Structure it: "Shorten without changing any claim, and tell me every wording change that affects meaning. Plot every observation; no silent averaging."

Choose tools. Going down the ladder: the journal's author guidelines give the word limits and figure rules. The approved assistant does the language pass and drafts the script, because both are checkable. He runs the script on his own computer. The interpretation stays his.

Build it. He sends the paragraph with: "Cut to 120 words. Keep every claim, number, hedge and citation. Then list every change that could alter the meaning." It lists two. He rejects one ("suggests" → "shows") and keeps the other. For the figure he asks for a script that plots all individual values, labelled by replicate, and checks it on a small table where he knows the answer: four rows in, four points out. His log gets two lines, one per task. His statement, for the journal's AI declaration field: "The authors used [tool, model] to shorten the Discussion and to draft the plotting code for Figure 3. The authors reviewed all edits against the original text and verified the figure against the source data. The authors take full responsibility for the content."

What it taught him: "tighten" is where meaning changes hide. Next time he'll ask for the list of meaning-changing edits from the start.

In both situations the assistant saved real time, and none of the judgements moved to it. The log made the disclosure a five-minute job.

What goes wrong

Listing AI as an author. Symptom: a co-author suggests thanking "ChatGPT" in the author list or the CRediT roles. Fix: authors must be able to take responsibility, which a tool can't. Describe the use in the disclosure or Methods instead.

Pasting an unpublished manuscript into a personal account. Symptom: the full draft goes into a consumer chat to "polish the English". Fix: use your institution's approved assistant for unpublished work, or only paste material that is already public (T12-L01).

Relying on generated citations. Symptom: a reference that looks perfect and doesn't exist, or exists and says something else. Fix: never ask the assistant to supply references. Give it the papers, and check every citation against the original.

Disclosing vaguely. Symptom: "AI tools were used in the preparation of this manuscript." Fix: name the tool, the task, the part of the work, and how you checked. Your use log gives you all four.

Reconstructing the log at submission. Symptom: you can't remember whether the tool or you dropped a paper. Fix: one log line per use, written the moment you finish.

Assuming your co-authors agree. Symptom: at submission, a co-author objects to AI in the analysis. Fix: agree what's acceptable at the start of a project, and share the log before the draft goes round.

Using AI on work you are reviewing. Symptom: a manuscript or proposal you received as a reviewer goes into an assistant "just for a summary". Fix: don't. The material is confidential: Springer Nature forbids uploading it to public AI tools, and the DFG allows AI in reviews only under the conditions of its reviewer guideline, never by pasting a proposal into a public tool.

Do it yourself (30 minutes)

Pick one real piece of your work where you used, or plan to use, an AI assistant.

Structure it (5 min). Write down the target venue and what AI did or will do: which task, which part of the work.
Understand it (10 min). Open the venue's AI policy. Find the lines that apply to your use. Is it allowed, allowed with disclosure, or not allowed? Where does the venue want the statement (form field, Methods, acknowledgements)? If the policy doesn't say, write down whom you'll ask.
Choose tools (2 min). Decide where your use log lives: a text file next to the manuscript, or a table in your lab notebook.
Build it (13 min). Fill in the log for every AI use so far, from your chat history. Then write the disclosure statement: tool, task, part of the work, how you checked, and that you take responsibility.

Exit check

Written finding: one disclosure statement for a real piece of your work, with:

the use log behind it, one line per use;
the policy it was checked against: its name, link, the date you read it, and the sentence that applies;
the location where the venue wants it, or the person you'll ask.

It passes when a co-author can match every sentence of the statement to a log entry and to the policy.

Rule to remember

You are the author. It is a tool you must be able to describe.

Use AI in Your Research and Say Exactly How