AI in your research workflow

This mini-book is for Level 1 User. AI may help with your own reading, writing, and code, but it cannot accept authorship responsibility or turn an unchecked result into research evidence. You remain accountable for what you submit.

2. The submission form asks what AI did

You are preparing a manuscript or technical research report. Last week, you asked an approved AI assistant to rank twenty public papers for closer reading. Yesterday, it tightened a paragraph in your discussion. Today, it produced a short script for a plot. You checked the paragraph against your results and ran the script, so the work feels finished.

Then the submission form asks whether generative AI was used, where it was used, and how its output was checked. Your notes say only "used AI for editing." You cannot remember the service shown in the interface, which papers were triaged, or how much of the plotting script survived. A co-author thinks language help needs no declaration; your target journal may say otherwise.

The problem is not merely finding better wording for the form. You need a workflow that keeps AI assistance visible while you work, preserves the route back to every cited source, and leaves scientific judgement with the human authors.

3. After this you can

• Classify a reading, drafting, or coding task as assistive, substantive, or not permitted under the applicable policy.
• Verify every proposed citation and factual claim in the original public source.
• Record what an AI tool did in a short use log while the work is in progress.
• Check generated or edited code against known inputs and expected results.
• Write a specific disclosure statement that matches a target venue's current instructions.

4. Prerequisites

• T01-L02 · Trust but verify.
• The current author instructions for one target journal or publication venue, plus your institution's AI and research-integrity policy.
• Twenty public bibliographic records or twenty synthetic research records, one paragraph you are permitted to edit, and a small public or synthetic dataset.
• An institutionally approved AI assistant, or a colleague who can play the assistant role from the prompts on this page.
• A text file or spreadsheet for a use log.

Do not paste unpublished manuscripts, peer-review submissions, participant or patient data, confidential company research, credentials, or licensed full text into a consumer service. Approval depends on the service, account, contract, purpose, and data. If any one of those is unclear, use synthetic material and ask the responsible person.

5. The idea in one page

Treat every AI-assisted research step as three connected decisions: may I use it, how will I check it, and what must I record or disclose? A task is not automatically acceptable because it is called editing, coding, or search.

Start with the applicable rules

Read the target journal's author instructions and submission form, then your institution, funder, collaboration, ethics, and data rules. Record the page URL and access date. If they conflict, do not average them: follow the stricter applicable requirement or ask the journal and your research-integrity contact. Agree the boundary with co-authors before AI touches shared work.

Current-policy check - last verified 4 September 2026. Policies change, and individual journals may add requirements. ICMJE says authors should disclose whether and how AI-assisted technologies were used, must not list them as authors, and remain responsible for submitted material. Elsevier's journal policy distinguishes basic spelling and grammar checks from substantive changes, asks for a manuscript declaration for reportable preparation uses, and asks authors to describe AI used in research methods in the Methods section. Nature Portfolio uses a risk framework that separates assistive language or organisation from interpretive use and from prohibited replacement of scholarly judgement. These are examples, not a universal rule. Open your target venue's current page at submission.

Separate assistance from evidence

For reading, AI can help label or order a fixed set of records. Its ranking is a triage suggestion, not a search strategy, inclusion decision, or literature finding. Preserve the original query and record set. Open the original paper before relying on a claim. A title, DOI, or quotation produced by the assistant is only a lead until you locate it in an authoritative index and then inspect the source itself.

For drafting, a language pass can improve clarity while preserving your meaning. Stop when it introduces an interpretation, causal claim, limitation, comparison, or citation that you did not establish. Substantive drafting is contested and may be permitted only with disclosure and demonstrable human judgement. The final sentence must be one you understand, can support, and deliberately accept.

For coding, generated code is an untested draft. Read it, run it only in a safe environment, compare its inputs and outputs with known cases, and retain the final script beside the data and figure. AI assistance does not lower the validity or reproducibility standard for code.

Some boundaries are firm: an AI tool cannot be an author because it cannot approve the work or answer for its integrity. Do not delegate peer-review judgement to AI. A submitted manuscript from someone else is confidential; follow the venue's reviewer policy and do not upload it where confidentiality is not assured.

Keep the log smaller than the regret

Add one row immediately after each use: date; displayed tool or model name; purpose; input description and data class; saved prompt/output location; what you checked; what entered the work; and likely disclosure location. This supports an accurate disclosure and an audit trail. It does not prove that another person can recreate the output: services and model routes can change, and the same prompt may return different text. Exact versions, parameters, input hashes, reruns, and reported differences belong to T15-L02 · Reproducibility with AI in the loop.

6. The worked example: one workflow, two research settings

Both versions use the same sequence. The Lab version supports a manuscript about a fictional materials experiment. The Company version supports a public evidence report about a fictional workplace programme. No confidential, personal, or unpublished source material enters the assistant.

Lab framing: twenty papers and a synthetic experiment

Mira first saves the target journal's AI policy URL, access date, required disclosure location, and Methods requirements. She exports twenty public records from PubMed using her documented search, labels them L01 to L20, and keeps the search string and export unchanged.

She supplies only permitted titles and public abstracts with this prompt:

Triage the 20 supplied records against this question: [question].
Return record ID, possibly relevant/not clearly relevant, one reason quoted from
the supplied abstract, and needs-full-text-check yes/no. Do not create references,
fill missing methods, or decide inclusion. Write "not stated" when needed.

The assistant marks six records possibly relevant. Mira opens all six originals, confirms their identifiers, and makes the inclusion decisions herself. Two are excluded because the full method differs from what the abstract alone suggested. No assistant-generated citation enters the manuscript.

Next, she submits one paragraph that she wrote from the synthetic result. Her instruction is limited: "Improve readability without adding or removing claims, numbers, uncertainty, or citations. List every meaning-changing proposal separately." She compares the revision sentence by sentence with her result table and source notes. She rejects a new causal verb and accepts two shorter sentences.

Finally, she asks for a plotting-script draft using this synthetic table:

condition,replicate,value
control,1,10
control,2,12
treatment,1,14
treatment,2,16

She reads the script, confirms that it plots all four observations rather than invented averages, runs it in a clean folder, and compares the displayed values with the CSV. She saves the final code, input, output, and software environment. That check supports this run; it is not evidence that an unspecified future environment will reproduce it.

Her use log now reads:

Mira drafts the disclosure from the journal's requested fields, naming the actual approved service shown in her records rather than writing "AI was used." She sends the statement and log to her co-authors for agreement; they remain the authors and approve the final manuscript.

Company framing: twenty sources and a public evidence report

Jonas's R&D team is preparing a technical report for external publication. He saves the venue's AI instructions and his company's research-integrity and information-handling rules. He gathers twenty public reports and standards records from their issuing organisations, labels them C01 to C20, and freezes the source list.

He uses the same triage prompt, replacing "full-text check" with "issuing-source check." The assistant marks five records possibly relevant. Jonas opens the five documents on the issuers' sites, checks dates and versions, and makes the relevance decisions. One record describes a superseded edition, so he removes it. The assistant's summary is never cited.

For drafting, Jonas supplies one paragraph he wrote from public evidence and uses the same language-only instruction. He accepts a clearer sentence order but rejects a generated claim that the programme "improves productivity," because his sources support reported experiences rather than a causal effect.

For coding, he uses the same four-row synthetic CSV with pilot and comparison labels. He reads and runs the chart script, checks all plotted values against the file, and saves the final inputs, code, output, and environment. He does not connect the script to a confidential analytics system.

His three log rows mirror Mira's: source triage with original-document checks, a language pass with a rejected causal claim, and a plotting draft with a known-input test. His venue asks for one AI statement in the report and methodological detail for code assistance. He names each purpose, states the human checks, and gets all human report authors to approve the wording before submission.

The important parallel is not the subject matter. In both settings, AI narrows attention, proposes wording, and drafts code; a human opens sources, owns interpretations, tests execution, keeps the record, and follows the actual venue policy.

7. What goes wrong

AI is listed as an author

Symptom: the byline or contributor list names a chatbot because it produced substantial text.

Fix: remove it from authorship. Human authors must meet the venue's authorship criteria, approve the final work, disclose assistance as required, and accept responsibility.

A confidential manuscript enters a consumer tool

Symptom: you paste your unpublished draft or someone else's peer-review submission before checking confidentiality and approval.

Fix: stop the upload, follow the local incident route if disclosure occurred, and use only public, synthetic, or explicitly approved content. Apply T12-L01 · What you must never paste.

Generated citations look complete

Symptom: a polished author, title, journal, year, and DOI move directly into the reference list.

Fix: locate the record independently, open the original source, and confirm that it supports the exact claim. Delete any reference you cannot verify.

The disclosure says only "AI was used"

Symptom: a reader cannot tell the tool, purpose, affected part, or human oversight.

Fix: answer the venue's requested fields from the use log. Distinguish manuscript preparation from research-method use and place each description where the policy requires.

A successful run is called reproducible

Symptom: the code ran once, so the method is described as reproducible without saved inputs, environment, version, or rerun evidence.

Fix: claim only what you observed, retain the checked materials, and use T15-L02 before making a stronger reproducibility claim.

Co-authors discover the workflow at submission

Symptom: the team debates acceptability after the text and code are finished.

Fix: agree permitted uses, data boundaries, logging, checks, and disclosure before shared work begins; revisit the agreement if the venue changes.

8. Do it yourself: a disclosure in 30 minutes

Use one real piece of your work, but do not put its content into an AI tool for this exercise.

Minutes 0-5: name the target venue and locate its current AI instructions. Save the page URL, page title, access date, and the exact submission or manuscript location requested for disclosure.

Minutes 5-10: read your institution's relevant policy. List any stricter rule or unresolved conflict. If the target policy is absent or ambiguous, write the question you will send to the journal or research-integrity contact instead of guessing.

Minutes 10-18: reconstruct AI use from existing records: date, displayed tool or model name, purpose, input class, affected section or code, human check, and what survived. Do not invent forgotten details; mark them unknown and resolve them from logs or collaborators.

Minutes 18-25: draft the statement in the venue's requested form. Name literature triage, language editing, code generation, or other uses separately. State the human verification and responsibility without claiming that review made the output error-free.

Minutes 25-30: check every phrase against the policy and your records. Ask each co-author to confirm completeness. If policy compliance cannot yet be shown, label the statement not ready and record the missing decision.

9. Exit check

Deliver exactly one artifact: a disclosure record containing the target journal policy title, URL and access date; the required disclosure location; a specific disclosure statement for one piece of work; and one line confirming how each statement field was checked against the policy and your use records.

It passes when another author can open the same policy, match the tool or service, purpose, affected work, and human checks to your records, identify any unresolved item, and see that the statement is placed where the venue requires. A statement marked not ready can pass the exercise only when it clearly names the missing decision and responsible contact; it is not ready for submission.

10. Rule to remember

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

11. Further reading & tools

• Taught: T01-L02 · Trust but verify - checks claims, numbers, quotations, and references before reuse.
• Taught: T12-L01 · What you must never paste - defines the privacy and confidentiality boundary before using a service.
• Taught: T15-L02 · Reproducibility with AI in the loop - extends the use log into recorded versions, parameters, hashes, reruns, and reported differences.
• Catalogued: ICMJE use of AI by authors (opens in a new tab) - current guidance on disclosure, authorship, checking, attribution, and human responsibility.
• Catalogued: ICMJE use of AI by reviewers (opens in a new tab) - current guidance on journal policy, confidentiality, permission, and disclosure in peer review.
• Catalogued: Elsevier generative AI policies for journals (opens in a new tab) - one publisher's current author, reviewer, disclosure, code, and Methods requirements; check the target journal too.
• Catalogued: Nature Portfolio AI policy (opens in a new tab) - one publisher's current risk-based distinction between assistive, interpretive, and prohibited uses.
• Catalogued: PubMed user guide (opens in a new tab) and Crossref REST API documentation (opens in a new tab) - public routes for locating bibliographic records; a record is not proof that a source supports your claim.
• Catalogued: Tools index - product references; confirm current organisational approval, data handling, and venue rules before use.