Write an AI disclosure that matches the work and the journal

1. Identify the policy and required location

ICMJE calls for explaining AI use in the cover letter and relevant manuscript section. Elsevier specifies a separate manuscript-preparation declaration and reporting research-process use in Methods. These are concrete policy examples; check the target journal’s current instructions rather than assuming identical wording or placement.

Record the policy address, date checked and article type. Read any manuscript template and portal questions alongside the policy. If one asks a broader question than another, resolve the difference rather than making inconsistent declarations. A disclosure that meets a visible field limit may still omit information requested elsewhere. Keep the submitted answer and the longer internal use record linked by a clear version identifier.

Reference: ICMJE: use of AI by authors · Elsevier: generative AI policies for journals

2. Separate language work from research work

List each task in ordinary words: translating an author-written abstract, reorganising a paragraph, suggesting search terms, drafting code or generating an explanatory illustration. State what input was used and which part of the final work was affected. Do not combine all of these under “language polishing” if the tool also influenced the analysis.

The proposed table is an inventory aid, not a permission matrix. A task appearing here is not automatically allowed by a journal or an institution. A writing declaration cannot replace the methodological account needed to understand how a result was produced. Similarly, a description of research software does not necessarily explain a separate translation step.

On small screens, scroll the table horizontally, or focus it and use the arrow keys.

3. Capture facts while they are available

For each use, record the service name, date, available model or version identifier, purpose, affected material and review performed. If the interface does not expose an identifier, record that limitation. Do not substitute a provider’s latest model name or a name the tool generated in its own response. A subscription label alone may not identify the model behind a particular result.

Keep the evidence proportionate to the task and within the approved storage arrangement. A short use record may be sufficient for a limited wording change; research code needs the actual code and execution evidence. Do not attach raw prompts containing confidential material to a public supplement merely to make the record look detailed. Identify what can be shared and who may authorise it.

4. Draft a statement from a fictional translation record

Assume a fictional team used an imaginary service called Tool T on 10 August 2026 to translate its author-written Arabic abstract into English. The interface did not expose a model identifier. Two named team members checked the translated abstract against the Arabic source and corrected a negation error. All authors then approved the corrected wording for submission. The tool was not used for data analysis in this invented scenario.

A draft statement could read: “We used the fictional Tool T to prepare an English translation of our Arabic abstract on 10 August 2026. The interface did not provide a model identifier. The authors compared the translation with the source, corrected identified discrepancies and approved the submitted wording. This use did not include data analysis.” This is original example language, not a publisher’s mandatory template. In a real statement, replace every detail with what actually occurred.

5. Recognise when a disclosure sentence is insufficient

Now change the fictional facts: Tool T also suggested code used to calculate a study result. The translation-only statement is no longer accurate. The team needs to identify the code, describe the relevant method and show how the result was checked. If no execution record exists, the missing evidence remains a problem even if the disclosure names the service.

For a proposed automated classification, the account may need the task definition, inputs, output interpretation, available settings and how errors were assessed. Determine the appropriate reporting detail with the scientific team and the journal. Do not invent parameters, claim access to an undisclosed model version or present a generic disclosure paragraph as proof that the analysis is reproducible.

6. Reconcile the disclosure with the final files

ICMJE makes people responsible for AI-assisted content and rejects treating AI-generated material as a primary source. Check substantive claims against their actual evidence, not a tool’s assertion that a reference exists.

Before submission, compare the internal task record with the cover letter, manuscript, image captions where required and portal answers. Review later revisions too: new assistance can make an earlier statement incomplete. Ask the relevant contributors to confirm the account of their work. If the policy remains unclear, retain the exact question and the journal’s response.

A final statement should be specific enough to explain the assistance without claiming checks that did not happen. Preserve the submitted version and record any later correction through the journal’s process. The intended outcome is a truthful, readable account of tool use and human responsibility, with remaining limitations visible.

Reference: ICMJE: use of AI by authors

Questions for a practical final check

• The policy and required locations are recorded.
• Writing, translation and research tasks are separated.
• Unknown model information is identified honestly.
• The statement matches the actual use record.
• Review claims have supporting evidence.
• All submitted files contain consistent disclosures.