- Keep an AI-use log and source/input checks with the draft.
Appropriate assistance
Treat brainstorming, code explanations and readability suggestions as proposals to evaluate. Independently verify generated citations, quotations and calculations at original sources or tools. A language model is not an authoritative factual database.
Protect information
Do not upload identifiable participant data, confidential peer-review text or restricted documents without authorization. Review each service’s retention and data-use policy before uploading material.
Responsibility and disclosure
Under ICMJE guidance, AI tools do not meet authorship accountability requirements. Follow institutional or journal policy for disclosure of tools, purpose and affected sections; policies differ between disciplines.
Final verification
Independently check results and references, assess possible bias and keep a record of tool use. No automated detector score alone conclusively establishes text provenance or misconduct.
Maintain a usage record
Record tool and version when available, date, task, permitted input and verification method. Identify assisted passages, code or calculations and who checked them. This record supports review of institutional and journal disclosure requirements.
Verify references and code independently
Locate suggested titles or DOIs on the publisher or repository and match authors and the actual supporting passage. Link resolution alone does not establish correct attribution. Test code on controlled inputs and known-answer cases; successful execution is not scientific validation.
Match use to sensitivity
Public permitted data, unpublished documents and identifiable data need different decisions. Use only necessary information and check actual service terms and project permissions. If permitted use is unclear, consult the responsible institution rather than assuming confidentiality.
Worked case and implementation decisions
The following is a fictional teaching case. Do not use its numbers or wording as actual study findings.
In the exercise an AI tool suggests a plausible but nonexistent reference. Check title, authors and DOI against the original record; if the document cannot be verified, do not present it as a valid source. Disclose actual assistance, scope and human checking under the outlet policy. Do not upload confidential data, restricted manuscripts or participant responses to external tools without permission. Tool output is neither specialist approval nor an accountable author. Record version and date to assess later changes.
| Stage | Teaching example | Verification question |
|---|---|---|
| Task | Defined coding or writing assistance | Does policy permit this use? |
| Source | Nonexistent suggested reference | Has the original been checked? |
| Input | Synthetic instead of confidential data | Is transfer authorized? |
| Review | Logic, numbers and attribution | Can authors explain the result? |
| Disclosure | Tool, version, role and human checks | Does it match actual work? |
Exercise output: Keep an AI-use log and source/input checks with the draft.
Sources and further reading
Official sources for verification and further reading