Check policy and scientific meaning before using an AI translation

1. Ask a specific policy question

Elsevier’s current author guidance explicitly addresses translation: it requires bilingual scientific checking, protection of the submitted material and disclosure. This is a publisher-specific example, not permission from every journal. Read the target journal’s current instructions for the actual use proposed.

Write the question concretely: “May we use this service to translate our own unpublished abstract, under these processing conditions, and where should we report it?” Check whether the relevant policy distinguishes author preparation from peer review, research methods or image generation. A statement about ordinary spelling checks does not necessarily answer a question about translating a scientific argument. Save the policy link and any clarification with the date consulted.

Reference: Elsevier: generative AI policies for journals

2. Establish what may enter the service

Identify the owner of the text and the person authorised to decide how it is processed. Check the actual service and account arrangement, including retention and access settings, against the institution’s requirements. Do not infer permission from the tool’s brand, price or a colleague’s successful login. If a necessary condition is unresolved, use an approved alternative or postpone that processing step.

For practice, use wholly fictional sentences like the example below. For a real draft, consider whether names, unpublished participant details, identifiable quotations or sensitive appendices are present. Removing names alone does not establish that the remaining text is suitable to share. Keep the permission record separate from the translation-quality record: good wording cannot repair an unauthorised disclosure.

3. Freeze the source and specify the task

Save an identified source version before translation. Ask for faithful rendering of that version, preserving the study design, direction of claims, uncertainty, numbers, units and citations. Tell the reviewer which sections were translated and which were only edited. If the source itself is unclear, resolve the author’s meaning first; fluent guessing does not improve a scientific ambiguity.

Create a short terminology sheet for recurring concepts, abbreviations and official names. Include definitions when a term has several plausible translations. Keep bibliographic titles and affiliation names tied to their verified forms where appropriate. A glossary is a consistency aid, not evidence that every use of a word fits its context.

4. Audit a deliberately flawed fictional translation

The fictional Arabic source says: “شملت الدراسة الرصدية 120 بالغاً. غاب 24 منهم عن الموعد، أي 20%. ولم تثبت الدراسة أن التذكير تسبب في زيادة الحضور.” A deliberately faulty English output calls the study a randomised trial, reports 24% missed appointments and says that reminders proved effective. These changes are planted errors for the exercise, not observed performance of a named tool.

A faithful rendering is: “The observational study included 120 adults. Of these, 24 missed their appointment (20%). The study did not establish that reminders caused increased attendance.” Check the number directly: 24 divided by 120 equals 0.20, or 20%. Neither the arithmetic nor the translation supplies a causal analysis or an effectiveness conclusion.

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

5. Review the actual manuscript, including difficult details

Assign someone who can understand both languages and the scientific content to compare the source and translation directly. Mark omissions, additions and meaning changes. Check table headings, decimal separators, inequalities, confidence intervals, units, participant labels and whether an association became a causal statement. Record unresolved terms rather than selecting the most impressive-sounding wording.

For a long manuscript, maintain a section-by-section review record with version, reviewer and outcome. Ask the methods contributor about statistical uncertainty where language alone cannot settle it. Back-translation or a second AI response may reveal a difference, but should not be counted as completed scientific review by itself. Recheck passages changed after the reviewer’s last pass.

6. Report the work that actually happened

ICMJE places responsibility for AI-assisted submitted material on people and calls for disclosure of its use. Translation does not transfer that responsibility to a service.

Keep a concise record of the service, available version information, date, sections processed and human checks completed. Do not invent a model identifier that the service did not expose. Use the journal’s required disclosure location and describe the task accurately; a statement that the team only checked spelling would misdescribe a translated abstract. Retain appropriate evidence privately without uploading sensitive logs as routine supplementary material.

Before submission, compare the final translated file with the approved source and the portal’s language fields. A later copy-and-paste error can undo an earlier review. If policy clarification or scientific interpretation remains open, show that status in the team’s record instead of marking the translation ready.

Reference: ICMJE: use of AI by authors

Questions for a practical final check

• The target journal’s actual translation policy is checked.
• The material and processing arrangement are authorised.
• The source version is identifiable.
• Design, negation, numbers and uncertainty match.
• A named human review covers the submitted version.
• The disclosure accurately describes the translation work.