PlagiatScanner

AI transparency · guidance · KIT-10

How to disclose AI-assisted translation in academic work

In brief: Record the source text and permission context, tool and date, purpose and extent, linguistic or subject review, and the place of disclosure. Translation does not change ownership of ideas or evidence: a machine-rendered passage remains tied to the original source and still requires citation. Your institution and assessment determine the required form; the matrix below supplies a transparent working record.

Source–review–disclosure matrix

Evidence fields for translated academic material
Stage Record Question
Source Author, work, passage, language and lawful access Can the original proposition be found?
Tool Provider, product, displayed model and date Which function produced the draft?
Extent Abstract, quotation, research material or chapter Exactly what was processed?
Review Terms, meaning, numbers, modality and caveats Who checked it, using what evidence?
Disclosure Method, note, appendix or declaration Does placement meet local rules?
Final version Retained changes and human editing Is the route to the submitted wording clear?

Distinguish the translation context

Your own prose: You translate an abstract or chapter you wrote. The main issue is linguistic assistance, but the versions still need to match in claims, numbers and limitations. A quotation: You render a passage from another publication. The original publication remains the source, and the reader needs to know whether the translation is yours or tool-assisted. Research material: Interviews, free-text responses or documents are translated for analysis or reporting. Translation choices then become part of method because they can influence categories and interpretation.

Search vocabulary presents another context. Translated terms may alter database retrieval. Preserve important term choices in the search protocol, particularly where apparently similar words have different disciplinary meanings. A generated translation is a candidate expression, not evidence that a field uses it in that way.

Validate meaning rather than fluency

Back translation is not a truth machine. Two systems may preserve the same mistake or smooth away the same ambiguity. Compare original and target sentence by sentence. Mark technical terms, negation, quantities, modal verbs, temporal relations and qualifications. Consult disciplinary dictionaries, terminology databases or parallel texts from authoritative organisations. A central passage may justify review by a person competent in both the subject and language.

  1. Freeze and identify the unaltered source passage.
  2. Save the first machine-produced version separately.
  3. Mark ambiguity and competing renderings.
  4. Resolve key terms using dependable subject sources.
  5. Retain the final text and reasons for consequential changes.
  6. Recheck citation details against the original publication.

If you cannot competently evaluate a language, do not claim full validation. State the review that genuinely occurred and seek qualified support where the reliability of the passage is important.

Translation does not erase citation

A direct quotation remains a quotation after translation. Follow the chosen citation style for identifying the translation and list the original work. “Author's translation” is often used when the researcher produced the wording; machine assistance adds a separate provenance point that should be described according to the applicable policy. There is no single notation suitable for every department or style.

Paraphrases also require engagement with the original. Do not paraphrase a generated rendering while never opening the source. Verify its meaning and cite the publication. Never accept invented page numbers or bibliographic records. The avoiding plagiarism guide explains the continuing relationship between attribution and independent wording.

Proportionate disclosure examples

Your abstract: “The English version of my German abstract was pre-translated with [tool and displayed model] on 22 August 2026. I checked terminology, figures and qualifications against the final German version and edited the English text.”

A quotation: “The English passage from [source and page] was rendered into German with [tool] and edited against the original; omissions and emphasis are marked.”

Interview material: “Selected extracts were pre-translated with [tool] for presentation. Coding was conducted against the original-language material, and ambiguous key terms are documented in the translation log.”

These examples neither authorise processing confidential text nor create a universal disclosure rule. Check consent, data protection, contracts and assessment instructions before using an external service.

Confidentiality and unpublished texts

Do not upload identifiable interviews, internal records or unpublished manuscripts simply for convenience. Establish purpose, lawful handling, contractual controls and institutional approval. Pseudonymisation may reduce risk but is not automatically anonymisation; combinations of detail can still identify people.

The evidence trail may describe the data category and controlled procedure without reproducing confidential originals in a public appendix. The guide to confidential data in AI tools provides a triage framework. Further documentation routes are collected in the AI transparency hub.

Use a terminology ledger

Maintain a bilingual record for central terms: original expression, chosen rendering, rejected alternative, supporting usage and reason. Search the final document for consistency. The ledger is especially useful across several chapters, translators or tools. It prevents one concept receiving accidental variants or different concepts collapsing into one term. It documents your decision but is not an external authority. Where context requires variation, record the contexts separately rather than enforcing false uniformity.

Sources

Sources reviewed 4 September 2026.

Related pages