PlagiatScanner

AI transparency · guidance · KIT-09

How to disclose AI-assisted language editing

In brief: Describe the depth of intervention: proofreading, stylistic revision, structural rewriting or addition of content. Identify the tool, affected sections, type of suggestion, checks and the person who decided the final wording. Your own assessment rules determine whether disclosure is mandatory; the scale below helps you report what actually happened rather than hiding different activities under “language improvement”.

Four levels of intervention

  1. Level 1 – surface: Suggestions address spelling, punctuation or typographical errors while propositions and syntax remain substantially intact.
  2. Level 2 – style: Sentences are shortened, repetition reduced or terminology harmonised without intentionally changing claims or evidence.
  3. Level 3 – structure: Sentences are combined, paragraphs reordered or transitions replaced; emphasis and reasoning may shift.
  4. Level 4 – content: New explanations, examples, interpretations or claims appear. This is no longer merely language correction.

Record: “[Tool] was used at level [1–4] in [location]. I compared suggestions with the prior draft, checked terminology and evidence, and made the final decision on each retained change.”

Why “language only” can alter meaning

Language and substance do not always separate cleanly. Replacing “is associated with” by “causes” converts a cautious relationship into causation. Removing “may” can erase uncertainty; simplifying a definition can discard an essential exception. Review should therefore go beyond grammar. Compare proposition, strength of evidence, distance from quoted language and technical vocabulary.

The deeper the intervention, the more detailed the provenance record should become. A brief tool statement may be proportionate for level 1 if local rules permit it. Rewritten paragraphs call for an accessible earlier version and a record of consequential choices. Any newly introduced claim must be treated like other content: traced to reliable evidence, evaluated by the author and removed if unsupported.

Create a lean revision trail

Freeze a draft before processing. Note date, service, displayed model, sections and instruction. Then preserve the accepted changes in tracked changes, version control or clearly dated files. Depending on the rules and scale, representative examples may be enough or a full comparison may be required. A chat screenshot cannot by itself show which suggestion entered the submitted document.

  • Name the pre-edit and final versions unambiguously.
  • Record the aim: error correction, readability, concision or terminology.
  • Identify sections rather than saying only “the whole dissertation”.
  • Flag changes that affect reasoning separately.
  • Verify any new or strengthened claim against its source.
  • Keep confidential passages out of uncontrolled public records.

Where several tools were involved, give each a separate entry. A local spelling checker and a generative rewriting service may have quite different functions and data implications.

Disclosure examples for different depths

Surface: “I used [tool] on 20 August 2026 for spelling and punctuation suggestions in the final draft. The request excluded content generation and substantive rewriting; I reviewed suggestions individually.”

Style: “In Chapter 4 I requested suggestions for shortening long sentences and harmonising pre-defined terminology. The order of argument and cited evidence remained unchanged, and every retained version was compared with my original.”

Substantive intervention: “I requested alternative structures for two paragraphs in section 5.2. I adopted one ordering, removed generated claims and rewrote the transitions. Before-and-after drafts and decision notes are retained.” Calling this last activity “proofreading” would misrepresent its scope.

A content-focused review after editing

  1. Has possibility, association or uncertainty become certainty?
  2. Do quotations, paraphrases and page references still match the originals?
  3. Are defined terms used consistently and accurately?
  4. Are the author's findings distinct from literature findings?
  5. Does the prose fit the discipline rather than a generic polished register?
  6. Can the author explain every important argumentative choice?

Disclosure does not authorise assistance prohibited for a particular assessment. Check the brief and institutional policy; this page is not a universal university rule.

Use the AI declaration template for an overall statement. The avoiding plagiarism guide explains how evidence and independent paraphrase remain necessary. Adjacent documentation is collected in the AI transparency hub.

What automated checks cannot establish

A finished text cannot reliably reveal which wording was changed by which feature. Style indicators or classifiers therefore do not replace version history. Conversely, a complete log cannot demonstrate that every accepted edit is academically correct. Provenance and content review answer different questions; the declaration should not confuse them.

Unpublished manuscripts, personal data and confidential material require particular care. Establish whether processing is permitted before uploading them. A locally retained, redacted comparison may be safer than a public full-text record. The appropriate data protection or assessment contact should decide the handling for the specific project.

Sampling a very large revision set

For thousands of suggestions, retain the full tracked record and define an inspectable review sample: for example every tenth changed paragraph plus every passage containing numbers, quotations, definitions or conclusions. Record the selection rule and findings. If meaning changes occur, widen the review. Sampling cannot replace a mandatory complete check, but it states the reach of your quality control honestly. Do not write “fully verified” where only a reasoned subset received subject review.

Sources

Sources reviewed 4 September 2026.

Next guides