Build-your-own declaration
Start with the assessment, not a stock disclaimer
A declaration is only useful when it fits the context in which the work is assessed. Locate the module handbook, assessment brief, academic integrity policy, authorship statement and any instructions issued by the supervisor or department. They may define a particular form, position or level of detail. If two documents appear to conflict, the responsible academic office can identify the controlling version; a generic web template cannot.
Next, reconstruct what happened. A student who asked for alternative database keywords has a different disclosure task from a student who generated and revised paragraphs. Code completion, automated transcription, machine translation and generative image work each leave different evidence. Describe the relevant activity instead of placing every digital feature under the label “AI”. If a familiar spelling checker introduced generative rewriting, record the function actually used rather than relying on the product category.
- For ideation, note the question posed and how suggestions were selected.
- For writing, identify the passages and distinguish language editing from changes to reasoning.
- For research, show how every suggested source was independently located and checked.
- For analysis, retain the data state, model details, instructions, parameters and validation.
- For software, identify affected files or functions, human revisions and tests.
- For media, record generation, editing, labelling and rights checks.
Turn vague assurances into auditable information
“AI was used only as a tool” sounds reassuring but provides no boundary. A more informative version would say: “On 18 August 2026 I used [tool, displayed version] to propose synonyms for three search concepts in section 2. I selected five terms, searched the named databases myself, and cited only publications that I opened and verified.” A reader can now distinguish suggestion, selection, retrieval and evidence.
Similarly, replace “AI improved the essay” with a location and an intervention: “I requested sentence-shortening suggestions for sections 3.2 and 3.3. The tool did not choose claims or references. I compared each suggestion with the previous draft and made the final editorial decision.” This wording does not prove that every decision was sound; it does make the claimed division of labour testable against drafts and logs.
Avoid overstating precision. If the interface displayed no durable model identifier, do not guess one. Record the service, provider, access date, visible model label and the absence of further information. If a system changes over time, the prompt alone may not reproduce the output. That is why context, input category, output retained and human decision can be more valuable than a bare screenshot.
What disclosure does not fix
Do not publish sensitive prompts merely to appear transparent. Draft research, personal data, interview material, confidential partner information and unpublished findings may require restricted handling. Follow the relevant data plan and ask the authorised institutional contact how a redacted record should be supplied. An honest declaration can state that evidence is retained under controlled access without exposing the material publicly.
Generated prose is not a substitute for the underlying scholarly source. Verify claims and quotations in the original publication, including page or section details. The guide on why AI output is not a source explains that distinction. For a chronological record, use an AI use log throughout the project rather than reconstructing everything at submission.
A submission-ready audit
- Completeness: Does the statement cover every generative function that materially affected the submitted work?
- Location: Can each intervention be connected to a chapter, file, figure, dataset or stage?
- Verification: Are the checks concrete and limited to those actually completed?
- Consistency: Do the declaration, appendices, version history and final artefact tell the same story?
- Local fit: Does the wording, placement and evidence match the rules for this assessment?
Read the result as an examiner would. Can they tell what the system supplied, what you decided, and how important claims were checked? If not, add one factual sentence rather than another promise of honesty. The AI transparency hub connects disclosure to prompt, model and verification records. For the separate technical question of text signals, see the AI detection overview; classification cannot reconstruct authorship decisions.