Skip to content
PlagiatScanner

AI transparency · Template · KIT-01

AI use log template for essays and theses

Direct answer: Record each academically relevant AI interaction while it is still fresh. A useful row identifies the tool and version information shown, the date, task, purpose, input category, kind of output, human checks, edits, material finally adopted and the location of supporting evidence. The downloadable table and worked example below create an audit trail; they do not grant permission to use AI or override the rules governing your assessment.

A downloadable table for recording work as it happens

The fields below distinguish a generated suggestion from a student decision. Download the blank CSV or recreate these columns in a locally approved spreadsheet. Add institution-specific fields when your course documentation asks for them.

Download the blank CSV template

FieldEntry to makeFunction in the record
Activity ID and dateA unique short code and interaction or access dateConnects the row with files and appendix items
Tool detailsProduct, provider and any model or version visibly displayedDescribes the working environment without guessing
Task and purposeChapter, file or operation and the specific reason for useDefines the scope of assistance
Input categoryYour prose, keywords, code, data or public source materialSurfaces privacy and rights questions
Output and adoptionType of response and what, if anything, entered the workSeparates generation from the final submission
Checks and editsSource review, tests, comparison or disciplinary validationRecords human responsibility
Evidence locationFilename, appendix label or approved protected locationMakes the item retrievable later

Worked example: generating search terms, not references

IDAcademic contextAI contributionHuman review and use
R-014, 2 September 2026Chapter 2 literature search on municipal climate adaptationNamed tool and displayed version recorded. Prompt requested German and English search terms; response contained eighteen suggestions.Selected eight terms and searched the named scholarly databases. References came only from catalogues and publications opened by the student. Extract stored as Appendix A-3; no factual claim copied from the response.
R-015, 3 September 2026Readability of two original sentences in section 2.4Requested three shorter formulations without adding evidence or claims.Compared alternatives with the source text, restored specialist terms and wrote the final wording. Version comparison retained in the working folder.

This record goes beyond preserving the prompt. It makes selection, database searching and adoption visible. A reader can see that suggested terms were leads rather than citable evidence, and that a generated wording option was not automatically the final sentence.

Build the log into the research routine

Create the file before the first relevant interaction and complete a row soon after each use. A list reconstructed just before submission is more likely to confuse dates, versions and purposes. Short identifiers such as R-001 make cross-referencing manageable. Use the same identifier in an export filename when an export is permitted. Repeated operations may sometimes be grouped, but the period, volume, selection criteria and resulting use should remain clear. Local rules can require a different level of detail.

Write purposes as observable activities. “Develop possible database keywords for section 2.1” communicates more than “research support.” “Offer three ways to shorten my existing paragraph without changing its claims” is more informative than “improve writing.” Note a rejection as well as an adoption. Recording that all proposed references were discarded is evidence of a decision and prevents the log from implying that every response became part of the paper.

Do not retain or publish material merely for the sake of a complete log. Personal data, confidential research, unpublished ideas and third-party content can require special handling. Check institutional approval, service terms and the sensitivity of the material before entering or exporting it. Where a full conversation would be inappropriate, agree early whether a redacted extract, a local summary or another form of evidence is acceptable.

Describe the human check, not just its outcome

“Checked by me” is difficult to assess. Replace it with a testable action: opened each paper in the publisher record, compared a quotation with the page in the original, reran a calculation independently, executed code against documented test cases, or reviewed every proposed edit in a version comparison. The log should not turn an AI-detector score into proof. It is designed to record the research process and the author’s critical intervention.

Claims and citations lead back to genuine publications. A model response is not evidence for an article, statistic or page reference. For code, pair the relevant input with the environment, human modifications and test result. For translation, identify the source version, terminology check and final editorial responsibility. When writing the paper, the detailed entries can support a concise methods account using the methods-section guide to AI use.

An activity log has evidential limits

A completed table does not by itself prove authorship, completeness or compliance. It cannot replace citations, consent or authorisation under assessment rules. A saved conversation records one captured interaction but may omit other stages and does not establish how the submitted text was produced. Describe only what the surviving record can support.

The European Commission’s guidance for responsible generative AI in research connects transparency with accountability and reproducibility. The German Research Foundation likewise states that using generative models does not remove researchers’ responsibility for content and form. Neither source imposes this particular spreadsheet on every student. The practical implication is narrower: disclose tool involvement honestly, preserve the author’s decisions and verify scholarly claims independently.

More documentation is not automatically better. Prompts can expose protected data, unreleased work or third-party material. Appendix design must therefore consider sensitivity as well as completeness. The separate guide on documenting prompts in an appendix helps choose between a full record, a representative extract and a structured summary.

Seven questions before submission or archiving

  1. Do the tool name, displayed version and date match the retained evidence?
  2. Is each relevant interaction tied to a defined task and purpose?
  3. Does every row separate the response, human verification, edits and final adoption?
  4. Were all citations and factual claims checked in the original publications?
  5. Are rejected outputs represented accurately rather than silently treated as used?
  6. Does the supporting material avoid inappropriate disclosure of personal or confidential content?
  7. Does the format meet the assessment brief, declaration wording and instructions of the responsible academic unit?

Finally, compare the log with the submitted work. Every item marked as adopted should be locatable, while significant documented assistance should not disappear from the disclosure. The AI transparency hub connects the surrounding workflows. The foundation page on academic writing and source practice places this record within a wider process of research, evidence and independent judgement.

Official foundations for this workflow

Sources reviewed 4 September 2026. Applicable institutional and assessment rules must be checked separately.