Skip to content
PlagiatScanner

AI transparency · How-to · KIT-02

Describe AI use in the methods section

Direct answer: Treat an academically relevant AI system as part of the method, not as a vague assistant. Identify the product and displayed version information, when it was used, the task, input and output, selection criteria, human verification, later edits and material eventually incorporated. The purpose-specific language below is a drafting aid. It must match your actual workflow and the rules for your assessment; disclosure cannot turn prohibited activity into permitted activity.

Seven elements readers need in the method

A methods section explains how evidence and conclusions were produced. Simply naming an AI brand does not reveal whether it proposed keywords, rewrote prose, generated code, classified material or influenced interpretation. A transparent account normally answers seven questions: which tool was used, what version information appeared and when, where in the research process it intervened, what kind of material went in, what came back, how a person evaluated it, and what contribution survived into the submitted work.

These details can be distributed across a paragraph. Start with a concise overview, then place the operational detail beside the affected method. Repeated prompts, discarded responses and access records may sit in an AI use log or an approved appendix. If your department requires a separate declaration rather than a methods statement, follow that location and wording. The important editorial test is that all records tell the same story.

Only use a sentence below when it describes what you did. Remove invented precision and add the real human checks. Expressions such as “fully verified” or “error-free” need a defined verification procedure; otherwise they overstate what the record establishes.

A wording bank organised by type of use

UseAdaptable methods languageDetail to add
Ideas and structure“On [date], I asked [tool] to propose alternative structures from my research question. I compared [scope] responses against [criteria] and designed the final outline.”Criteria, rejected approaches and affected chapters
Search preparation“The system suggested keyword combinations. I then searched [databases], opened the original records and verified every work that is cited.”Keep generated leads separate from sources
Language editing“For [sections], the tool produced options for [clarity/shortening]. The request excluded new claims and references; I reviewed each edit against my source version.”Depth of intervention and final editing
Translation“A working translation of [material] was produced using [tool]. I checked terminology against [disciplinary resource] and edited the final version.”Languages, purpose, permission and protected content
Programming“The system proposed code for [function]. I changed [parts] and ran it in [environment] using [types of tests]; only the recorded revision was used.”Files, dependencies, test cases and failures
Data analysis“The system supported [defined operation] using [category of data]. I reproduced calculations with [independent method], while interpretation and conclusions remained my own.”Data approval, parameters and validation

Image generation needs a slightly different account: describe source material, generation route, selection, editing and labelling. For speech transcription, identify the recording basis, consent where relevant, the correction process and the handling of personal information. Avoid padding the methods section with a catalogue of product features. Report functions that materially affected the evidence, text or analysis.

Rewrite a vague statement as a traceable account

Unhelpful: “Artificial intelligence assisted the analysis.” A reader cannot tell what was analysed or whether the system wrote code, assigned categories or interpreted findings.

Traceable version: “On 21 and 22 August 2026, I used [product and displayed version information] to propose Python code for preprocessing publicly available council minutes according to a coding framework I had already defined. The system did not determine categories or research conclusions. I reviewed the code line by line, replaced its handling of missing values and first ran the revised script in [environment] against three documented test files. The revision, parameters and test results used for the analysis are retained under the file versions cited in the repository. I made all coding decisions and interpreted the results.”

The revised paragraph separates software support from the analytical framework, code revision, tests and interpretation. The number of files belongs there only if it is accurate. If a different validation was performed, state that procedure instead. Methodological precision comes from correspondence with the working evidence, not from adding plausible-looking numbers.

Connect the prose to a reproducible supporting record

A methods paragraph summarises what readers need to understand the research route. The AI use log template can hold individual interactions, dates, selection decisions and checks. An appendix may include significant prompts or redacted extracts when that is authorised and safe. Use identical tool names, dates, activity IDs and task labels across these layers. A generic declaration that conflicts with detailed files undermines rather than strengthens transparency.

A probabilistic service may not produce identical wording when prompted again. Reproducible reporting therefore captures the conditions actually known: access date, product, displayed model information, significant prompt, visible settings, input category, selection rules and subsequent human work. Do not infer a model number from memory. If the interface did not expose a stable version, say so and preserve the other observable details.

The European Commission’s updated guidance for generative AI in research links responsible use to transparency, accountability and reproducibility. It does not create one universal student citation style. It does offer a sound editorial direction: document enough of the system’s role and its limits for a reader to assess the method rather than relying on an assurance that AI was “only helpful.”

Six method-writing mistakes to correct

  1. “Used only for support”: identify the operation and the material affected.
  2. “The AI researched the literature”: distinguish generated search leads from database work and publications you opened.
  3. “All errors were removed”: report checks and known limits instead of a blanket result.
  4. A version guessed later: include only displayed or otherwise evidenced information.
  5. “No data were stored”: avoid unsupported provider claims; review the service, account, settings and institutional policy.
  6. “Produced independently” without boundaries: separate system suggestion, human revision and final decision.

An AI detector cannot reconstruct this process. It produces an assessment of text features, not a reliable history of research decisions. Base a methods account on contemporaneous notes, file versions, queries and actual validation. A full-looking chat record still does not prove which response was incorporated into a paper.

Review the statement for truth and compliance

Use two passes. The evidence pass compares dates, products, tasks, inputs and checks against the log and files. The rule pass compares placement, detail, declaration and appendix with the brief, assessment regulations and current departmental guidance. Where confidential data are involved, resolve the acceptable evidence form with the responsible unit before publishing an appendix. Completeness is not a reason to expose protected content.

  • Every system contribution maps to a research or writing operation.
  • The response, selection, validation and adoption are distinct.
  • References and factual assertions lead to original sources.
  • Privacy and rights questions are not hidden behind broad assurances.
  • The description claims neither universal permission nor a universal disclosure rule.

The AI transparency hub covers purpose-specific records. The foundation guide to academic writing and source practice places tool documentation within the larger evidential process. Once the method is accurate, a shorter summary can be adapted from the AI use declaration template.

Official sources behind the guidance

Sources reviewed 4 September 2026. The specific rules governing an assessment must be verified locally.