A short field guide for AI-assisted research

Six practical prompts support consequential judgement without logging every interaction. If you are new to GenAI, use the three-step first encounter first: choose safe material, request alternatives, and check one suggestion. A governance record never substitutes for checking the data, sources, analysis or interpretation.

Before using AI

Permission gate — before any material is transferred. Establish whether ethics approvals, consent conditions, institutional and collaborator rules, confidentiality or contractual obligations, and the chosen service’s handling arrangements permit this particular use. An anonymised-looking transcript may still be disclosive. If permission is absent or unclear, do not upload it; use an authorised service or an alternative that does not disclose restricted material.

1. Define the research object. State the substantive question, relevant population or source base, measures or concepts, and intended kind of conclusion. A survey association, archival interpretation and qualitative theme do not share one mechanical test of validity.

2. Set the research boundary. Decide what AI may propose or draft, what must be independently assessed, and what cannot be shared or silently altered. Use the Research Charter where a stable agreement is useful.

When a consequential move occurs

3. Separate proposal from evidence. Capture a proportionate, permitted trace of a proposal and the researcher’s disposition at the time it changes the project, rather than constructing an origin story months later. Do not retain confidential content merely to improve attribution records.

4. Identify a meaningful challenge. Ask which comparison, unexpected case, source contradiction or design-sensitive check could narrow the preferred claim. Carry out the appropriate work; writing a challenge in a log does not constitute executing it.

Before revision and release

5. Authorise the claim. State the strongest warranted wording, the prohibited overstatement and necessary uncertainty. Claim authorisation means deciding what the present evidence licenses the manuscript to say. Reopen the boundary explicitly when later evidence or theory genuinely changes it.

6. Assess attribution and inspect the object. Preserve consequential decision traces throughout the work; assess later contribution statements against those traces at the end. Inspect the delivered manuscript, its figures and metadata, and perform a separate public-release rights check.

Stop or return upstream when…

• restricted material is about to enter an unapproved AI service;
• generated output is being mistaken for executed, read or verified evidence;
• an alternative explanation, source contradiction or negative result has been ignored;
• a descriptive association is rewritten causally without a warrant;
• editing removes an inferential limit or a defensible distinctive contribution;
• repeated AI review produces stylistic agreement but no specified improvement in validity;
• a retrospective account is presented as a contemporaneous trace.

Records are optional and proportional

For a first exploratory exchange, no formal form may be needed. Keep a short dated note only if a suggestion changes something consequential in the research. Longer projects may benefit from a Research Charter, selective Decision Log entries and a Claim-Preservation Contract when writing is revised. Use more detailed records when evidence or contribution is contested. No quantity of documentation establishes analytic validity on its own.

Useful terms: claim authorisation is deciding what the evidence permits you to say; a claim-preservation contract is a short record protecting that wording and its qualifications during edits; reason provenance is evidence about how and by whom a research decision entered the project. No prompt alone guarantees accurate output.

For the stage-by-stage view, use the crosswalk. For a complete constructed teaching decision, see the illustration.