Ten-step protocol for generative-AI-assisted social-science research
A researcher’s responsibilities are organised below as ten stages, not ten compulsory forms. Reason provenance runs through the whole workflow. Preserve traces of consequential moves as they occur; stage 10 is a final assessment of which contribution claims those traces can support, not an instruction to reconstruct history after the manuscript is complete.
Before any stage involving AI: establish whether sharing research material with the chosen service is permitted by ethics approvals, participant consent, institutional and collaborator policies, confidentiality obligations, contracts and the service’s data handling arrangements. Do not transfer restricted material while permission remains unresolved.
Avoid mechanical compliance. Delegation creep, ceremonial checking, reviewer-driven convergence and novelty erosion can arise even in well-documented processes. See Risks & Failure Modes and the short field guide.
The protocol crosswalk connects stages to evidence limits and selective records. Disciplinary expertise remains essential; verification varies with the research design.
1. Start with a substantive research problem
Identify a disciplinary question, public event, empirical anomaly, theoretical puzzle or substantive issue that matters independently of generative AI. Record the researcher’s prior commitments and expertise where relevant.
2. Use AI to widen the proposal space
Solicit alternatives to questions, sources, comparisons, interpretations and checks. Treat every output as a candidate, not evidence. When a proposal materially redirects the project, retain a permitted contemporaneous trace and the researcher’s reason for selecting or declining it.
3. Specify the object and constraints
Define units of analysis, comparison groups, exclusions, key constructs, source scope and the inferential target (the specific kind of conclusion the design is intended to support). Make non-negotiable ethical and disciplinary constraints explicit. Record consequential specifications and changes when they happen.
4. Separate artefacts from verified evidence
Generated code, citations, themes, excerpts and claims become potentially evidential only after design-appropriate checks. A quantitative analysis may require execution and diagnostics; an interview interpretation may require close engagement with extracts, analytic context and discordant accounts; an archival argument may require source criticism (checking a document’s authorship, purpose, context and limitations). Verification is not equivalent to completing a template.
5. Challenge the preferred interpretation
Use an alternative explanation, relevant comparator, contrasting source, negative case or sensitivity check capable of genuinely narrowing the thesis. Record consequential challenges and their outcomes at the time; refusing uncomfortable evidence is not successful governance.
6. Record inferential changes
Identify important claims that remain, narrow, change, or are withdrawn. Attach the reasoning and observed trigger to consequential revisions. A numerically correct result may not identify the movement, mechanism or population asserted in the original interpretation.
7. Authorise the present claim
Decide exactly what the checked evidence licenses the manuscript to say. Preserve authorised and prohibited claims, required caveats, population limits and conceptual distinctions. A claim-preservation contract is a tool for expressing that decision, not a source of evidential validity.
8. Govern manuscript revision
Reject stylistic changes that strengthen unsupported claims, remove necessary uncertainty or obscure a defensible novel contribution. Distinguish valid evidence-driven narrowing from convergence towards familiar language. If new evidence or theory changes the claim, reopen the contract and record the change rather than silently overriding it.
9. Inspect the finished scholarly object
Read the export as a reader will: verify figures and labels, consistency among abstract/results/conclusion, references, metadata, permissions and actual public-release suitability. Final inspection cannot be delegated to a generic success label.
10. Assess bounded contribution provenance
Return to ex-ante commitments, contemporaneous records, frozen artefacts and clearly labelled ex-post reconstructions gathered along the way. Decide which claims about AI, researcher and collaborator contributions are defensible and which alternative histories remain unresolved. Do not rewrite later recollection as if it were an original trace.
Governing sequence
Substantive problem → candidate proposals → researcher specification → verified evidence → challenge → inferential revision → authorised claim → governed rewriting → final inspection → bounded attribution assessment.
Reason-provenance recording accompanies consequential moves throughout that sequence. The objective is proportionate, contestable judgement—not maximal record keeping or maximal AI use.