Using generative AI without surrendering research judgement
A practical resource for social scientists who want to use generative AI while keeping verification, inference, consequential decisions and contribution provenance visible.
How can generative AI contribute to social-science research without confusing proposals with evidence, fluent interpretations with valid inference, or better prose with stronger research claims?
Begin with the task you have
For a first practical exchange, see the Start Here guide and its constructed researcher–AI exchange. For teaching or independent practice, start with the constructed illustration or workshop activities. The example is hypothetical; the seven PISA worked cases are bounded accounts of a research programme, not validation of the framework.
Three governing distinctions
Proposals are not evidence
Generated analyses, citations, themes and explanations remain candidates until they pass checks appropriate to the design.
Evidence sets the claim boundary
Claim authorisation means deciding exactly what the available evidence licenses the manuscript to say.
Contribution history needs its own evidence
Reason provenance distinguishes what was recorded when a consequential decision occurred from what was reconstructed later.
Before any AI interaction involving research materials, check whether institutional policy, ethics, consent, contractual terms and the chosen service permit their disclosure. See the field guide.
Two connected governance problems
Reason provenance assesses the evidence for bounded accounts of how decisions, contributions and revisions entered the research process. Preserve traces of consequential moves when they occur; do not postpone collection until the end.
Manuscript governance protects warranted claims, necessary qualifications and distinct contributions during later rewriting. A valid revision may improve the prose; it may not silently improve the evidence.
The two responsibilities connect without becoming interchangeable. Neither a detailed log nor a polished manuscript replaces sound research design and substantive assessment.
When the process itself becomes the problem
Repeated simulated review may encourage stylistic convergence, bury the most original finding or reward ceremonial logging. The risks and failure modes explain when to stop, defend a warranted distinction or return to the actual evidence.