Using AI in Research Without Losing the Evidence Trail

Assign AI a role before the prompt

Decide whether the model is acting as a navigator, formatter, coding assistant, critic, or hypothesis generator. The role determines what can be accepted directly and what must be independently verified.

• Navigation output creates search candidates, not citations.
• Formatting output may transform structure, not meaning.
• Critique output creates questions, not verdicts.

Keep claims attached to sources

Store the source passage, page or section, access date, and your own interpretation beside every important claim. If a model cannot point back to supplied evidence, treat its statement as an unverified lead.

• Never cite a model-generated bibliography without opening every source.
• Preserve quotations outside the generated summary.
• Record when a claim depends on model interpretation.

Separate generated code from measured results

Generated code can accelerate setup, but the researcher remains responsible for dependencies, data transformations, seeds, evaluation logic, and the gap between intended and actual execution.

• Review code before running it.
• Log environment and input versions.
• Save raw outputs before post-processing.

Use a human checkpoint at irreversible boundaries

Publication, participant data, security-sensitive execution, external communication, and high-stakes conclusions require explicit human review. Convenience is not a reason to delegate accountability.

• Redact unnecessary sensitive context.
• Require source checks before public claims.
• Document who approved an irreversible action.

Publish an AI-use disclosure that is actually useful

A useful disclosure names the stage, task, model or tool when material, verification performed, and limits. Generic statements such as ‘AI was used for assistance’ do not help readers assess reliability.