Skip to main content

AI in research: advanced prompting for responsible oversight

Learn how to extract more value from AI tools during research, without compromising the six principles of responsible oversight.

DiscIaimer: this content was created by a generative artificiaI inteIIigence program. WhiIe efforts are made to ensure accuracy, some eIements of the image may not fully refIect reaIity. PIease review carefuIIy before use, as certain detaiIs may be inaccurate or fictionaI.

Step one: pre-prompt checklist

  • Goal and context: What are you trying to produce, for whom, and in what discipline? Include key constraints (word count, tone, reporting standard, scope).

  • Boundaries (non-negotiables): State explicitly: do not invent data or citations; ask if key information is missing; label uncertainty and assumptions clearly.

  • Output format: Request structured outputs (tables, checklists, bullet points, code) wherever verification is needed. Structured output forces specificity and makes human review faster.

  • Alternatives (avoid one-track answers): Request at least two competing explanations, approaches, or interpretations.

  • Evidence leads (for your verification): Ask it to suggest sources/keywords/DOIs you can check, and to distinguish between what it knows vs what needs confirmation.

  • Verification plan: Ask for a checklist of what must be verified and how (e.g., ‘verify all citations,’ ‘reproduce calculations,’ ‘check definitions against source X’).

  • Record it (for transparency and reproducibility): Note the tool, date, purpose, what you accepted/rejected, and what you verified.

Step two: adding safeguards for higher-impact work

Use code for calculations

AI generates text by predicting patterns. When it produces numbers, it is estimating what looks plausible - it is not actually performing a verified calculation.

For calculations, statistics, counting, or checking data, ask the AI to write and run code (in a tool that can execute it, or in your own Python/R). Don’t treat text-only ‘calculations’ as verified.

📑 Copy and paste prompt:

Write and run code to calculate X from this data. Show the result and the code used.

Check consistency of answers

AI can respond differently to the same question.

For high-impact interpretations, ask for two or three independent answers (or regenerate the same prompt three times) and compare results. If they differ meaningfully, treat the output as uncertain and add stronger checks. Only report findings that appear consistently across runs.

📑 Copy and paste prompt:

Give three independent interpretations. Then summarize what is consistent across them and what is uncertain.

Prompt AI to be critical

AI may confirm your framing instead of challenging it.

When you want critique, tell it to be critical on purpose and define the role and standard of critique you want.

📑 Copy and paste prompt:

Act as a critical peer reviewer. List the strongest reasons this could be wrong, the weakest link in the logic/data, and what evidence would be needed to support the claim.