
Study design and research governance
The guidance below highlights impactful applications of AI during interpretation, with practical considerations and prompt templates you can copy and paste to adapt to your own needs.
Aim: Turn results into conclusions you can defend - without overreach.

Generate alternative plausible reasons for the pattern and list analyses that would distinguish them.
Link each conclusion to evidence and uncertainty including what would prove it wrong.
Write clear boundary limits for ‘when this applies that match your design and data.
Prevent ‘story drift’: require AI to output:
what the results support
what they do not support
assumptions/uncertainty
alternatives
what evidence would change the conclusion.
Use ‘argue against this conclusion’ prompts to reduce confirmation bias - but verify every critique against your actual results.
Don’t jump from ‘related to’ to ‘causes’ (correlation vs causation). Justify any cause-and-effect statements (causal claims) explicitly.
📑 Copy and paste prompt: claims table |
|---|
You are a decision-support assistant to help me avoid overclaiming. Use only the information I provide below. Do not introduce new facts, mechanisms, or citations. If something is not supported by what I provide, label it ‘Not supported by provided results.’ Results summary (no raw data): [paste] Study design: [paste] Create a Claims Table with: Claim | What results support | What they do NOT support | Assumptions | Uncertainty/limitations | Alternative explanations (clearly labeled as hypotheses) | What evidence would change the claim | Safer phrasing options. End with: ‘Items requiring human verification’ (bullet list). |
📑 Copy and paste prompt: anti–story drift and devil’s advocate |
|---|
Act as a skeptical reviewer to help me check my reasoning. This is for critique and brainstorming only, not a substitute for expert judgement. Use only my summary; do not add facts. If you speculate, label it ‘Hypothesis.’ Results summary: [paste] Draft conclusion: [paste] Output: (A) Supported by the results (only) (B) Not supported / overreach risks (C) Assumptions + uncertainty to state (D) Alternative explanations (label as hypotheses) (E) Specific checks that could differentiate explanations Then provide a ‘Devil’s advocate critique’ with each point tagged: [Verify with data]. |