A three-rule prompt to stop AI hallucinating on documents

Executive overview

AI models are trained to be helpful, so when they can't find information in a document, they guess — confidently. This makes document extraction one of the highest-risk AI use cases.

Three prompt rules ground the model in the document and cut hallucination rate significantly. Pair them with the right reasoning model and a structured verification pass.

Explicitly telling AI it's okay to say "not found" is the single most effective hallucination-reduction technique.

Choose the right model

• Use the highest-tier reasoning model available for document extraction tasks
• Default models are optimised for helpfulness, not accuracy under uncertainty
• High-reasoning models are less likely to fill gaps with training data

The three grounding prompts

• Grounding: "Base your answer only on the uploaded documents and nothing else"
• Permission to say not found: "If information isn't found, say 'not found in the documents' — don't guess"
• Citation requirement: For each claim, include document name, page or section, and a relevant quote

Two bonus prompts

• Middle ground / unverified: If the AI finds something related but isn't fully confident, ask it to mark the claim as "unverified" — prioritise those for human review
• High-stakes mode: "Only respond with information if you are 100% confident it came from the file" — trades recall for precision; use for contracts, financial analysis, legal documents

Three verification methods

1. Self-check (low intensity): In the same thread, prompt: "Rescan the document for each claim — give me the exact quote that supports it; if you can't find the quote, retract the claim." The rescan instruction forces the model to re-read rather than validate its own output.
2. Multi-model check (medium intensity): Export the first model's analysis and the source document; feed both into a different AI (e.g. if the first was ChatGPT, use Claude or Gemini). Prompt: "Review this analysis against the uploaded document — flag any claims not directly supported."
3. NotebookLM (high intensity): Upload the source document and the AI's analysis to NotebookLM; ask which claims are not supported. Returns clickable source citations for fast human validation. Powered by Gemini, so use ChatGPT or Claude as the primary extraction model.