In this video, Dylan Davis explains how to prevent AI models like ChatGPT, Gemini, and Claude from “hallucinating” or guessing when extracting information from uploaded documents. He provides a framework centered on model selection, specific prompting rules, and verification methods.
1. Choose the Right Model
The first step is to use high-level reasoning models to reduce errors. As of the video’s date, recommended models include:
- ChatGPT: GPT-5.2 with extended reasoning.
- Claude: Opus 4.5 with extended reasoning.
- Gemini: Gemini 3 Pro.
2. The 3-Rule Grounding Prompt
To stop AI from using its general training data or making things up, you should include these three rules in your prompt:
- Strict Grounding: Tell the AI to base its answers only on the uploaded documents and nothing else.
- Permission to be Uncertain: Explicitly state that if the information isn’t found, the AI should say “not found” rather than guessing.
- Mandatory Citations: Require the AI to provide the document name, page/section, and a direct quote for every claim it makes.
Bonus Rules:
- Mark Unverified: Ask the AI to flag any information it is “unsure” about as “unverified” so you know what to double-check first.
- High Stakes Mode: For legal or financial work, tell the AI to only respond if it is 100% confident. This reduces the amount of data you get but ensures higher accuracy.
3. Verification Methods
Once the AI provides an output, use these three levels of verification to ensure accuracy:
- Self-Check: Ask the same AI to “rescan the document” and provide exact quotes for every claim. Forcing a rescan prevents it from just agreeing with its previous summary.
- Cross-Model Check: Take the first AI’s analysis and the source document, then feed them into a different AI model. Ask the second model to flag any claims not supported by the document.
- NotebookLM: Upload your document and the AI’s analysis to Google’s NotebookLM. Ask it which claims are unsupported; it provides clickable citations to the exact spot in the source text, making manual verification much faster.