Conversation with Gemini
Summarize : m.youtube.com
In this video, Dylan Davis provides a comprehensive guide on how to prevent AI models like ChatGPT, Gemini, and Claude from "hallucinating" (guessing) when extracting information from documents.
The Problem: Helpful vs. AccurateAI models are trained to be helpful assistants. If they cannot find information in a document, they often guess confidently rather than admitting they don't know, filling in gaps with their existing training data [ 01:17].
Step 1: Choose the Right ModelTo minimize errors, always use high-level reasoning models rather than default versions [ 01:47]. Recommended models include:
- ChatGPT: GPT-o1 (or the latest reasoning model) [ 02:13].
- Claude: Opus 3.5/4.5 with extended reasoning [ 02:20].
- Gemini: Gemini 1.5/2.0 Pro [ 02:25].
Step 2: The 3-Rule Grounding PromptAdding these three specific rules to your prompt forces the AI to stay anchored to your file [ 03:02]:
- Strict Grounding: "Base your answer only on the uploaded documents and nothing else" [ 03:13].
- Permission to Fail: "If information isn't found, say 'not found.' Do not guess" [ 03:30].
- Mandatory Citations: Require the AI to provide the document name, page/section, and a direct quote for every claim [ 04:07].
Bonus Tips:
- The Gray Area: Ask the AI to mark information as "unverified" if it is only "pretty sure," helping you prioritize what to double-check manually [ 05:20].
- High Stakes: For legal or financial work, tell the AI to only respond if it is 100% confident [ 06:07].
Step 3: Verification MethodsDavis outlines three ways to check the AI's work [ 06:36]:
- Self-Check: In the same chat, ask the AI to "rescan the document" to find supporting quotes for its own previous claims [ 06:56].
- Cross-Model Check: Take the first AI's analysis and the source document, then feed them into a different AI model to flag unsupported claims [ 07:35].
- NotebookLM: Upload your document and the AI’s analysis to Google's NotebookLM. It is specifically designed for grounded search and provides clickable source links for easy validation [ 08:25].
Video Link: The 3-Rule Prompt That Stops ChatGPT, Gemini, and Claude From Guessing
The 3-Rule Prompt That Stops ChatGPT, Gemini, and Claude From Guessing
Dylan Davis · 84K views
Gemini is AI and can make mistakes.
|
