01, Problem
What Causes ChatGPT Hallucinations?
Hallucinations occur when ChatGPT generates plausible but incorrect information. Common triggers include vague prompts, lack of source constraints, and absence of verification steps. Unlike simple errors, hallucinations confidently fabricate details like fake statistics, non-existent studies, or incorrect facts.
Common Hallucination Triggers:
- Knowledge gaps or outdated training data
- Vague or ambiguous prompts
- No source grounding requirements
- Pressure to provide confident answers despite uncertainty
- Lack of verification or self-audit steps
02, Core
Core Prompting Principles
Effective hallucination prevention hinges on three principles: explicit uncertainty handling, source-grounding, and iterative verification.
Safe Mode Template:
You are a rigorous research assistant. For every response:
1. If you lack definitive knowledge or cannot cite a specific source, state "I don't have verified information" upfront.
2. Only provide facts you can trace to established knowledge through reasoning.
3. When uncertain about any detail, flag it explicitly with [UNCERTAIN].
4. Prefer saying "no answer" over speculation.
03, Verification
Structured Verification Techniques
Pre-Answer Verification Checklist:
Before providing your final answer, complete this verification process:
1. **Source Check**: Can I cite 2+ specific, credible sources for every major claim?
2. **Logic Check**: Does each step follow logically from established premises?
3. **Edge Cases**: Have I considered counterexamples or exceptions?
4. **Uncertainty**: Are there areas where reasonable experts might disagree?
Only after passing ALL checks, provide the answer. If any check fails, explain why and what additional information is needed.
Double-Verification Pattern:
Step 1: Generate your best answer to [QUESTION].
Step 2: Immediately switch to critic mode. Review your answer and identify:
- Any statements without clear sourcing
- Logical gaps or questionable assumptions
- Potentially outdated information
Step 3: Revise the answer based on your critique, or state what cannot be reliably answered.
FAQ