Why hallucinations are dangerous
AI hallucinations look exactly like correct answers. They're confident, fluent, and plausible. Without firebreaks, they spread through documents, decisions, and code until someone catches them—often too late.
Firebreak 1: Grounding Sources
Technique: Provide authoritative sources in the prompt and require AI to cite them.
For PMs
Use ONLY the following sources to answer:
- [paste relevant doc 1]
- [paste relevant doc 2]
If the answer isn't in these sources, say "Not found in provided sources."
Question: [your question]
Format: Answer with inline citations [Source 1], [Source 2].
If you cite something not from these sources, flag it as [UNGROUNDED].
For Engineers
Generate code based on this specification:
[paste spec]
Requirements:
1. Every function must map to a spec requirement (cite section)
2. If you generate code not in the spec, comment it as // UNSPECIFIED
3. If spec is ambiguous, list assumptions before code
Do NOT generate features not in the specification.
Firebreak 2: Verify Loops
Technique: Have AI check its own output against criteria before finalizing.
Self-Check Prompt
You just generated this output:
[paste AI output]
Now verify:
1. List every factual claim made
2. For each claim, cite the source OR mark as [UNVERIFIED]
3. For any [UNVERIFIED] claims, either find a source or remove the claim
4. Rewrite the output with only verified claims
Show your verification work, then provide the cleaned output.
External Check Prompt
I'm checking this AI-generated content for hallucinations:
[paste AI output]
For each factual claim:
1. Is it verifiable? (yes/no)
2. If yes, what's the source?
3. If no, what would verify it?
Flag any claim that can't be verified as [NEEDS VERIFICATION].
Firebreak 3: Refusal Patterns
Technique: Train AI to refuse rather than guess.
Uncertainty Framing
Answer this question with the following rules:
- If you're confident, answer directly
- If you're uncertain, say "I'm not certain, but..." and explain why
- If you don't know, say "I don't have reliable information about this"
NEVER guess. NEVER make up statistics. NEVER cite sources you're not sure exist.
Question: [your question]
Explicit Refusal Triggers
If asked about:
- Specific numbers without source data: REFUSE, suggest data sources
- Future predictions as facts: REFUSE, offer scenarios instead
- Legal/medical/financial advice: REFUSE, recommend professional
- Internal company information: REFUSE, note you don't have access
Apply these rules to: [your question]
- ✓Grounding: Provide sources and require citations
- ✓Verify: Run self-check prompts on factual outputs
- ✓Refusal: Teach AI to say "I don't know" instead of guessing
- ✓Flag: Mark unverified claims visibly in output
- ✓Check: Manually verify 3 claims per output as spot-check
High-Risk vs. Low-Risk Tasks
High-Risk (use all firebreaks)
- External communications
- Financial analysis
- Legal/compliance content
- Technical specifications
- Customer-facing content
Low-Risk (grounding sufficient)
- Internal brainstorming
- First drafts for heavy editing
- Code comments
- Meeting summaries
“"We caught a hallucinated regulation citation that would have been in our compliance training. The verify loop saved us from a serious error."”
Related Resources
- Judgment Pack — practice verification techniques
- Bias Spotting Checklist — broader quality checks
Apply this now
Practice prompt
Take an AI output with factual claims and run the verify loop prompt.
Try this now
Add grounding sources to your next factual prompt and compare output quality.
Common pitfall
Trusting AI's self-reported confidence—it sounds certain even when wrong.
Key takeaways
- •Ground AI in sources and require citations—no sources, no answer
- •Run verify loops on factual content—AI checks its own work
- •Teach AI to refuse uncertain answers—silence beats confident mistakes
See it in action
Drop this into a measured run—demo it, then tie it back to your methodology.
See also
Pair this play with related resources, methodology notes, or quickstarts.
Further reading
Next Steps
Ready to measure your AI impact? Start with a quick demo to see your Overestimation Δ and cognitive load metrics.