What is Grounding in AI?
Grounding is the practice of anchoring AI responses in verifiable source data rather than relying solely on the model's training knowledge. A grounded response can cite its sources; an ungrounded response might be a hallucination.
Why It Matters
In enterprise contexts, an AI confidently stating incorrect information is worse than saying "I don't know." Grounding forces the AI to base claims on actual documents, enabling fact-checking and building trust with users and auditors.
Grounded vs Ungrounded Response
❌ Ungrounded
"The termination clause allows 30-day notice."
No source cited — could be hallucinated
✓ Grounded
"The termination clause allows 30-day notice. [Source: Contract Section 8.2, page 12]"
Verifiable claim with citation
Grounding Techniques
- • RAG — Retrieve relevant documents before generating
- • Citation Requirements — Prompt must include source references
- • Quote Extraction — Require exact quotes from source
- • Confidence Scores — Flag responses without strong source support
- • Verification Step — Second LLM call to verify claims against sources