Hallucination (AI)
When an LLM produces output that sounds plausible but is factually wrong or fabricated, e.g. citing a non-existent record or inventing a policy detail.
Quick answer
What is Hallucination (AI)?
When an LLM produces output that sounds plausible but is factually wrong or fabricated, e.g. citing a non-existent record or inventing a policy detail.
Last updated:
Hallucination is the #1 enterprise blocker for LLM deployment. The model's job is to produce fluent text. And when it doesn't know the answer, it produces fluent text anyway. This is dangerous in CRM contexts: a sales rep could read a confidently fabricated customer history and act on it.
Mitigations include: (1) grounding the model with retrieved facts via RAG; (2) prompting the model to say "I don't know" when uncertain; (3) using model "temperature" settings near zero for deterministic outputs; (4) verification layers that fact-check generated content against source records; (5) human-in-the-loop review for high-stakes outputs.
For Salesforce, anti-hallucination is implemented at multiple layers: Einstein Trust Layer's grounding, gptfy's Security Layer's source-citation requirement, and architecture patterns where the LLM can only return content that's traceable to a source record.
Related terms
Browse all terms- GroundingSupplying an LLM with authoritative, current, customer-specific data inside the prompt so its response is anchored in real information, not training data.
- RAG (Retrieval-Augmented Generation)An LLM is given relevant retrieved documents as context before generating a response, grounding outputs in your specific data, not just the model's training.
See it in your Salesforce org
See Hallucination (AI) running in GPTfy
Book 30 minutes with a GPTfy engineer to see how Hallucination (AI) actually works inside a Salesforce org like yours.
Book a demo