What is hallucination?
Hallucination is incorrect, fabricated or unsupported model output. It is not simply a model 'lying'; the model does not necessarily know the answer is wrong.
Language models generate likely sequences and do not possess a human experience of truth or intent.
Why do models make errors?
Ambiguous questions, missing Context, stale data, conflicting sources and pressure to answer can raise error rates.
The whole system—including retrieval, tools and output handling—can introduce errors.
Types of error
Errors include fabricated claims or citations, wrong numbers, entity confusion and answers inconsistent with retrieved documents.
- Unsupported claims
- Irrelevant or fabricated citations
- Document misinterpretation
- Numerical or temporal error
- Incorrect tool result handling
Why does this matter in an enterprise?
In legal, financial, HR or operational work, a wrong answer can become a real decision or action.
High-risk scenarios need abstention, escalation and visible uncertainty.
How does RAG help?
RAG retrieves relevant information before generation and can reduce unsupported answers, but it cannot reduce risk to zero.
Poor retrieval, stale documents or inappropriate chunks can create bad Context.
Grounding and citation
Grounding connects a claim to evidence. A citation aids review but does not prove the claim is supported by that source.
Source freshness, authority and claim-level support must be checked.
The role of evaluation
Evals should measure faithfulness, factuality, citation quality and behavior when no answer is available.
They must run again when models, sources or retrieval change.
Can errors be eliminated?
No. Systems can reduce probability and impact, but zero error is not a realistic guarantee for generative AI.
Trusted sources, controlled retrieval, evals, guardrails, action limits and human approval form layered defenses.
Frequently asked questions
Is hallucination the same as lying?
No. The model is not necessarily aware that an answer is wrong.
Does RAG eliminate errors?
No. It lowers risk and depends on retrieval and source quality.
Does a citation guarantee correctness?
No. The cited evidence must actually support the claim.
Sources
- Google CloudCheck grounding with RAGOfficial Documentation
- Google CloudGrounding overviewOfficial Documentation
- NISTArtificial Intelligence Risk Management Framework: Generative AI Profile26 July 2024 · NIST AI 600-1