Hallucination
A hallucination is generated AI output that presents false, unsupported, invented, or internally inconsistent information as though it were valid.
NIST uses the more formal term confabulation for this risk and describes it as confidently presented erroneous or false generated content. NIST Generative AI Risk Management Profile
OpenAI similarly uses hallucination to describe plausible but false statements produced by language models. OpenAI research on language-model hallucinations
What It Can Look Like
A model might invent:
- A fact
- A statistic
- A citation
- A URL
- A quotation
- A policy
- A person's role
- A technical capability
- A feature that does not exist
- A logical explanation for an incorrect answer
The response may sound completely confident.
Why It Happens
Generative models produce likely outputs based on learned patterns and available context.
They are not inherently performing a verified database lookup every time they answer.
When the system lacks enough reliable information, a plausible continuation can still be generated.
Hallucination vs. Incorrect Source Data
Not every incorrect AI answer is a model hallucination.
The response may be wrong because:
- Retrieved source is outdated
- Database contains incorrect information
- API returned the wrong record
- Prompt supplied false context
- User gave incorrect information
Diagnosing the cause matters.
Can RAG Eliminate Hallucinations?
No.
Retrieval-Augmented Generation can provide better evidence, but a model can still:
- Misread the source
- Ignore it
- Combine sources incorrectly
- Make unsupported claims beyond it
RAG should therefore be evaluated rather than assumed to produce factual correctness.
Why It Matters
The acceptable hallucination risk depends on the use case.
An invented detail in brainstorming may be inconvenient.
An invented:
- Payment status
- Medical fact
- Legal requirement
- Reservation
- Product specification
can have much greater consequences.
How to Reduce Risk
Useful approaches can include:
- RAG
- Clear Prompts
- Reliable source material
- Structured outputs
- AI Evaluation
- Human-in-the-Loop
- Limiting actions
- Explicit uncertainty
- Source citations
No single technique guarantees factual accuracy.