01
Where this one gets misread
Retrieval and citations lower the rate; they do not remove it. The dangerous case is not the obviously wrong answer but the plausible one carrying a real-looking citation to a real document that does not support it.
Glossary
What is AI hallucination?
An AI hallucination is output that sounds plausible and is unsupported or wrong. Retrieval and citations lower the risk of one without removing it.
Updated 21 Aug 2026
02
Questions to ask
Ask for a worked example on your own material, and the evidence needed to reproduce it.
- What is the rate on your material, not a benchmark?
- What does it do when nothing supports an answer?
- Is abstention tested as a behaviour?
- Can a citation point at a passage that does not support its claim?
03
How Marella uses the term
We use “AI hallucination” only where a product mechanism or an evaluation method backs it up, and we say when the behaviour depends on how a deployment is configured.
- Backed by a product mechanism or an evaluation method
- Deployment differences flagged
What this page does not prove
- B1A definition is not a claim about how the product performs.
- B2Vendor implementations vary.
- B3Test the term against a representative workflow.
Related pages
Test the claim on your documents
Pick a real piece of work, agree what a good answer looks like, then go through the results together.