🔮
Fabrication
Completely invented facts entities, statistics, citations, or events that do not exist. The most dangerous hallucination type because the claim is entirely false with no basis in reality.
Example: "According to the WHO 2023 study by Dr. Patel et al. (n=12,000)..." where the study does not exist
🔀
Substitution
Real entities or facts that are swapped or misattributed real names, real studies, or real statistics applied to the wrong context, year, or conclusion.
Example: Correctly citing a real study but stating the wrong finding, or attributing one author's work to another
📅
Temporal Drift
Presenting outdated information as current regulatory requirements that have changed, drug dosages that have been revised, case law that has been overturned, or statistics from a prior year presented as the current figure.
Example: "The current VAT rate for X is 20%" when the rate was revised to 17% in a recent budget
🔢
Numerical Error
Incorrect figures, statistics, calculations, or quantities even small errors in medical dosages, financial figures, or legal penalties can have significant consequences.
Example: Correct drug name, correct indication, but wrong recommended dosage or wrong contraindication threshold
🌀
Over-generalisation
A claim that was true in a specific, limited context is stated as universally applicable applying findings from one population to all populations, or applying a rule that has many exceptions without noting them.
Example: "All Type 2 diabetes patients should avoid X" when the actual guidance is conditional on comorbidities
📚
RAG Faithfulness Failure
In retrieval-augmented systems, the generation contradicts or diverges from the retrieved source documents the model ignores or misrepresents its own retrieved context to produce a more "fluent" response.
Example: Retrieved document says "X is contraindicated in pregnancy" model generates "X is safe during pregnancy" for a smoother response flow
