Hallucination Rate
Definition
A hallucination rate describes how often model outputs are considered hallucinations according to a defined criterion.
For a hallucination rate, it must be clear what counts as a hallucination and what the frequency refers to. Jesson et al. define for in-context learning a probability of outputs outside a likely range of the underlying mechanism. This is a concrete modeling, not a universal error metric.
Simply explained
Imagine an error bag into which only marbles go according to a defined test rule. Without a rule and a suitable total quantity, the size of the bag says little. A statistical rate remains a simplified summary.
Why it matters in practice
For comparing models, the task and evaluation rule must be documented along with the result. An isolated rate is not sufficient as a quality judgment.
Example from business
Possible example: A team defines before an assistant test which outputs count as unsupported claims and documents the tested tasks and evaluations.
Limits and pitfalls
- Different hallucination definitions must not be silently equated. In particular, the model-related probability in the cited work is not automatically a share of false sentences.
Origin & history
Jesson et al. published the cited first version in 2024. They distinguish true and posterior estimated hallucination rates in a Bayesian framework for in-context learning.
Use in mid-sized companies
- A documented comparison of several configurations with the same operational test task is conceivable. The evaluation should include not only the rate but also individual errors and evaluation criteria.
Frequently asked questions
Is there a context-free hallucination rate?
The cited definition is tied to mechanism, prompt, and chosen likely range.
Is every hallucination rate a sentence error rate?
No. Counting unit and definition must be explicitly stated.
Sources
- Estimating the Hallucination Rate of Generative AI · Jesson et al. (2024): Estimating the Hallucination Rate of Generative AI
Editorially reviewed on: