Hallucination Budget
Definition
A hallucination budget defines the acceptable risk of errors for a specific AI task.
The term transfers the idea of an error budget to factually incorrect or fabricated AI outputs. For this, error type, measurement method, and consequences must be defined. There is no universally acceptable error rate.
Simply explained
In a game, a wrong guess is less consequential than a false statement in an important document. The allowed error margin depends on the task.
Why it matters in practice
Makes quality requirements and necessary controls of a use case explicit.
Example from business
Hypothetical example: A company automatically checks product numbers against master data before output instead of tolerating freely invented entries there.
Limits and pitfalls
- An average value must not conceal severe individual errors. The limit is not a promise that errors are technically excluded.
Origin & history
The term is linked to error budgets from the reliable operation of technical systems.
Use in mid-sized companies
- Agree on separate error tolerances, controls, and abort rules for different tasks.
Frequently asked questions
Can a budget be zero?
As a requirement, yes; however, this does not guarantee that an uncontrolled language model operates error-free.
Sources
- Estimating the Hallucination Rate of Generative AI · Jesson et al. (2024): Estimating the Hallucination Rate of Generative AI
Editorially reviewed on: