What is an AI hallucination?

A hallucination is when a language model produces text that is fluent and confident but factually wrong or unsupported by its sources. It happens because the model predicts likely tokens rather than retrieving verified facts.

How is the risk score calculated?

The score combines four weighted factors — the factual sensitivity of your domain, the consequence of an error, whether the output is grounded in supplied sources, and whether a verification step exists. Higher factual stakes and weaker checks raise the score.

Does a low score mean my system is safe?

No. A low score means your design choices reduce hallucination risk, but no LLM is immune. Always keep at least one verification layer for any output that informs decisions or reaches users.

What is the single most effective mitigation?

Grounding the model in retrieved, citable source documents (RAG) and instructing it to answer only from those sources, or to say it does not know, removes the largest class of hallucinations.

Is my input sent anywhere?

No. The scoring runs entirely in your browser using a transparent rule set, and nothing you type is uploaded or stored.

AI Hallucination Risk Checker

The AI Hallucination Risk Checker helps you estimate how likely a large language model is to produce confidently wrong output for your specific use case — and what to do about it. Hallucination is the single biggest reliability problem in production LLM systems, and the risk is not uniform: a creative brainstorming assistant tolerates invention, while a medical or legal summariser does not. This tool turns four practical factors into a clear 1-10 score with concrete mitigations, entirely in your browser.

How it works

The checker weighs four inputs. Domain sensitivity captures how factual your field is — law, medicine, and finance are high-stakes, while marketing copy is low. Consequence of error measures what happens when the model is wrong, from a harmless typo to financial or safety harm. Grounding asks whether the model answers from retrieved source documents or from its own parametric memory; ungrounded generation is the largest hallucination driver. Verification asks whether any human or automated check sits between the model and the end use.

Each factor is mapped to a numeric weight, summed, and normalised to a 1-10 scale with a plain-English risk band. The tool then assembles a prioritised mitigation list, surfacing the techniques that close your biggest gaps first — for example, recommending retrieval-augmented grounding when your inputs say the model is answering from memory in a high-stakes domain.

Tips and notes

Treat the score as a design signal, not a guarantee. The most reliable architectures combine three layers: retrieval grounding so the model cites real sources, an instruction to abstain (“say you don’t know rather than guess”), and a verification step — a human reviewer, a second model acting as a checker, or a deterministic validator for structured output. For anything that reaches customers or informs decisions, never ship a single-pass, ungrounded generation. Re-run the checker whenever you change your prompt, your data sources, or your review process, because each of those directly moves the score.