The AI verification glossary
Last updated: July 26, 2026
This glossary defines the terms behind verifying AI output: hallucination and confabulation, grounding and source-grounding, claim extraction, the four verdicts (supported, contradicted, disputed, unverifiable), false-alarm rate, precision and recall, RAG, de-identification, and why an AI-content detector is not a fact-checker.
Hallucination concepts
These terms describe how and why AI models produce confident wrong answers. Start here if you are trying to name a failure you have seen.
- AI hallucination — when a model states something false or fabricated with the same fluent confidence it uses for true facts.
- Confabulation — a more precise term some researchers prefer: the model fills a knowledge gap with a plausible-but-unfounded detail, without awareness that it is wrong.
- Knowledge cutoff — the date a model's training data ends, after which it confidently answers with stale information.
- Prompt injection — hidden instructions in untrusted input that hijack a model, another reason you cannot trust output at face value.
Verification concepts
These terms describe how you check an AI answer and what the check can tell you. They are the core of how Mephistopheles works.
- Fact-checking — testing individual claims against evidence and returning a verdict per claim.
- Grounding and source-grounding — tying a claim to an independent, citable source that actually backs it.
- Verdict taxonomy — the four outcomes a claim can receive: supported, contradicted, disputed, unverifiable.
- Unverifiable — a real, honest verdict meaning no authoritative source was found, not a system failure.
- Hallucination-risk score — a single number summarising how much of an answer looks unsupported.
Method and measurement terms
These terms describe the machinery and the honest math behind a verifier. They explain why a checker's own error rate matters as much as the errors it catches.
- Claim extraction — breaking an answer into atomic, individually checkable statements.
- Retrieval-augmented generation (RAG) — fetching outside documents to inform an answer; it reduces but does not eliminate hallucination.
- Precision and recall — the two numbers that measure a detector: how often its flags are right, and how many real errors it catches.
- False-alarm rate — how often a checker flags a true statement; the metric that decides whether you can trust a flag at all.
- AI-content detector — a tool that guesses whether text was machine-written; importantly, not a fact-checker.
- De-identification — stripping personal identifiers before sending sensitive text to external sources for grounding.
Frequently asked questions
What is the difference between an AI hallucination and a confabulation?
Both describe a confident false statement from an AI. Hallucination is the mainstream term. Confabulation is a term some researchers prefer because it borrows from psychology, where it means filling a memory gap with a plausible but false detail without awareness. A 2023 PLOS Digital Health paper argues confabulation is the more accurate metaphor since models have no senses to hallucinate with. In practice the terms are used interchangeably. See AI hallucination and confabulation.
Is an AI content detector the same as a fact-checker?
No, and the difference matters. An AI content detector estimates whether text was written by a machine. A fact-checker tests whether the claims are true. Human-written text can be full of false claims; AI-written text can be perfectly accurate. Detecting authorship tells you nothing about truth. See AI detector vs fact-checker.
Why does 'unverifiable' count as a real answer?
Because a verifier that guesses when it has no source is worse than useless. When no authoritative, independent source can confirm or refute a claim, the honest output is unverifiable — a flag for a human to check, not a silent pass. Marking a claim unverifiable protects you from a false 'supported'.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.