Mark Gales

Most hallucination-detection methods rely on at least one of three things: the model's internal probabilities, a labeled reference answer, or a more powerful second model acting as judge. SelfCheckGPT, which Gales's Cambridge group introduced in 2023, removed all three — the idea is to ask the model the same thing several times and watch whether the answers stay self-consistent. Inconsistency across samples is read as a signal of fabrication; consistency, as a signal of grounded retrieval from training.

Worth following when

you need to audit a commercial LLM you don't own and can't pair with a second model.

Topics

zero-resource hallucination detection; self-consistency as an evaluation signal; the speech-modeling tradition behind much of modern LLM uncertainty work.

Key works

SelfCheckGPT (2023); decades of speech foundation work feeding into the uncertainty methodology.