Hallucination Detection Methods

Detection asks a narrower question than why models hallucinate: given an output, can it be trusted? Robust systems usually combine several of the methods below. Self-consistency checks rely on a simple observation. When a model knows a fact, repeated samples tend to agree, while fabricated details tend to vary. SelfCheckGPT (Manakul et al., 2023) formalized this. It generates several responses and measures whether each sentence of the main answer is supported by the others. The method needs no external source. The drawbacks are the extra model calls and the errors it misses when a model repeats the same wrong answer every time. Source-grounded checks compare the output against reference text, which suits retrieval-augmented systems. The answer is broken into atomic claims, and each claim is tested for entailment against the retrieved passages. Entailment means the passage logically supports the claim. The test uses either a natural language inference model or an LLM. These checks measure faithfulness to the sources, which is not the same as truth if the sources are wrong. Uncertainty signals use the model's own probabilities. Low token probabilities or high entropy on key spans such as names, numbers and dates can indicate guessing. Semantic entropy (Farquhar et al., published in Nature in 2024) groups sampled answers by meaning before measuring uncertainty, so different wordings of the same answer are not counted as disagreement. Token-level signals need access to log probabilities, which some APIs limit. Judge models ask a strong LLM to grade an answer for unsupported claims, usually with the sources and a rubric attached. They scale well, but judges have their own biases and error rates, so check them against human labels. A common misconception is that asking a model 'are you sure?' reliably catches errors. Models may abandon correct answers or defend wrong ones. Structured checks are more dependable.