Detect LLM hallucinations

Hallucination detection scores how likely an answer is fabricated, so you can flag, block, or ask-for-review before it reaches a user. There is no single detector; pick by whether you have a ground-truth/context to check against.

Choose a method by what you have

You have retrieved context (RAG / QA): check faithfulness / groundedness - is every claim in the answer supported by the retrieved context?

• NLI/entailment: run each answer sentence against the context with an entailment model; unsupported sentences are suspect.
• LLM-as-a-judge: ask a strong model "is this answer fully supported by the context? list any unsupported claims." Cheap, effective, calibrate it (see add-llm-evals).

You have no reference (open-ended generation): use consistency across samples - a model that knows the answer says the same thing across re-samples; a hallucinating model varies.

• Self-consistency (SelfCheckGPT-style): sample the answer N times; measure agreement. High disagreement = likely hallucination.
• Semantic entropy: cluster the N samples by meaning (bidirectional entailment), then compute entropy over the semantic clusters. High semantic entropy = the model is uncertain about the fact, not just the wording. This is the current state of the art for reference-free detection.

Cheap signals to combine: token log-probs / low confidence, and "I don't know"-style hedging. Weak alone, useful as features.

See references/methods.md for how each method works, cost/latency tradeoffs, and which libraries implement them.

Wire it into production

1. Run the detector inline on high-stakes answers (block/route to human on high hallucination score) or sampled on traffic (dashboard + alert).
2. Emit the score as a span attribute so hallucination rate is trackable over time (see instrument-llm-observability).
3. For RAG, log which claims were unsupported so failures are diagnosable (ties to monitor-rag-quality).

Calibrate (do not skip)

Any detector has false positives/negatives. Label ~50 answers (hallucinated vs not) and tune the threshold to your tolerance - blocking flow favors recall (catch more), UX flow favors precision (fewer false blocks). Re-check after model changes.

Anti-patterns

• Trusting a single LLM-as-judge call as ground truth without calibration.
• Using consistency methods on deterministic (temperature 0) output - you need sampling diversity for them to work.
• Checking only the final answer in RAG, never which retrieved claim was (un)supported.
• Treating low token-probability as proof of hallucination - it is a weak signal, not a verdict.

Grounding

Self-consistency: SelfCheckGPT, Manakul et al. 2023 (arXiv:2303.08896). Semantic entropy: Farquhar et al., Detecting hallucinations in large language models using semantic entropy, Nature 2024 (paper); cheaper variant, Semantic Entropy Probes (arXiv:2406.15927). LLM-as-a-judge: Zheng et al. 2023 (arXiv:2306.05685).