Hallucination Detection

Detect and prevent LLM hallucinations: factuality checking, grounding verification, and off-brand content filtering.

Last updated 2026-06-12

After this section you can

Classify a hallucination as factual, citation, reasoning, or entity, and name the detection layer that catches each type
Compute a detector’s daily review load from base rate, precision, and recall — flagged cases, false alarms, and leaked hallucinations — and defend the operating point in an interview
Build a two-layer grounding + consistency checker in stdlib Python, and state which production components (NLI cross-encoder, NER) replace the toy checks
Run a weekly calibration loop: sample and human-label 200 cases, recompute per-layer precision/recall, and retune thresholds after base-rate shifts

SD-10

Hallucination & Off-Brand Detection

Detect fabricated facts, fake citations, and brand-policy violations before they reach users.

HALLUCINATION TAXONOMY

FACTUAL HALLUCINATION HIGH RISK

Model states incorrect facts with high confidence

Example: "The Eiffel Tower is 500m tall" (actual: 330m)

Risk: Users trust authoritative tone

CITATION HALLUCINATION CRITICAL

Model invents fake sources, URLs, or paper titles

Example: Cites "Smith et al. 2023, Nature" that doesn’t exist

Risk: Critical in legal/medical/academic contexts

REASONING HALLUCINATION SUBTLE

Logical steps seem plausible but contain errors

Example: Correct premises but invalid conclusion

Risk: Hardest to detect — requires domain expertise

ENTITY HALLUCINATION HIGH RISK

Wrong names, numbers, dates, or identifiers

Example: Confuses CEO names, swaps product versions

Risk: High in customer-facing applications