Hallucination Rate

The hallucination rate is the proportion of generated statements that contain factually incorrect, nonsensical, or unfaithful information relative to a ground truth source. It is typically calculated as:

• Formula: (Number of hallucinated statements / Total number of factual statements generated) × 100
• Scope: Applies to both intrinsic hallucinations (contradicting the provided context) and extrinsic hallucinations (fabricating information not verifiable in any source)
• Granularity: Can be measured at the passage, sentence, or atomic claim level for precision

Several algorithmic approaches exist for detecting hallucinations at scale without human annotation:

• Natural Language Inference (NLI): Uses an entailment model to check if a generated claim is logically supported by the evidence text, outputting entailment, contradiction, or neutral labels
• SelfCheckGPT: A sampling-based approach that generates multiple responses to the same prompt and measures semantic consistency between them; high variance indicates potential hallucination
• Chain-of-Verification (CoVe): The model generates its own fact-checking questions about its output, answers them independently, and compares results to identify inconsistencies
• LLM-as-Judge: Using a stronger model to critique and score the factual accuracy of a weaker model's output against a reference

Hallucination rates are not uniform and vary significantly based on task and domain:

• Open-domain QA: Higher rates due to reliance on parametric knowledge without retrieval grounding
• Summarization: Moderate rates, with extractive approaches hallucinating less than abstractive ones
• Closed-book generation: Highest hallucination risk, as the model must rely entirely on its training data memory
• Domain specificity: Medical and legal tasks show elevated rates without RAG, as they require precise, up-to-date factual recall
• Long-form generation: Rates increase with output length due to exposure bias and compounding errors

Despite advances in automation, human judgment remains the gold standard for hallucination rate measurement:

• Atomic Claim Decomposition: Annotators break generated text into individual factual claims, then verify each against source documents
• Inter-annotator agreement: Measured via Cohen's Kappa or Krippendorff's Alpha to ensure labeling consistency across evaluators
• Fine-grained taxonomies: Modern protocols classify hallucinations into subtypes including entity errors, relation errors, temporal errors, and numeric errors
• Benchmark datasets: TruthfulQA, HaluEval, and FActScore provide standardized test sets for comparing hallucination rates across models

An evaluation score measuring the degree to which generated text is factually consistent with and directly inferable from the provided source context. Unlike general accuracy, faithfulness specifically penalizes extrinsic hallucinations—information added by the model not present in the source.

• Uses Natural Language Inference (NLI) models for automated scoring
• Critical for evaluating summarization and RAG systems
• A score of 1.0 indicates all claims are directly supported by the context

A measure of whether all factual claims in a generated summary or answer are supported by the source text, ensuring no contradictions or fabricated details are present. Factual consistency evaluation often employs entailment scoring between source and generated text.

• Detects both intrinsic (source contradiction) and extrinsic (fabrication) hallucinations
• Benchmarked using datasets like FactCC and SummaC
• Foundational metric for RAG output quality

The alignment between a model's predicted probability of correctness and its actual empirical accuracy. A well-calibrated model assigns low confidence to likely hallucinations and high confidence to correct outputs.

• Measured via Expected Calibration Error (ECE)
• Poor calibration leads to overconfident hallucinations
• Post-hoc methods like temperature scaling improve calibration without retraining

A method for reducing hallucination where a language model first drafts a response, then generates a series of independent verification questions to fact-check its own initial output, and finally produces a corrected answer.

• Implements a self-critique loop without external tools
• Reduces hallucination rate by up to 28% in benchmarks
• Each verification step targets specific factual claims in the draft

A guardrail mechanism that uses a language model's perplexity score on generated text to detect and filter out low-quality, nonsensical, or potentially hallucinated outputs. High perplexity indicates the model finds its own output improbable.

• Operates as a lightweight, inference-time filter
• Effective for catching degenerate text and confabulations
• Can be combined with semantic coherence checks for stronger filtering

The process of using a Natural Language Inference model to calculate a probability score indicating whether a given evidence text logically implies a target claim. This forms the backbone of automated fact-checking in RAG evaluation.

• Outputs three classes: entailment, contradiction, neutral
• Models like DeBERTa fine-tuned on MNLI are standard
• Aggregated scores produce document-level faithfulness metrics