Averaging Bias: Human Faithfulness Annotations are not Locally Faithful

2026-07-31

AI Summary

This study addresses a critical gap in the evaluation of text summarization faithfulness, which traditionally relies on human global annotations whose alignment with the strict conjunctive rule—that every sentence must be factually accurate—remains unclear. The authors identify and formally name a previously undocumented phenomenon, “Averaging Bias,” wherein summaries containing local factual errors still receive positive overall faithfulness ratings from human annotators. To investigate this, they leverage five large language models to perform sentence-level faithfulness scoring across four established benchmarks, complemented by manual validation. Their analysis reveals that human global judgments correlate significantly more strongly with the average of sentence-level assessments than with the conjunctive criterion, thereby demonstrating the presence of systematic Averaging Bias in widely used evaluation benchmarks.

Abstract

Evaluation of faithfulness of text summarization treats a model generated summary as faithful only if every of its sentences is supported by the source document: a strict conjunctive rule under which a single unsupported sentence makes the whole summary unfaithful. Yet most faithfulness benchmarks collect only one global human annotation label per summary. We ask whether such global human labels actually implement the conjunctive rule. We hypothesize that annotators may accept a summary as faithful when most sentences are faithful, not only when all are faithful. To test our hypothesis, we use five large language model (LLM) judges as per-sentence raters across four widely used faithfulness benchmarks. We find that global human labels correlate better with the average of per-sentence LLM judgments than with the implementation of the strict conjunctive rule. A manual review confirms that a substantial fraction of summaries labeled faithful by humans contain genuine local factual errors. We call this tendency Averaging Bias. Our results reveal that human labels on widely used faithfulness benchmarks contain measurable Averaging Bias, calling for carefully structured designs for trustworthy human annotations

Problem

• faithfulness
• human annotation
• averaging bias
• text summarization
• evaluation

Innovation

• Averaging Bias
• faithfulness evaluation
• human annotation
• text summarization
• LLM judges