Whose Data, Whose Benchmark? A User-Centric Framework for Auditing Evaluation Resources

arXiv:2610.19268v1 Announce Type: cross Abstract: Benchmarks shape how model capabilities are measured and interpreted, yet their scores can obscure whose conventions, contexts, and judgments define successful performance. We introduce a user-centric audit framework that distinguishes users who contribute benchmark content from those whose needs evaluation results are intended to represent. The framework operationalizes seven dimensions across benchmark content, metadata, documentation, and scoring procedures. Across 16 resources, we find that American-centered conventions account for 62.4-94.4% of matched signals in 11 text benchmarks, while geographic visibility in five image-text resources is concentrated and strongly associated with national GDP share (mean Pearson r = 0.780; Spearman p = 0.778). Documentation reveals a systematic source-target gap: all 16 resources operationalize source-user context, but none operationalize target-user context. Paired scoring experiments further show that meaning-preserving answer variants reduce DROP exact-match accuracy from 100.0% to 33.3%, while TruthfulQA exhibits metric-dependent reductions in reference similarity. Our framework and documentation checklist make user assumptions auditable, clarifying what benchmark scores represent and where evaluation claims are supported.

The item highlights biases in benchmarks, but implications for human flourishing are uncertain.

This is a methodological audit showing measurement bias in AI benchmarks, not a direct societal harm. The implication is moderately concerning: flawed evaluation standards could entrench inequities in whose needs AI serves well, but the paper itself is a corrective, evidence-strong contribution.

The text describes a systematic bias in AI benchmarks towards American conventions and wealthy nations. This bias can mislead evaluations of AI capabilities and their real-world applicability. The lack of representation of target users in benchmark documentation exacerbates this issue.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   8214b33a-33cd-4762-9798-cd65e0ede240
           https://arxiv.org/abs/2610.19268
   ↓
EVIDENCE   3 extracted excerpts
   ↓
MODEL RUN  3 runs, methodology 1.0.1
   ↓
SCORE      -62  (Adverse)
   ↓
CONFIDENCE 77%