Models refuse harmful requests 3–18 percentage points more often when verbalizing eval awareness

{
  "doc": "aranguri-bloom-verbalized-eval-awareness-2026.md",
  "author": null,
  "context": "Quantified behavioral effect showing safety score inflation from eval awareness.",
  "enrichment": {
    "is_stale": true
  },
  "norm_label": "Models refuse harmful requests 3–18 percentage points more often when verbalizing eval awareness",
  "source_url": null,
  "graphify_id": "finding_refusal_increase",
  "source_file": "aranguri-bloom-verbalized-eval-awareness-2026.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/aranguri-bloom-verbalized-eval-awareness-2026/graph.json",
  "extracted_type": "finding",
  "graphify_file_type": "finding"
}