Little evidence of steganography between AV and AR; meaning-preserving transformations cause only small FVE drops.

{
  "doc": "natural.md",
  "context": "NLA explanations appear to encode information transparently in natural language rather than hidden channels.",
  "category": "ai",
  "norm_label": "Little evidence of steganography between AV and AR; meaning-preserving transformations cause only small FVE drops.",
  "graphify_id": "steganography_evaluation",
  "source_file": "natural.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/natural/graph.json",
  "extracted_type": "finding",
  "source_location": "§Measuring behavioral properties of NLAs",
  "graphify_file_type": "finding"
}