{
  "recordVersion": "1.0",
  "title": "Hidden characters in 10,000 AI responses",
  "performedAt": "2026-08-12",
  "publishedAt": "2026-08-13",
  "method": {
    "scannerRuleset": "literal-unicode-1",
    "script": "scripts/scan-generated-corpus.ts",
    "syntheticMarkersInserted": false,
    "responseTextIncludedInThisRecord": false,
    "description": "Unmodified model responses were scanned with the same deterministic Unicode classification rules used by the browser checker. The scan recorded exact supported code points and did not use an AI-authorship classifier."
  },
  "corpus": {
    "responses": 10000,
    "characters": 13719843,
    "words": 2107786,
    "modelLabels": 18,
    "sources": [
      {
        "name": "Google DeepMind SynthID Text human evaluation data",
        "url": "https://github.com/google-deepmind/synthid-text/blob/main/data/human_eval.jsonl",
        "inputSha256": "946bee15531384a8d8cdc75afffe44eda660aa0509e654c936f573093dbd1052",
        "selection": "All 3,000 paired rows",
        "groups": [
          { "label": "watermarked", "responses": 3000 },
          { "label": "unwatermarked", "responses": 3000 }
        ]
      },
      {
        "name": "OpenBMB UltraFeedback",
        "url": "https://huggingface.co/datasets/openbmb/UltraFeedback",
        "selection": "Four completions from each of the first 1,000 rows",
        "groups": [
          { "label": "generated", "responses": 4000 }
        ]
      }
    ]
  },
  "results": {
    "responsesWithSupportedFindings": 15,
    "responseFindingRate": 0.0015,
    "supportedFindingCount": 38,
    "groups": [
      { "label": "Google watermarked Gemma", "responses": 3000, "responsesWithFindings": 0, "findingCount": 0 },
      { "label": "Google unwatermarked Gemma", "responses": 3000, "responsesWithFindings": 0, "findingCount": 0 },
      { "label": "Other generated responses", "responses": 4000, "responsesWithFindings": 15, "findingCount": 38 }
    ],
    "findings": [
      { "codePoint": "U+FE0F", "name": "Emoji variation selector", "count": 28, "responses": 13, "cleanupAction": "review-preserve" },
      { "codePoint": "U+200D", "name": "Zero-width joiner", "count": 7, "responses": 3, "cleanupAction": "review-preserve" },
      { "codePoint": "U+00A0", "name": "No-break space", "count": 2, "responses": 1, "cleanupAction": "normalize" },
      { "codePoint": "U+2061", "name": "Mathematical function application", "count": 1, "responses": 1, "cleanupAction": "review-preserve" }
    ]
  },
  "interpretation": [
    "Supported literal Unicode findings were uncommon in these corpora.",
    "The Google watermarked and paired unwatermarked groups both contained zero supported literal findings.",
    "This experiment did not run Google's keyed SynthID detector and therefore does not measure SynthID detection accuracy.",
    "A zero result from the browser checker does not prove that text is human-written or free of provider-only statistical watermarks."
  ]
}