{
  "description": "New arithmetic using published aggregate counts; not a new experiment or participant-level reanalysis",
  "source": {
    "citation": "Mai KT, Bray S, Davies T, Griffin LD (2023). Warning: Humans cannot reliably detect speech deepfakes. PLOS ONE 18(8): e0285333.",
    "article_url": "https://doi.org/10.1371/journal.pone.0285333",
    "table_3_url": "https://doi.org/10.1371/journal.pone.0285333.t003",
    "table_4_url": "https://doi.org/10.1371/journal.pone.0285333.t004",
    "retrieved_on_utc": "2026-10-17",
    "source_table_license": "CC BY 4.0; Mai et al. (2023)",
    "source_license_url": "https://creativecommons.org/licenses/by/4.0/"
  },
  "input_csv_sha256": "2de8c16c860506cd5fec67c4faa533b815dfe6a70b81ac203391b4ac218ce2a9",
  "single_clip_confusion_counts": {
    "genuine_called_genuine": 1761,
    "genuine_called_fake": 837,
    "fake_called_genuine": 681,
    "fake_called_fake": 1841
  },
  "paired_choice_counts": {
    "incorrect": 775,
    "correct": 4605
  },
  "metrics": [
    {
      "name": "overall_accuracy",
      "numerator": 3602,
      "denominator": 5120,
      "fraction": 0.703515625,
      "percent": 70.3515625,
      "meaning": "Correct / all single-clip judgments"
    },
    {
      "name": "fake_hit_rate",
      "numerator": 1841,
      "denominator": 2522,
      "fraction": 0.7299762093576526,
      "percent": 72.99762093576526,
      "meaning": "Called fake / judgments of fake clips"
    },
    {
      "name": "fake_miss_rate",
      "numerator": 681,
      "denominator": 2522,
      "fraction": 0.2700237906423473,
      "percent": 27.002379064234734,
      "meaning": "Called genuine / judgments of fake clips"
    },
    {
      "name": "genuine_specificity",
      "numerator": 1761,
      "denominator": 2598,
      "fraction": 0.6778290993071594,
      "percent": 67.78290993071595,
      "meaning": "Called genuine / judgments of genuine clips"
    },
    {
      "name": "genuine_false_alarm_rate",
      "numerator": 837,
      "denominator": 2598,
      "fraction": 0.32217090069284066,
      "percent": 32.21709006928407,
      "meaning": "Called fake / judgments of genuine clips"
    },
    {
      "name": "overall_error_rate",
      "numerator": 1518,
      "denominator": 5120,
      "fraction": 0.296484375,
      "percent": 29.648437500000004,
      "meaning": "Incorrect / all single-clip judgments"
    },
    {
      "name": "false_alarms_share_of_errors",
      "numerator": 837,
      "denominator": 1518,
      "fraction": 0.5513833992094862,
      "percent": 55.13833992094862,
      "meaning": "Genuine-as-fake errors / all single-clip errors"
    },
    {
      "name": "fake_share_of_test_judgments",
      "numerator": 2522,
      "denominator": 5120,
      "fraction": 0.492578125,
      "percent": 49.2578125,
      "meaning": "Artificial study mixture; not real-world prevalence"
    },
    {
      "name": "in_sample_fake_precision",
      "numerator": 1841,
      "denominator": 2678,
      "fraction": 0.6874533233756535,
      "percent": 68.74533233756534,
      "meaning": "Actually fake / called fake; study mixture only"
    },
    {
      "name": "in_sample_real_precision",
      "numerator": 1761,
      "denominator": 2442,
      "fraction": 0.7211302211302212,
      "percent": 72.11302211302211,
      "meaning": "Actually genuine / called genuine; study mixture only"
    },
    {
      "name": "paired_choice_accuracy",
      "numerator": 4605,
      "denominator": 5380,
      "fraction": 0.8559479553903345,
      "percent": 85.59479553903346,
      "meaning": "Correct fake selection / paired-choice trials"
    },
    {
      "name": "paired_choice_error_rate",
      "numerator": 775,
      "denominator": 5380,
      "fraction": 0.14405204460966542,
      "percent": 14.405204460966543,
      "meaning": "Incorrect fake selection / paired-choice trials"
    }
  ],
  "balanced_accuracy_percent": 70.3902654332406,
  "published_author_percentages_for_comparison": {
    "table_3_overall_accuracy": 70.35,
    "table_3_genuine_specificity": 67.78,
    "table_3_fake_hit_rate": 73.0,
    "table_4_paired_choice_accuracy": 85.59
  },
  "validation_checks": {
    "single_clip_total_matches_table_3": true,
    "genuine_total_matches_table_3": true,
    "fake_total_matches_table_3": true,
    "called_genuine_total_matches_table_3": true,
    "called_fake_total_matches_table_3": true,
    "paired_total_matches_table_4": true,
    "single_clip_accuracy_matches_published_rounding": true,
    "genuine_specificity_matches_published_rounding": true,
    "fake_hit_rate_matches_published_rounding": true,
    "paired_accuracy_matches_published_rounding": true
  },
  "limitations": [
    "Judgments repeat within participants and stimuli; these aggregate cells cannot support independence-based confidence intervals.",
    "Table 2 lists 260 single-clip participants; 20 stated trials each would imply 5200 judgments, whereas Tables 3 and 5 report 5120. The 80-judgment difference remains unexplained by the available aggregates.",
    "Language and familiarization conditions are pooled; no subgroup or participant-level inference is possible from these cells.",
    "OSF raw responses and their separate license were not verified in this retrieval run; only published CC BY article tables were used.",
    "Test-mixture precision and performance do not estimate real-world prevalence, caller authenticity, fraud success, or safeguard effectiveness.",
    "Paired choice is a distinct same-sentence, one-real/one-fake task, not an operational verification intervention."
  ]
}