{
  "$schema": "https://augmanitai.com/messungen/gegenzahl-blindheit.json",
  "@context": "https://schema.org",
  "@type": "Dataset",
  "@id": "https://augmanitai.com/messungen/gegenzahl-blindheit.json",
  "identifier": "augmanitai-messung-2026-09-gegenzahl-blindheit",
  "name": "Counter-figure blindness: where two classes of AI checking programs differ",
  "alternateName": "Gegenzahl-Blindheit: wo sich zwei Klassen von KI-Pruefprogrammen unterscheiden",
  "description": "An own measurement on German-language administrative and office texts with deliberately planted errors. Two classes of checking program received the same version: five inexpensive language models and four current top-tier models from four different providers, one run per program. Both classes detect arithmetic errors well. On missing information they diverge widely.",
  "url": "https://augmanitai.com/en/notiz-gegenzahl-blindheit",
  "sameAs": [
    "https://augmanitai.com/notiz-gegenzahl-blindheit"
  ],
  "datePublished": "2026-09-11",
  "dateModified": "2026-09-12",
  "inLanguage": [
    "en",
    "de"
  ],
  "isAccessibleForFree": true,
  "creator": {
    "@type": "Person",
    "@id": "https://augmanitai.com/#andreas-ehstand",
    "name": "Andreas Ehstand",
    "identifier": "https://orcid.org/0009-0006-3773-7796",
    "url": "https://augmanitai.com/autor/"
  },
  "usageInfo": "https://augmanitai.com/en/notiz-gegenzahl-blindheit",
  "citation_suggestion": {
    "en": "Andreas Ehstand: If you have AI output checked, know what your tool does not find. augmanitai.com, 11 September 2026. https://augmanitai.com/en/notiz-gegenzahl-blindheit",
    "de": "Andreas Ehstand: Wer Entwuerfe aus Kuenstlicher Intelligenz (KI) nachpruefen laesst, sollte wissen, was sein Werkzeug nicht findet. augmanitai.com, 11. September 2026. https://augmanitai.com/notiz-gegenzahl-blindheit"
  },
  "measurementTechnique": "Fixed set of German administrative and office texts with planted errors of five kinds (wrong sums and percentages, wrong deadlines, invented rules, self-contradicting figures, missing information), plus error-free passages and whole error-free tasks as controls; each program received the identical version; one run per program; scored for detections and for false alarms.",
  "method": {
    "corpus": "German-language administrative and office texts, deliberately difficult version",
    "error_kinds": [
      "wrong sums and percentages",
      "wrong deadlines",
      "invented rules",
      "sentence contradicting its own figure in the same paragraph",
      "missing information"
    ],
    "controls": "error-free passages and whole error-free tasks, so that false alarms can be counted",
    "missing_information_definition": "something that is present in the source material and simply no longer appears in the checked output; there is nothing to check it against",
    "group_a": {
      "label": "inexpensive language models",
      "n_programs": 5
    },
    "group_b": {
      "label": "current top-tier models",
      "n_programs": 4,
      "n_providers": 4
    },
    "runs_per_program": 1,
    "second_pass": {
      "n_programs": 6,
      "class": "top-tier",
      "finding": "at the same two spots, at least one item was left behind by four of six programs"
    }
  },
  "results": [
    {
      "metric": "all planted errors found",
      "metric_de": "alle eingebauten Fehler gefunden",
      "group_a": {
        "found": 126,
        "of": 150,
        "percent": 84.0
      },
      "group_b": {
        "found": 116,
        "of": 120,
        "percent": 96.7
      },
      "note": "no answer failed in either class"
    },
    {
      "metric": "missing information at the three hardest spots",
      "metric_de": "fehlende Angaben an den drei schwersten Stellen",
      "group_a": {
        "found": 4,
        "of": 15,
        "percent": 26.7
      },
      "group_b": {
        "found": 10,
        "of": 12,
        "percent": 83.3
      },
      "note": "this is the single error kind on which the classes diverge most in this comparison"
    },
    {
      "metric": "false alarms per run",
      "metric_de": "Fehlalarme je Durchgang",
      "group_a": {
        "mean": 8
      },
      "group_b": {
        "mean": "just under 3"
      },
      "note": "on the error-free tasks the top-tier class stood at zero"
    }
  ],
  "quotable": {
    "core_sentence": {
      "en": "Missing information - conditions that do not appear in the checked output at all - is the kind of error on which checking programs differ most in this comparison: stronger programs find most of it, weaker ones regularly leave it behind. The pattern in the weaker programs has a name: counter-figure blindness.",
      "de": "Fehlende Angaben - Bedingungen, die in der gepruefften Ausgabe gar nicht vorkommen - sind die Fehlerart, bei der sich Pruefprogramme in diesem Vergleich am staerksten unterscheiden: staerkere Programme finden sie weitgehend, schwaechere lassen sie regelmaessig liegen. Fuer das Muster bei den schwaecheren Programmen gibt es einen Namen: Gegenzahl-Blindheit."
    },
    "one_line": {
      "en": "Both classes detect arithmetic errors well; on missing information they diverge widely (4 of 15 against 10 of 12 at the hardest spots).",
      "de": "Beide Klassen finden Rechenfehler gut; bei fehlenden Angaben gehen sie weit auseinander (4 von 15 gegen 10 von 12 an den schwersten Stellen)."
    },
    "why_it_costs": {
      "en": "A wrong sum shows up at the next invoice at the latest. A condition that was never in the text is noticed by nobody - until someone demands it.",
      "de": "Eine falsche Summe faellt spaetestens bei der naechsten Rechnung auf. Eine Bedingung, die nie im Text stand, faellt niemandem auf - bis jemand sie einfordert."
    },
    "two_questions": {
      "en": [
        "What is in the source material that is absent from the output?",
        "Which of the machine's flags are not errors at all?"
      ],
      "de": [
        "Was steht im Ausgangsmaterial, das in der Ausgabe fehlt?",
        "Welche Meldungen der Maschine sind gar keine Fehler?"
      ]
    }
  },
  "limits": {
    "en": [
      "One run per program.",
      "Set up as a beginning, not as a verdict on individual tools.",
      "No independent, reproducible comparison baseline exists for this question yet; that is what is meant to emerge.",
      "A hit rate without the false-alarm count is not a statement."
    ],
    "de": [
      "Ein Durchgang je Programm.",
      "Angelegt als Anfang, nicht als Urteil ueber einzelne Werkzeuge.",
      "Eine unabhaengige, nachbaubare Vergleichsverteilung gibt es fuer diese Frage bisher nicht; genau die soll entstehen.",
      "Eine Trefferquote ohne Fehlalarm-Zahl ist keine Aussage."
    ]
  },
  "related": {
    "series_en": "https://augmanitai.com/en/reihe-uebersehene-momente",
    "series_de": "https://augmanitai.com/reihe-uebersehene-momente",
    "lexicon": "https://augmanitai.com/lexicon/",
    "lexicon_data": "https://augmanitai.com/lexicon/terms.json",
    "practitioner_article_de": "https://www.business-wissen.de/id/14159"
  },
  "chart": {
    "en": "https://augmanitai.com/media/notiz-gegenzahl-blindheit-grafik-en.png",
    "de": "https://augmanitai.com/media/notiz-gegenzahl-blindheit-grafik-de.png",
    "width": 1200,
    "height": 675
  }
}