{
  "schemaVersion": "1.0",
  "slug": "auditing-understanding",
  "title": "The opacity alibi: when “too complex” stops the questions",
  "url": "https://evidencepress.org/articles/auditing-understanding/",
  "standfirst": "A machine-checked result is not a lesson learned. Before declaring AI discoveries beyond human understanding, we should ask whether we have tested understanding fairly.",
  "summary": "An enormous proof, a deceptively simple switch puzzle and two identical-looking causal worlds reveal why checking a result and understanding it are different achievements.",
  "datePublished": "2026-10-30",
  "dateModified": "2026-10-30",
  "articleClass": "essay",
  "status": "published",
  "byline": "Evidence Press",
  "banner": {
    "src": "/assets/articles/auditing-understanding.svg",
    "alt": "A random signal controls X which controls Y, or controls Y which controls X. Observations agree, but forcing X on makes Y certainly on in the first and on half the time in the second.",
    "caption": "Constructed causal schematic, not experimental data. Both models produce matching lights; forcing X on distinguishes their predictions. The two answers are conditional on exactly these two models."
  },
  "audio": {
    "src": "/assets/audio/auditing-understanding.mp3",
    "transcript": "/assets/audio/auditing-understanding.txt",
    "provenance": "/assets/audio/auditing-understanding.provenance.json",
    "durationSeconds": 704.304,
    "voiceLabel": "OpenAI API synthetic voice (fable)"
  },
  "topics": [
    "AI-assisted science",
    "scientific understanding",
    "interpretability",
    "causal inference"
  ],
  "readingMinutes": 8,
  "wordCount": 1759,
  "newResearchClaims": false,
  "claimBoundary": "An accessible adaptation of The Opacity Alibi research prospectus, with standard constructed parity and causal-identification examples. No human study, frontier-model evaluation, demonstrated audit advantage or universal limit on human understanding is reported. Commentary does not imply external peer review or formal verification.",
  "sources": [
    {
      "citation": "Heule, Kullmann and Marek (2016), Solving and Verifying the Boolean Pythagorean Triples Problem via Cube-and-Conquer",
      "url": "https://arxiv.org/abs/1605.00723"
    },
    {
      "citation": "Doshi-Velez and Kim (2017), Towards a Rigorous Science of Interpretable Machine Learning",
      "url": "https://arxiv.org/abs/1702.08608"
    },
    {
      "citation": "Poursabzi-Sangdeh and colleagues (2021 revision), Manipulating and Measuring Model Interpretability",
      "url": "https://arxiv.org/abs/1802.07810"
    },
    {
      "citation": "Schut and colleagues (2023 preprint; published in PNAS 2025), Bridging the Human–AI Knowledge Gap",
      "url": "https://arxiv.org/abs/2310.16410"
    }
  ],
  "relatedReleases": [],
  "relatedArticles": [
    {
      "slug": "assurance-infrastructure",
      "relation": "Why a checked object and a justified scientific claim need separate evidence"
    }
  ],
  "corrections": [],
  "license": "CC0-1.0",
  "markdownUrl": "https://evidencepress.org/articles/auditing-understanding/index.md",
  "sourceUrl": "https://github.com/ipitchford/evidence-press-site/blob/main/articles/auditing-understanding/body.md",
  "editUrl": "https://github.com/ipitchford/evidence-press-site/edit/main/articles/auditing-understanding/body.md",
  "metadataEditUrl": "https://github.com/ipitchford/evidence-press-site/edit/main/articles/auditing-understanding/meta.json"
}