{
  "lesson": "27-data-provenance-training-governance",
  "title": "Data Provenance and Training-Data Governance",
  "questions": [
    {
      "stage": "pre",
      "question": "Why is the compliance window for training-data governance at collection time, not after training?",
      "options": [
        "Once data enters model weights, surgical erasure is not possible; retraining from scratch is the only complete remediation and is prohibitively expensive",
        "Because GDPR right-to-erasure does not apply to AI",
        "Because regulators do not audit deployed models",
        "Because cookie-consent frameworks ban post-collection use"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Roughly how many statutory items does California AB 2013 require developers to publish in a dataset summary?",
      "options": [
        "12",
        "3",
        "7",
        "30"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which AB 2013 item is new relative to the Gebru et al. 2018 datasheet structure?",
      "options": [
        "Sources or owners of datasets",
        "Time period during which data was collected",
        "Whether the system uses or continuously uses synthetic data generation",
        "Cleaning, processing, or other modifications"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is the convergent 2025 DPA principle for LLM training on publicly available first-party content?",
      "options": [
        "Training is prohibited under all circumstances",
        "Legitimate interest can justify training on publicly available first-party content with opt-out and appropriate safeguards; consent is not required",
        "Explicit user consent is always required",
        "Only synthetic data may be used"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which of these is NOT a partial remediation for the irreversibility problem listed in the lesson?",
      "options": [
        "Unlearning (approximate removal measured by MIA)",
        "Influence-function-based localization and selective update",
        "GDPR-mandated automatic weight deletion",
        "Fine-tune-suppression of outputs derived from the data"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What does the Data Provenance Initiative's 'Consent in Crisis' (July 2024) find?",
      "options": [
        "Consent rates are stable across all sources",
        "Synthetic data has fully replaced web data",
        "Robots.txt restrictions are decreasing year over year",
        "Publishers are adding robots.txt restrictions at an accelerating rate; the openly-trainable-upon commons is contracting rapidly, with around 25% of top training sources adding restrictions 2023 to 2024"
      ],
      "correct": 3,
      "explanation": ""
    }
  ]
}