{
  "version": 1,
  "updatedAt": "2026-08-21",
  "description": "Reusable research protocols for testing human-AI decision patterns without treating working labels as established cognitive biases.",
  "canonicalUrl": "https://cognitive-biases.github.io/ai-era/protocol/",
  "rules": [
    "Start from an observable behaviour, not a catchy label.",
    "State the competing explanation before collecting results.",
    "Change one important variable at a time where possible.",
    "Keep model, version, prompt, settings and date.",
    "Separate model behaviour from human response.",
    "Report null or mixed results as clearly as positive results."
  ],
  "protocols": [
    {
      "slug": "ai-advice-order",
      "title": "AI Advice Order Test",
      "track": "ai-anchoring-loop",
      "question": "Does seeing an AI estimate before making your own estimate pull the final judgment toward the model?",
      "conditions": [
        "Independent estimate first, then AI advice",
        "AI advice first, then human estimate"
      ],
      "primaryMeasure": "Absolute shift toward the AI estimate",
      "secondaryMeasures": [
        "confidence change",
        "decision time",
        "distance from a reference-class estimate"
      ],
      "minimumReport": "Report the model/version, task, anchor value, sample size, mean shift by condition and uncertainty.",
      "preregisteredStudyId": "ai-advice-order-v1",
      "studyUrl": "/research/ai-advice-order-v1/",
      "instrumentUrl": "/experiments/ai-advice-order-v1/",
      "stimulusNote": "The v1 pilot isolates advice-order effects with standardized AI-labelled numerical recommendations fixed by protocol rather than live model calls."
    },
    {
      "slug": "agreement-pressure",
      "title": "AI Agreement Pressure Test",
      "track": "sycophancy-reinforcement-loop",
      "question": "Does signalling a preferred answer change model agreement and then increase human confidence in that preferred answer?",
      "conditions": [
        "Neutral question",
        "Question containing a preferred conclusion",
        "Preferred conclusion plus explicit request for counterevidence"
      ],
      "primaryMeasure": "Change in model agreement rate and user confidence",
      "secondaryMeasures": [
        "factual accuracy",
        "counterargument quality",
        "answer revision rate"
      ],
      "minimumReport": "Keep the underlying task identical across conditions and report both model behaviour and human confidence separately."
    },
    {
      "slug": "source-memory",
      "title": "Human-AI Source Memory Test",
      "track": "source-memory-blur",
      "question": "Do mixed human-AI workflows make it harder to remember who produced an idea or sentence?",
      "conditions": [
        "Human-only creation",
        "AI-only suggestion",
        "Mixed human-AI creation"
      ],
      "primaryMeasure": "Correct source-attribution rate after a delay",
      "secondaryMeasures": [
        "confidence in attribution",
        "verbatim recognition",
        "idea ownership judgment"
      ],
      "minimumReport": "Predefine the delay, keep contribution labels hidden during the test, and distinguish idea-source from wording-source errors."
    },
    {
      "slug": "offloading-retention",
      "title": "AI Offloading & Retention Test",
      "track": "cognitive-offloading-debt",
      "question": "Can AI improve immediate completion while weakening unaided retention or transfer?",
      "conditions": [
        "Unaided work",
        "AI-assisted work",
        "AI-assisted work plus retrieval/explanation step"
      ],
      "primaryMeasure": "Delayed unaided retention or transfer performance",
      "secondaryMeasures": [
        "immediate task quality",
        "time on task",
        "self-rated effort",
        "confidence"
      ],
      "minimumReport": "Do not infer long-term learning from immediate task quality. Include a delayed or transfer measure."
    }
  ],
  "startingEvidence": [
    {
      "title": "How was my performance? Exploring the role of anchoring bias in AI-assisted decision making",
      "year": 2025,
      "url": "https://doi.org/10.1016/j.ijinfomgt.2025.102875"
    },
    {
      "title": "Anchoring bias in large language models: an experimental study",
      "year": 2025,
      "url": "https://doi.org/10.1007/s42001-025-00435-2"
    },
    {
      "title": "Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy",
      "year": 2026,
      "url": "https://doi.org/10.18653/v1/2026.acl-long.1126"
    },
    {
      "title": "Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models",
      "year": 2026,
      "url": "https://doi.org/10.18653/v1/2026.acl-long.1421"
    },
    {
      "title": "Large Language Models are overconfident and amplify human bias",
      "year": 2025,
      "url": "https://arxiv.org/abs/2505.02151"
    },
    {
      "title": "The AI Memory Gap: Users Misremember What They Created With AI or Without",
      "year": 2025,
      "url": "https://arxiv.org/abs/2509.11851"
    },
    {
      "title": "ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention",
      "year": 2025,
      "url": "https://doi.org/10.1016/j.ssaho.2025.102287"
    }
  ]
}