{
  "reviewed": "2026-10-29",
  "measure": "Task time indexed within each experiment; no-AI baseline = 100; lower means less time",
  "comparability": "Different tools, tasks, populations, dates and estimators. These point estimates must not be pooled or treated as current universal effects.",
  "studies": [
    {
      "id": "copilot",
      "publication_year": 2023,
      "experiment_period": "May–June 2022",
      "participants_randomized": 95,
      "task": "Implement a JavaScript HTTP server",
      "no_ai_index": 100,
      "ai_index": 44.2,
      "reported_time_change_percent": -55.8,
      "reported_95_percent_interval_for_time_change": [-89, -21],
      "mean_minutes_with_ai_among_completers": 71.17,
      "mean_minutes_without_ai_among_completers": 160.89,
      "method": "44.2 = 100 × (1 − reported 55.8 / 100); normalized using the rounded published reduction, conditional on task completion.",
      "source": "https://arxiv.org/abs/2302.06590",
      "boundary": "A self-contained task; does not measure software lifecycle costs or establish a general code-quality gain."
    },
    {
      "id": "metr",
      "publication_year": 2025,
      "experiment_period": "February–June 2025",
      "participants": 16,
      "issues": 246,
      "task": "Resolve issues in familiar open-source repositories",
      "no_ai_index": 100,
      "ai_index": 119,
      "reported_time_change_percent": 19,
      "reported_95_percent_interval_for_time_change": [2, 39],
      "method": "119 = 100 × (1 + reported 19 / 100); represents the study's estimated time effect, rather than a pooled raw mean.",
      "source": "https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/",
      "boundary": "A specific early-2025 setting. The February 2026 update describes later measurement as an unreliable signal of the current effect.",
      "update": "https://metr.org/blog/2026-02-24-uplift-update/"
    }
  ],
  "conceptual_figures": {
    "portfolio": "Original architectural illustration of distinct experiences sharing capabilities; not a measured deployment or a claim that all products use all capabilities.",
    "implementation": "Original framework separating development method from runtime implementation; examples are illustrative."
  },
  "references": [
    {"id": "fatigue", "year": 2005, "url": "https://doi.org/10.1509/jmkr.2005.42.4.431", "supports": "Capability and usability tradeoffs before and after product use"},
    {"id": "subtraction", "year": 2021, "url": "https://doi.org/10.1038/s41586-021-03380-y", "supports": "Eight experiments on overlooking subtractive changes"},
    {"id": "nintendo", "url": "https://iwataasks.nintendo.com/interviews/wii/wii_remote/0/0/", "supports": "First-party design history used as an analogy"},
    {"id": "parnas", "year": 1972, "url": "https://doi.org/10.1145/361598.361623", "supports": "Modularity and information hiding"},
    {"id": "heart", "year": 2010, "url": "https://research.google/pubs/measuring-the-user-experience-on-a-large-scale-user-centered-metrics-for-web-applications/", "supports": "User-centered metrics and goals-to-signals-to-metrics process"},
    {"id": "agents", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents", "supports": "Practitioner guidance on workflow and agent complexity tradeoffs"}
  ]
}
