{
  "@context": {
    "@vocab": "http://schema.org/",
    "exposure": "http://schema.org/exposure",
    "laborMarketImpact": "http://schema.org/laborMarketImpact",
    "hasPart": {
      "@id": "http://schema.org/hasPart",
      "@type": "@id"
    },
    "mainEntity": {
      "@id": "http://schema.org/mainEntity",
      "@type": "@id"
    },
    "author": {
      "@id": "http://schema.org/author",
      "@type": "@id"
    },
    "publisher": {
      "@id": "http://schema.org/publisher",
      "@type": "@id"
    },
    "image": {
      "@id": "http://schema.org/image",
      "@type": "@id"
    },
    "datePublished": {
      "@id": "http://schema.org/datePublished",
      "@type": "http://www.w3.org/2001/XMLSchema#date"
    },
    "Question": "http://schema.org/Question",
    "Answer": "http://schema.org/Answer",
    "DefinedTerm": "http://schema.org/DefinedTerm",
    "DefinedTermSet": "http://schema.org/DefinedTermSet",
    "HowTo": "http://schema.org/HowTo",
    "HowToStep": "http://schema.org/HowToStep",
    "position": "http://schema.org/position",
    "VideoObject": "http://schema.org/VideoObject",
    "AudioObject": "http://schema.org/AudioObject"
  },
  "@type": "Article",
  "@id": "/appendix-labor-market-impacts-ai",
  "headline": "Appendix to “Labor market impacts of AI”",
  "datePublished": "2026-03",
  "author": {
    "@type": "Organization",
    "name": "Anthropic",
    "url": "https://anthropic.com"
  },
  "publisher": {
    "@type": "Organization",
    "name": "Anthropic",
    "url": "https://anthropic.com"
  },
  "image": "/images/page1_anthropic_labor_market_impacts_ai.png",
  "articleBody": "This appendix details the methodology and additional results related to measuring labor market impacts of AI, focusing on task-level exposure, job-level exposure, unemployment impacts, task granularity, and comparison of occupational exposure measures.",
  "hasPart": [
    {
      "@type": "DefinedTermSet",
      "@id": "/appendix-labor-market-impacts-ai#definedTerms",
      "name": "Defined Terms in Labor Market Impacts of AI",
      "description": "Key terms defined in the appendix to clarify methodology and concepts.",
      "hasDefinedTerm": [
        {
          "@type": "DefinedTerm",
          "name": "WorkUsage_t",
          "description": "Weighted work-related count for task t, combining observed work-related usage in Claude.ai and 1P API usage."
        },
        {
          "@type": "DefinedTerm",
          "name": "ClaudeWorkUsage_t",
          "description": "Count of tasks t in Claude.ai classified as work-related, restricted to work-related transcripts."
        },
        {
          "@type": "DefinedTerm",
          "name": "APIUsage_t",
          "description": "Observed usage in 1P API traffic for task t, indicating integration into production workflows."
        },
        {
          "@type": "DefinedTerm",
          "name": "AutoShare_t",
          "description": "Share of Claude.ai usage that is automative for task t."
        },
        {
          "@type": "DefinedTerm",
          "name": "Exposure of task (r̃_t)",
          "description": "Exposure of a task t calculated as an indicator of sufficient usage, theoretical AI capability, and automation factor."
        },
        {
          "@type": "DefinedTerm",
          "name": "Job-level exposure (R_o)",
          "description": "Weighted average exposure of tasks for job o, representing share of job performed or accelerated by AI."
        },
        {
          "@type": "DefinedTerm",
          "name": "IWA",
          "description": "Intermediate Work Activities, a higher-level categorization of tasks in O*NET."
        },
        {
          "@type": "DefinedTerm",
          "name": "DWA",
          "description": "Detailed Work Activities, a finer categorization of tasks in O*NET."
        },
        {
          "@type": "DefinedTerm",
          "name": "Eloundou β",
          "description": "AI capability score from Eloundou et al. (2023) measuring theoretical AI feasibility of tasks."
        },
        {
          "@type": "DefinedTerm",
          "name": "Coreweight",
          "description": "O*NET weighting scheme for tasks designated as core tasks, used in aggregation."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "/appendix-labor-market-impacts-ai#howToDefineExposure",
      "name": "How to Define AI Exposure at Task and Job Level",
      "description": "Steps to measure AI exposure starting from task usage data to job-level exposure.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Collect task usage data",
          "text": "Gather work-related task usage counts from Claude.ai and 1P API traffic, restricting to work-related transcripts."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Apply usage threshold",
          "text": "Only consider tasks with WorkUsage_t ≥ 100 or 0.0025% of traffic to filter out noise and rare tasks."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Calculate task exposure",
          "text": "Calculate task exposure r̃_t using theoretical AI capability (β_t) and automation factor (α_t) based on automative usage share."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Aggregate to job-level exposure",
          "text": "Compute job-level exposure R_o as weighted average of task exposures by fraction of time spent on each task."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "/appendix-labor-market-impacts-ai#howToAnalyzeEmploymentImpacts",
      "name": "How to Analyze Employment Impacts of AI Exposure",
      "description": "Methodology to analyze unemployment rates among workers with different AI exposure levels.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Define exposure groups",
          "text": "Designate workers in top quartile of AI exposure as high exposure and others as low exposure."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Compare unemployment rates",
          "text": "Compare unemployment rates of 22-25 year old workers in high and low exposure occupations over time."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Conduct difference-in-differences regression",
          "text": "Estimate changes in unemployment rates post-ChatGPT release controlling for pre-existing trends."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "/appendix-labor-market-impacts-ai#howToAddressTaskGranularity",
      "name": "How to Address Task Granularity in O*NET Data",
      "description": "Approaches to handle variability in task specificity in O*NET database.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Use higher-level task categories",
          "text": "Aggregate tasks to Intermediate Work Activities (IWAs) or Detailed Work Activities (DWAs) levels."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Group tasks by semantic similarity",
          "text": "Group tasks sharing the same IWA and with semantic similarity ≥ 0.7 to reduce redundancy."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Calibrate task specificity",
          "text": "Consider calibrating task specificity and interdependency for improved exposure measurement."
        }
      ]
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q1",
      "name": "What is the definition of AI exposure at the task level?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "AI exposure at the task level is defined by combining observed work-related usage in Claude.ai and 1P API traffic, applying a threshold of usage, and weighting by theoretical AI capability and automation share."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q2",
      "name": "How is job-level AI exposure calculated?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Job-level AI exposure is calculated as the weighted average of task-level exposures, weighted by the fraction of time workers spend on each task."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q3",
      "name": "What does the automation factor α_t represent?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The automation factor α_t upweights tasks that see relatively more automative uses, scaling coverage with the degree of automation."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q4",
      "name": "What are the main findings on unemployment impacts for young workers?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Unemployment rates for 22-25 year old workers in high AI exposure occupations have remained roughly constant compared to less exposed workers, with no significant increase post-ChatGPT release."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q5",
      "name": "How does task granularity affect exposure measurement?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Variability in task specificity can bias exposure measurement; grouping tasks by higher-level categories or semantic similarity can mitigate this issue."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q6",
      "name": "What alternative measures of occupational exposure were compared?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Several measures were compared including baseline observed coverage, Claude.ai usage, Eloundou et al. (2023) AI capability scores, baseline weighted by success rate, Ridge regression imputed usage, and aggregation at DWA and IWA levels."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q7",
      "name": "What is the significance of API usage in defining exposure?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "API usage indicates deeper integration into production workflows and is weighted heavily in defining task exposure."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q8",
      "name": "Why are some tasks grouped and allocated across jobs?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Because some tasks are shared across multiple occupations or differ only slightly in wording, they are grouped and task counts allocated equally across jobs based on employment shares."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q9",
      "name": "What data sources were used to analyze unemployment impacts?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Data from the Current Population Survey (CPS) and unemployment insurance claims aggregated by the Department of Labor were used."
      }
    },
    {
      "@type": "Question",
      "@id": "/appendix-labor-market-impacts-ai#q10",
      "name": "What is the role of the Eloundou et al. (2023) AI capability score?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It provides a theoretical measure of task feasibility with AI, used to identify tasks that are theoretically doable with LLMs or LLM plus tools."
      }
    }
  ]
}