Task-level approaches reveal that low-skilled tasks face the largest displacement risk while high-skilled work is more likely to be augmented, but much of the literature rests on static, pre-2020 mappings and limited causal evidence that rapid LLM advances have already outpaced.

Main Finding

Recent methodologies in AI–labor economics have shifted from occupation-level to task-level analysis, increasingly using NLP and large language models to measure AI exposure and predict effects on employment, wages, and productivity. These methods reveal heterogeneous impacts across tasks and skill levels: lower-skilled tasks face higher displacement risk while higher-skilled workers are more likely to be augmented. However, many studies rely on static assumptions and theoretical models, limiting causal inference and making some pre-2020 conclusions about which professions will endure already outdated by 2025.

Key Points

• Task-level focus
  • Research is moving from treating occupations as homogeneous to mapping task-by-task exposure to AI (e.g., Felten et al., 2023; Eloundou et al., 2023).
  • Task approaches capture within-occupation heterogeneity in automation/augmentation risk.
• Use of advanced ML and LLMs
  • Methods increasingly apply NLP (BERT, LSTM) and LLMs (GPT-4) to parse job descriptions, map skills/tasks, and predict automation risk (Xu et al., 2025; Hampole et al., 2025).
• Scenario and policy-oriented modelling
  • Think tanks and international organisations (TBI, IPPR, IMF) emphasise scenario planning with differing adoption initial conditions to inform reskilling and labour-market policy (TBI, 2024; IPPR, 2024; Korinek, 2023).
• Distributional findings
  • Robust pattern: AI’s effects vary by skill level — displacement risk concentrated among lower-skilled tasks, augmentation and wage gains more likely for high-skilled tasks (Brynjolfsson et al., 2023; Chen et al., 2024).
• Methodological limitations
  • Many studies use static assumptions (fixed comparative advantage, no adaptation), limited causal identification, and model-dependent projections (critique of Acemoglu & Restrepo, 2022; Webb, 2020).
  • Rapid technological change (post-2020 advances) has already made some earlier profession-level conclusions obsolete by 2025.

Data & Methods

• Data sources (commonly used / implied)
  • Job postings and descriptions, occupational task databases (e.g., O*NET style), employer/household surveys, administrative payroll data, firm-level productivity measures.
• Measurement approaches
  • AI exposure indices: task-level scores indicating susceptibility to AI automation or augmentation (Felten et al., 2023; Eloundou et al., 2023).
  • NLP/LLM pipelines: use BERT/LSTM/Transformer models to extract tasks/skills from free-text job ads and to map tasks to AI capabilities (Xu et al., 2025; Hampole et al., 2025).
  • Scenario modelling: counterfactual simulations with different adoption speeds, policy responses, and initial conditions to bound possible employment/wage/productivity trajectories (TBI, IPPR, IMF).
• Analytical strategies and shortcomings
  • Cross-sectional exposure correlations and panel-difference analyses are common but often lack strong causal identification (endogeneity of adoption, unobserved confounders).
  • Static equilibrium/representative-agent models neglect dynamic reallocation, task re-bundling, and firm-level heterogeneity.
  • Rapid model and capability change (LLMs and multimodal AI) challenge out-of-sample validity of indices developed pre-2020.

Implications for AI Economics

• Research priorities
  • Prioritise causal, dynamic analyses: exploit natural experiments, firm-level adoption variation, and longitudinal microdata to identify causal effects on employment, wages, and productivity.
  • Update and validate task–AI mapping continuously using state-of-the-art LLMs and human-in-the-loop validation to keep exposure indices current.
  • Incorporate task reallocation and worker adaptation into models (reskilling, on-the-job learning, task switching).
• Policy implications
  • Design flexible reskilling and retraining programs targeted at high-risk tasks and low-skilled workers, informed by task-level exposure maps.
  • Use scenario planning to prepare labour-market institutions for a range of adoption speeds; policy should focus on smoothing transitions (unemployment insurance, portable benefits, active labour-market policies).
  • Consider incentives for firms to invest in augmentation (human+AI) rather than pure replacement where socially beneficial.
• Measurement and forecasting
  • Move away from occupation-level forecasts toward task-level, continuously updated indicators linked to real-world adoption measures (firm purchases, API usage, procurement).
  • Develop standards for documenting AI-capability changes and their mapping to tasks so conclusions remain transparent and updatable.
• Cautions for interpretation
  • Be wary of projections based on pre-2020 assumptions; rapid AI capability improvements (LLMs, multimodal models) alter which tasks are automatable.
  • Emphasise uncertainty and distributional outcomes — aggregate productivity gains can coincide with concentrated displacement without complementary policy.