Large language models are concentrated in relatively secure, low-precarity occupations: occupations classified as low-precarity have a mean LLM exposure of 0.386 (95% CI 0.356–0.417) versus 0.205 (95% CI 0.136–0.275) for very-high-precarity jobs, suggesting LLMs are more likely to affect workers previously sheltered from technological disruption.

Main Finding

Occupations with the lowest measured precarity have the highest estimated exposure to large language models (LLM). Using Canadian Labour Force Survey data (2021–2024) and an O*NET‑based LLM‑exposure score, the authors find mean occupational LLM exposure = 0.34 (range 0–0.84). By a multidimensional precarity index, occupations classified as “low” precarity had a mean LLM exposure of 0.386 (95% CI 0.356–0.417) versus 0.258 (0.221–0.295) for “moderate”, 0.260 (0.194–0.328) for “high”, and 0.205 (0.136–0.275) for “very high” precarity. Apart from earnings adequacy, each separate precarity dimension (temporary employment, schedule unpredictability, working‑time mismatch) was associated with lower LLM exposure.

Key Points

• Research question: Are occupations with greater precarity more or less exposed to LLM‑amenable tasks?
• Precarity operationalized along four LFS dimensions: temporary employment, low wages (earnings inadequacy), irregular hours (schedule unpredictability), and involuntary part‑time (working‑time mismatch). A multidimensional index counts how many dimensions are in the top quartile of exposure.
• LLM exposure measure: adopted from Eloundou et al. (ONET task‑level coding of share of tasks that could realize ≥50% time savings via LLMs). Crosswalked ONET occupations to Canadian NOC codes to produce occupation‑level scores for 512 Canadian occupations.
• Analytical approach: pooled cross‑sectional occupational‑level analysis (2021–2024 LFS), four multivariate linear regressions (one per precarity dimension) plus a fifth on the multidimensional index; models adjusted for occupational composition (gender, age, education), industry (NAICS) and province; cluster‑robust SEs used to address heteroscedasticity.
• Main adjusted result: occupations with higher prevalence of temporary work, irregular hours, and involuntary part‑time work have significantly lower mean LLM exposure; earnings inadequacy showed a different (non‑conforming) pattern.
• Sensitivity check: reclassification by tertiles produced similar patterns.
• Data notes: analysis excludes territories, self‑employed, Indigenous reserves, full‑time armed forces, institutionalized/remote populations. LFS data are weighted and multiple imputation procedures are applied by Statistics Canada.

Data & Methods

• Data source: Statistics Canada Labour Force Survey (pooled months 2021–2024), occupational‑level analysis using 4‑ or 5‑digit NOC codes (512 occupations analyzed).
• Precarity variables (from LFS items):
  • Temporary employment (permanent vs temporary)
  • Low wages (hourly pay < 2/3 median)
  • Irregular hours (paid hours vary week‑to‑week)
  • Involuntary part‑time (work <30 hrs/week but want ≥30 hrs/week)
• Precarity coding: for each dimension, occupations ranked by proportion of workers exposed and grouped into quartiles (Q1 lowest exposure … Q4 highest); multidimensional index = count of times an occupation was Q4 across the four dimensions, categorized into low/moderate/high/very high overall precarity.
• LLM exposure: Eloundou et al. rubric applied to ONET task descriptors → proportion of tasks per occupation amenable to ≥50% time savings via LLM (continuous 0–1). Crosswalked ONET → Canadian NOC via established concordances.
• Models: multivariate linear regressions (LLM exposure score as outcome) controlling for occupational composition (gender, age groups, education), industry sector, and province; cluster‑robust SEs (512 clusters); diagnostics for heteroscedasticity, residual normality, leverage; sensitivity analyses with tertiles.
• Key summary statistics: mean occupational LLM exposure = 0.34 (min 0, max 0.84). Supplementary material provides occupation‑level scores.

Implications for AI Economics

• Distributional impacts: LLM exposure is concentrated in occupations with lower measured precarity (higher wages/education, more stable schedules). Early productivity or task‑augmentation gains from LLM are therefore likely to accrue to better‑paid, less precarious workers—potentially increasing within‑ and between‑occupation inequality unless redistributed.
• Complementarity vs substitution: the pattern is consistent with LLMs being complementary to cognitive/knowledge tasks in higher‑quality jobs rather than immediate substitutes in more precarious, routine work. Economic models should account for task‑level complementarities that favour skilled workers.
• Dynamics and diffusion risks: initial concentration in less precarious occupations does not preclude later diffusion to more precarious jobs (via task decomposition, tool bundling, or cost reductions), which could alter job quality and precarity over time. Dynamic models of technology diffusion and endogenous task reallocation are needed.
• Labor demand, wages and bargaining: if LLMs raise productivity mainly in higher‑quality occupations, aggregate wage effects could be heterogeneous—magnifying returns to skills and possibly compressing demand in intermediate tasks. Researchers should estimate wage and employment elasticities to LLM adoption at worker and firm levels.
• Health and social outcomes: because precariousness maps onto health risks, differential exposure to LLMs could indirectly affect worker health inequities. Analyses linking AI exposure to injury, mental health, and access to benefits are warranted.
• Measurement and policy: occupation‑level exposure is a coarse measure of potential impact. Policy should distinguish exposure from adoption and actual task automation/complementation. Policymakers should consider:
  • Targeted upskilling and lifelong learning programs for occupations likely to be affected.
  • Strengthening social protections for precarious workers to mitigate downstream distributional harms.
  • Monitoring adoption across sectors and occupations, with attention to task reallocation and managerial use of LLMs that may alter scheduling, surveillance, or casualization.
• Research priorities for AI economics:
  • Worker‑level longitudinal studies linking measured LLM adoption to employment, wages, hours, and health outcomes.
  • Firm‑level analyses of adoption decisions, complementarities between LLMs and human capital, and heterogeneity by firm size/industry.
  • Task‑level studies mapping how LLMs change task bundles and create new complementarities/substitutions across occupations.
  • Distributional models that incorporate diffusion timing, task reallocation, and bargaining/policy responses.
  • Estimation of causal effects (instrumental variables, difference‑in‑differences around firm or industry adoption events).

Limitations to bear in mind when using these results for economic modelling: occupational (ecological) analysis—no worker‑level causal claims; LLM exposure derived from US O*NET task coding and crosswalked to Canada (potential misclassification); exposure ≠ adoption or realized productivity gains; sample exclusions (self‑employed, territories, some Indigenous populations).