Eloundou β — Theoretical LLM Task Exposure
One-line
A task-level scoring of LLM theoretical capability used as input to most modern AI-exposure measures. β = 1 if an LLM alone can halve task time, 0.5 if LLM + tools can, 0 otherwise.
Origin
Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arXiv:2303.10130.
Constructed via a combination of human and GPT-4 annotation of all ~20,000 O*NET tasks. The framework distinguishes:
- Directly exposed (β=1). Task time can be reduced ≥50% by an LLM alone, with a 2,000-word input limit and no access to real-time facts.
- Exposed with tools (β=0.5). Task time can be reduced ≥50% with an LLM augmented by software for retrieval, image processing, etc.
- Not exposed (β=0). Time cannot be reduced ≥50% by LLM-based tooling.
Why this measure dominates the literature
- Granularity. Operates at task level (O*NET tasks), not occupation level, enabling weighted aggregation by employment.
- Capability-anchored. Independent of any particular firm’s deployment choices.
- Reproducible. Public ratings, established codebook.
- Pre-deployment. Compiled before mass LLM diffusion, so usable as a counterfactual measure.
What it gets wrong (or under-specifies)
- Theoretical, not realized. β=1 does not imply usage. Many β=1 tasks see no actual LLM use (theoretical-vs-observed-capability-gap).
- No automated/augmentative distinction. A task could be β=1 because an LLM speeds it up alone, but in practice usage is augmentative — the worker still does the work. The measure does not weight these differently.
- Static. Captures capability as of early 2023. Frontier model jumps since then are not reflected.
- No work-vs-leisure split. A task that is β=1 may see all its LLM use in non-work settings; the measure does not exclude this.
These limitations motivate observed-exposure-measure, which uses β as input but adds realized work-context usage as a second layer.
Distribution of Claude usage by β (per massenkoff-mccrory-labor-market-impacts-2026 Figure 1)
| Eloundou β | Share of observed Claude work-context usage |
|---|---|
| β = 1 | 68% |
| β = 0.5 | 29% |
| β = 0 | 3% |
97% of observed Claude work-context usage falls on tasks Eloundou et al. classified as theoretically feasible. This is the strongest empirical validation of the β framework to date — what theory said LLMs could do, real users actually do.
Empirical predictive power
Standalone — limited. The Eloundou β alone is not correlated with BLS 2024–2034 projected occupational growth. It becomes predictive only when weighted by realized usage (see observed-exposure-measure and ai-exposure-vs-bls-growth).
This is the headline methodological finding of the Massenkoff–McCrory paper: theoretical capability without deployment is invisible in labor-market forecasts. Usage data is the missing layer.
Related
- observed-exposure-measure — the realized-usage-weighted measure that builds on β.
- theoretical-vs-observed-capability-gap — the gap between β and observed deployment.
- massenkoff-mccrory-labor-market-impacts-2026 — primary application of β in this wiki.
- acemoglu-simple-macroeconomics — earlier macro analysis using a related task-exposure approach.
- task-based-framework — Acemoglu’s macro frame that consumes task-exposure measures.
- ai-task-exposure-decomposition — Acemoglu’s task-incidence decomposition.
- ai-exposure-vs-bls-growth — empirical test of β-only vs usage-weighted exposure.
- diffusion-vs-innovation — the macro frame for why theoretical-only measures under-perform empirically.