About 80 percent of US workers could have at least 10 percent of their tasks affected by LLMs.
Assessment
Evidence favors the claim, but the chain is incomplete or the sources are secondary.
The figure comes from Eloundou, Manning, Mishkin and Rock's 2023 study "GPTs are GPTs" (arXiv:2303.10130, later published in Science), which had human annotators and GPT-4 rate tasks in the O*NET occupational database for whether access to a large language model could cut the time to complete them by half or more at equal quality. Aggregating those ratings, the authors estimate that about 80 percent of US workers are in occupations where at least 10 percent of tasks are exposed in this sense.
The estimate is credible as a statement of potential exposure, though it rests largely on one study's rubric and raters. In its favor, the authors report substantial agreement between human and GPT-4 ratings, and independently constructed exposure indices broadly agree on which occupations are most exposed, suggesting the scores capture a real signal; continued gains in model capability since 2023 would, if anything, raise the share of tasks affected. The main reservations are methodological: the claim turns on whether these task-level exposure ratings reliably identify what LLMs could affect, and a body of critique argues that zero-shot model judgments of task exposure amount to ungrounded model priors unless externally validated. The precise number is therefore softer than the qualitative finding that exposure is very widespread.
The claim measures what LLMs could technically touch, and it is common ground that this kind of exposure indicates technical potential, not realized job impact or displacement: the figure is not a forecast of adoption or job loss, and reading it as one goes beyond what the study asserts.
Full reasoning: the evidence and decisions behind this verdict
The claim restates the headline finding of Eloundou et al. 2023 (arxiv.org/abs/2303.10130): "around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs." The working paper's text clarifies the construction: 80 percent of workers belong to an occupation with at least 10 percent of its tasks rated exposed, where a task is exposed if LLM access could reduce its completion time by at least 50 percent at equal quality. The authors' own publication page (openai.com/index/gpts-are-gpts/) asserts the same figure, and the paper subsequently passed peer review at Science, which raises the weight given to its methods.
How the material subclaims weigh. The load-bearing premise is that human and GPT-4 exposure ratings of O*NET tasks reliably estimate which tasks LLMs could affect. The study reports high human/GPT-4 agreement, but this is internal validation; a methodological literature (e.g. the position paper "Jobs' AI Exposure Should Be Measured from Evidence, Not Model Priors", arxiv.org/html/2605.15474v1, and an ILO research brief on workers' AI exposure) argues that zero-shot LLM classifications are not valid measurement instruments without external grounding, and that expert or annotator judgments of capability can be systematically optimistic. This is why the verdict is supported rather than verified: the specific 80 percent statistic has not been independently replicated, and a threshold statistic ("at least 10 percent of tasks") is sensitive to task-level rating noise. Partially offsetting this, independently built exposure indices, such as Felten et al.'s ability-based AIOE, broadly agree with these scores on which occupations are most exposed, which is convergent validity for the rankings if not for the exact percentage. The framework premise that task-based analysis of occupational databases can meaningfully estimate AI exposure is the standard approach of this literature (Frey-Osborne, Felten et al., Webb) and is not seriously disputed as a first-order method, though its limitations are.
Instance stances: both recorded instances (the arXiv abstract and OpenAI's publication page) affirm; no source found in this pass asserts the negation. Critics dispute what exposure measures and how reliable the ratings are, not the figure as a within-framework estimate, so the claim is not contested in the §10 sense. Related estimates are consistent in direction: Pew (2023) finds 19 percent of US workers in the most-exposed jobs, matching the paper's companion 19-percent-at-50-percent finding, and the IMF's roughly-60-percent figure for advanced economies uses a broader AI definition and a different threshold.
Credence 0.75: the claim's hedged form ("about 80 percent", "could have") tolerates measurement error, capability gains since GPT-4 push exposure upward, and directional corroboration is good; the discount reflects single-source dependence on the rubric and annotations for the specific threshold statistic. What would change the conclusion: an independent replication with grounded task-level validation (raising toward verified), or evidence that the human/GPT-4 ratings systematically overstate task-level applicability enough to move large numbers of occupations below the 10 percent threshold (lowering toward contested or contradicted).
Decomposition
The claims this one rests on directly. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →Task-based analysis of occupational databases can meaningfully estimate occupations' exposure to AI automation ↗︎
- requiresa load-bearing premise: the parent is false without itsteward instructions →Exposure ratings of O*NET occupational tasks by human annotators and GPT-4 reliably estimate which work tasks LLMs could affect ↗︎
- assumesbackground the parent's framing takes as givensteward instructions →LLM task exposure measures technical potential, not realized labor-market impact or job displacement ↗︎
- supportsthis provides evidence for the parentsteward instructions →Independently constructed AI occupational exposure indices broadly agree on which occupations are most exposed ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs
Our findings reveal that around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs, while approximately 19% of workers may see at least 50% of their tasks impacted.
Our findings indicate that approximately 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of GPTs, while around 19% of workers may see at least 50% of their tasks impacted.
OpenAI's publication page for the "GPTs are GPTs" working paper, summarizing the paper's headline exposure findings in the authors' own voice.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Aug 11, 2026. Every judgment on this page is accompanied by a reasoning trace.