Start with tasks, because jobs are bundles
A task is one unit of work with a result, such as reconciling invoices. A job is a bundle of tasks, and most of the research works at that level. The O*NET job duties search draws on “over 19,000 occupation-specific task statements”, and the Eloundou paper applies its rubric to O*NET occupational data.
A single percentage for an occupation does not show which tasks produce it. For a task-by-task pass on a real posting, use the Job Description AI Impact Analyzer. The 15-minute audit guide covers the manual version.
What exposure means, and what it does not
In the Eloundou et al. paper (arXiv v5, August 2023), exposure is “a proxy for potential economic impact without distinguishing between labor-augmenting or labor-displacing effects.” It measures what could change, not whether the change helps or replaces a worker.
The paper’s direct-exposure test is a time test. A task counts as directly exposed when an LLM could “decrease the time required to complete the DWA or task by at least half (50%).” Human annotators and GPT-4 applied the rubric, so the result records judgements about possible savings, not timed measurements.
Anthropic’s June 2026 Economic Index report separates observed exposure, which counts tasks already seen being done with Claude, from theoretical exposure. The report calls the second “an upper bound on what is possible instead of a measure of current use.” A task can score high on one and low on the other.
Automation and augmentation, as the sources define them
The sources do not share one vocabulary, so check which definition a number uses. Anthropic’s automation share counts conversations whose pattern is directive, such as “translate this document”, or a feedback loop, such as “edit this email… make it more casual”. It measures how people use the tool, not which tasks a model could take over.
The ILO Working Paper 96 means something different by augmentation: change inside the job. It says the most important effect is likely to be “augmenting work”, meaning “automating some tasks within an occupation while leaving time for other duties”. It reports augmentation potential as the greater of its two measures.
The economist David Autor makes the same point in task terms. In his 2015 review, he writes that “Automation does indeed substitute for labor”, and that automation also complements labor. A substituted task therefore does not settle how many jobs exist.
Anthropic’s June 2026 text gives no overall automation-versus-augmentation percentage. Its discussion of one table in Section 2 says the pattern “looks more labor-augmenting than labor-displacing.” The September 2025 edition reported 77% of business API uses with automation patterns, against about 50% for Claude.ai users. That figure is dated, not current.
What the published evidence says, with dates
Each figure belongs to one edition and one measure, so give its date with it and do not add rows together.
| Source, edition and measure | Figure as printed | Limit the source states |
|---|---|---|
| Eloundou et al., arXiv v5, August 2023 Potential exposure of US worker tasks to LLMs | About 80% of workers could have at least 10% of tasks affected; about 19% at least half | No predictions about the adoption timeline |
| Eloundou et al., same edition Tasks that could be done significantly faster | About 15% of US worker tasks; 47–56% with LLM-powered software and tooling | Same level of quality; potential, not realized savings |
| Anthropic Economic Index, June 26, 2026 Observed and theoretical exposure; usage patterns | No overall automation-versus-augmentation split in the extractable text | Theoretical exposure is an upper bound |
| Anthropic, September 15, 2025 (older edition) Automation patterns, business API versus Claude.ai | 77% of business API uses vs about 50% of Claude.ai users | September 2025 data only |
| ILO Working Paper 96, August 2023 Automation and augmentation potential, share of total employment | Automation 0.4% (low-income) and 5.5% (high-income); augmentation 10.4% and 13.4% | Upper-bound exposure; infrastructure not considered |
| ILO Working Paper 166, March 2026 Average GenAI exposure by income group | About 32% of employment in high-income and about 15% in low-income economies | Realized augmentation depends on digital access |
| WEF Future of Jobs Report 2025, January 2025 Employer projections for 2025–2030 | 170 million new jobs (14%); 92 million displaced (8%); net 78 million (7%) | Survey-based projections, not observed outcomes |
Two studies that are easy to overstate
The Stanford paper “Canaries in the Coal Mine?” (Brynjolfsson, Chandar and Chen, revised August 12, 2026) reports an early-career gap. Employment of workers aged 22 to 25 in AI-exposed occupations is 19% below its kept-pace counterfactual, using ADP payroll data through June 2026. The paper also says “We find no evidence of widespread, economy-wide job displacement”, and that the gap works “primarily through reduced hiring of young workers rather than increased separations.”
The authors call the results “early, descriptive indicators” and “rather than causal estimates”, and say the patterns “attenuate when controlling for education.” Repeating the gap as an effect of AI goes beyond what they claim.
NBER working paper 31161 (Brynjolfsson, Li and Raymond, revised November 2023) studied 5,179 customer-support agents as a generative AI assistant was introduced in stages. Issues resolved per hour rose 14% on average, and 34% for novice and low-skilled workers, with “minimal impact on experienced and highly skilled workers.” That is a productivity result in one support setting, not a count of jobs.
Before you repeat a number like these, check its wording in the Claim–Evidence Map and the guide to weak evidence.
Task features that push toward automated, augmented or human-led work
The table lists features the sources describe, plus cues that this site’s analyzer uses. The basis column says which is which. No row is a measured rule for a specific job.
| Feature in the task | Leans toward | Basis |
|---|---|---|
| Directive request, such as translating a document | Leans automated (this guide’s reading) | Basis: this guide’s heuristic. Anthropic’s automation share describes how people used Claude, not what a model could do with the task. |
| Routine clerical work on structured data | Automated or augmented (this guide’s reading) | Figure: ILO WP96 finds 24% of clerical tasks highly exposed, and 1–4% in other broad groups. Exposure gives no direction, so the lean is this guide’s reading. |
| Information work with a clear goal and a checkable output | AI-assisted (this guide’s reading) | Source: Tomlinson et al. find that the most common and successful AI-assisted work activities involve information work. The goal and checkability criteria are this guide’s reading. |
| Could take at least half less time with AI | Exposed (time test) | Source: Eloundou et al. direct-exposure test; it measures time, not who keeps the work |
| Final sign-off, a licence or legal responsibility | Human-led (capped at augmented) | This site’s heuristic: accountability phrases such as “signs off” |
| In person, hands-on or on site | Human-led | This site’s heuristic: presence cues |
| New, ambiguous or strategic work | Human-led or uncertain | This site’s heuristic; the ILO’s “big unknown” is the source-based version |
Invented example, checked 18 Oct 2026: take the task “reconcile supplier invoices against purchase orders weekly.” The analyzer reads a routine verb (+2), a records object (+1) and the cue weekly (+1), which gives band 4, Could be automated. Add “and sign off the discrepancies” and the accountability phrase caps the duty at band 2, Augmented, so a person keeps the decision. Record it as partly routine and partly accountable.
What exposure scores cannot tell you
An exposure score answers one narrow question: which tasks could AI speed up or carry, according to a rubric or a usage study. It cannot say when change will happen, whether a job will still exist, or what happens to the people who do it.
- Timing. Eloundou et al. write, “We do not make predictions about the development or adoption timeline of such LLMs.”
- Direction. Exposure does not separate augmentation from displacement. The ILO calls its middle group “the big unknown”, about 16.2% of global employment, and says its data “do not allow for a clear-cut classification of this group.”
- Access. The ILO’s 2026 paper attributes to the authors’ earlier Latin America research the finding that nearly half of the jobs that could benefit from augmentation cannot realize it, because of “a lack of access to basic digital technologies at work.”
- Ceiling. The ILO calls its exposure figures “an upper-bound estimate of exposure”.
Common mistakes. Do not read a percentage of tasks as a percentage of jobs lost. Do not treat exposure and automation as one measure. Do not apply the 14% customer-support gain to every role, and do not present a 2023 figure as current. For a check on a real posting, compare the analyzer’s reasons with your own reading of each duty.
Common questions
Does a high AI exposure score mean a job will be automated?
No. Exposure describes potential: the share of tasks that could be sped up or affected. Eloundou et al. say exposure does not separate augmenting from displacing effects, and they make no timeline predictions. A high score points to tasks worth examining, not to a verdict on a job.
How do exposure and automation differ?
Exposure asks whether AI could affect or speed up a task. Automation describes how AI is actually used on it. Anthropic’s automation share counts directive or feedback-loop conversations. A task can be exposed without being automated in practice, which is why the June 2026 report separates observed use from theoretical exposure.
Does augmentation mean my job is safe?
No source here says that. The ILO says the most important effect is likely to be augmentation: AI automating some tasks within an occupation while people keep other duties. That describes change inside the work, not job security.
Where do the 80% and 19% figures come from?
They come from Eloundou et al. (arXiv v5, August 2023). About 80% of US workers could have at least 10% of their tasks affected by LLMs, and about 19% could have at least half affected. They describe tasks affected, not jobs lost, and they carry no timeline.
Does AI reduce jobs for young workers?
The Stanford study finds a descriptive gap, not a proven cause. Employment of workers aged 22 to 25 in AI-exposed occupations was 19% below its kept-pace counterfactual on June 2026 data, with no comparable gap for experienced workers.
Can I apply the 50% time test to my own tasks?
You can ask it of each task: could AI cut the time for this task by at least half? That is the paper’s direct-exposure question. The paper’s rubric was applied by annotators and GPT-4, not by timing the work. Treat your answers as estimates, and write down what each one rests on.
Sources and further reading
- Eloundou et al., “GPTs are GPTs” — arXiv v5, August 2023
- Anthropic Economic Index report — June 26, 2026
- Anthropic, “Uneven geographic and enterprise AI adoption” — September 15, 2025
- ILO, “Generative AI and jobs”, Working Paper 96 — August 2023
- ILO, “Disruption without dividend”, Working Paper 166 — March 2026
- Stanford Digital Economy Lab, “Canaries in the Coal Mine?” — revised August 12, 2026
- NBER, “Generative AI at Work”, working paper 31161 — revised November 2023
- Tomlinson et al., “Working with AI”, arXiv 2507.07935 — v6, December 2025
- World Economic Forum, “The Future of Jobs Report 2025” — January 2025
- Autor, “Why Are There Still So Many Jobs?” — Journal of Economic Perspectives, Summer 2015