Exposure, automation and replacement are three different things
Short answer
A high score cannot tell you whether you will lose your job. Exposure means your work overlaps with what AI gets used for. Automation means a task is actually handed over. Replacement means a person is no longer employed. These are three separate questions, measured by three different kinds of evidence, and the evidence gets thinner at every step. Exposure is the best measured of the three, and it is still an estimate. Replacement, through the first quarter of 2026, was barely detectable in United States labour data, although the same researchers who report that null result also flag warning signs and say effects may yet appear.
Three words that are not synonyms
| Term | What it claims | How it is measured |
|---|---|---|
| Exposure | Your tasks overlap with what AI is used for or could plausibly do | Usage logs, or expert and model ratings of task descriptions |
| Automation | A specific task is actually handed to a machine | Almost nothing measures this directly at scale; usage patterns are used as a proxy |
| Replacement | A person is no longer employed | Payroll and labour force data, months or years after the fact |
Every widely quoted "AI and jobs" number sits in the first row. The public conversation usually reports it as if it belonged in the third.
The people who built the measures say so themselves
Eloundou and colleagues write that they define exposure "as a proxy for potential economic impact without distinguishing between labor-augmenting or labor-displacing effects", and that they make no predictions about adoption timelines.1 The International Monetary Fund's exposure index is described as agnostic about whether exposure means a job is helped or replaced.2 The Microsoft study behind this site's scores warns against reading its number as a probability of job loss, and its own classifications are model estimates rather than direct observation, which reading your score quantifies.3
What usage data says about how AI is actually used
Anthropic publishes a running measurement of how people use Claude, split into augmentation, where a person works with the model, and automation, where the model is directed to complete something. In the report it published in January 2026, covering Claude.ai consumer conversations sampled in November 2025, 52% were classified as augmentation and 45% as automation.4 As of August 2026 that was the most recent split it had published as numbers: its March and June 2026 reports say augmentation rose further on Claude.ai without giving updated percentages.4
The split moves between reports, and the direction has reversed more than once: the August 2025 sample had automation ahead at 49% to 47%, and the widely quoted 57% augmentation against 43% automation figure comes from February 2025 and has been superseded several times since.4 The channel matters more than the trend. In August 2025 data from Anthropic's own API, where developers build products on the model, 77% of traffic showed automation patterns and 12% augmentation.4 Consumer chat and production software are different worlds. Anthropic itself notes that "it's not always clear how these conversations might map onto changes in the real world".4
What the labour market shows so far
The Budget Lab at Yale tracks whether AI exposure is showing up in United States employment and wages, comparing exposed occupations against a synthetic control group built from unexposed ones. Their finding, through the first quarter of 2026, is that they "generally find no statistically or economically significant effects as of yet", for either employment share or real hourly wages.5
That is a genuine null result and it deserves to be reported plainly. It also deserves the qualifications the same authors supply. Evidence for workers aged 34 and under is mixed rather than null.5 And they state directly that an early absence of effects does not settle the question: "Even if no effects are evident in 2022 or 2023, they may yet become evident in 2026 or 2027."5 They also point, in a footnote, to a separate analysis finding larger declines in job postings for more exposed occupations since 2022, while noting the post-pandemic period was an unusual macroeconomic episode in which they do not find the same pattern.5
Weak job numbers are not evidence of AI displacement on their own. A companion Budget Lab piece argues that a principal reason for recent weak payroll growth is the slowdown in net immigration, which reduced labour force growth.6 That is a separate analysis by one of the same authors, not part of the exposure study above.
Outside the United States, a Danish study of about 25,000 workers across 7,000 workplaces in eleven highly exposed occupations found no significant effect on earnings or hours two years after ChatGPT's launch, with confidence intervals tight enough to rule out effects larger than a few per cent.7 Almost no one in that sample reported that AI chatbots had changed their own earnings.
What this means for you
A high score is a statement about your tasks, not about your employment. It tells you which parts of your week overlap with something that is getting cheaper and faster, which is genuinely useful information, and it tells you nothing about your employer's plans, your licence, your clients or your industry's demand.
The practical move is to look at which activities carry the overlap on your occupation page, then read what to do if your score is high. If you want the measurement argument rather than the career one, why risk is measured in tasks explains why the unit of measurement changes the answer by a factor of five.
- Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, 2023 (later published in Science, 2024). arXiv:2303.10130. Quoted: the definition of exposure as a proxy that does not distinguish augmenting from displacing effects, and the authors' refusal to predict adoption timelines.
- Mauro Cazzaniga and colleagues, Gen-AI: Artificial Intelligence and the Future of Work, International Monetary Fund Staff Discussion Note SDN/2024/001, January 2024. imf.org. The exposure index is explicitly agnostic between help and replacement, and its high and low categories are set at median values, making them relative rather than absolute.
- Tomlinson, Jaffe, Wang, Counts and Suri, Working with AI, Microsoft Research, 2025. arXiv:2507.07935, version 6 of 22 December 2025. The paper reports agreement between its classification pipeline and three human annotators at Cohen's kappa of 0.44, 0.35 and 0.38 for user goals and 0.53, 0.34 and 0.39 for AI actions, and describes these kappa scores as generally low. The scores are language-model classifications of conversations, not direct observation of work.
- Anthropic Economic Index reports, February 2025 through June 2026. anthropic.com/economic-index. Figures quoted: 52% augmentation and 45% automation on Claude.ai from the report published 15 January 2026, Economic primitives, which covers conversations sampled in November 2025; 49% automation against 47% augmentation in the August 2025 sample it compares against; 57% augmentation against 43% automation in the first report of February 2025, since superseded; and 77% automation against 12% augmentation in first-party API traffic in the September 2025 report, drawn from one million transcripts sampled in August 2025. As of 22 August 2026, the January 2026 pair was the most recent augmentation and automation split Anthropic had stated as numbers: the March 2026 and June 2026 reports say augmentation rose further on Claude.ai but publish no updated percentages. The augmentation and automation shares are classifications of conversation patterns, not measurements of jobs being automated, and the Claude.ai figures do not describe API traffic, which Anthropic reports as automation-dominant.
- Martha Gimbel, Joshua Kendall and Ryan Nunn, What We Do and Don't Know About How AI is Affecting the Labor Market, The Budget Lab at Yale, 2026. budgetlab.yale.edu. The same analysis states that AI exposure metrics are ultimately informed guesses about which tasks and occupations are most likely to be affected, and could be wrong from the start. Synthetic difference-in-differences on Current Population Survey microdata, results reported through the first quarter of 2026. The same analysis reports evidence for workers aged 34 and under as mixed rather than null.
- Ryan Nunn, AI Is Probably Not (Yet) the Reason for Labor Market Weakening, The Budget Lab at Yale, May 2026. budgetlab.yale.edu. A separate companion analysis, which names the slowdown in net immigration as a principal reason for weak recent payroll growth.
- Anders Humlum and Emilie Vestergaard, Large Language Models, Small Labor Market Effects (also circulated as Still Waters, Rapid Currents), National Bureau of Economic Research Working Paper 33777, 2025, revised 2026. nber.org. About 25,000 Danish workers across 7,000 workplaces in eleven highly exposed occupations; null effects on earnings and hours with confidence intervals ruling out changes larger than roughly 2% to 3%; nearly all adopters reported no effect on their own earnings.
Every figure above is quoted with the caveat its own source states. Last verified 2026-08-22. Scores and bands used on this site are documented on the methodology page.