How this dataset is built

What is on the site

547 jobs across 22 industries have a page, all listed on the jobs list. Each job has one of two kinds of page:

• 512 timelines. One per occupation: employment and wages across as many years as cited sources reach, the technology eras that reshaped the work, and cited forecasts for the next decade or so. Many timelines also show the present-day view described next.
• 35 job profiles. The present-day view on its own, for a job with no timeline yet: its main tasks, the named AI tools that touch each task, adjacent careers, and learning resources.

The occupation list and task descriptions come from O*NET, the US Department of Labor's occupation database. Not every O*NET occupation has a page yet.

Official statistics, researched claims and our own words

Not everything on a page is checked to the same standard. There are three kinds of content. Pages do not yet label which kind each item is; the Sources section on each page lists what it cites.

Official statistics

Counts, pay and projections published by a statistics agency, and the scores a research paper published: US employment and wages from the BLS occupational surveys (the OES files from 1997, OEWS since 2003), the BLS ten-year projections, the Australian, UK and Canadian series, the AI exposure scores, and O*NET task descriptions. Code copies these figures from the agency's or the paper's own files, with no AI drafting. Matching each US job to the closest Australian, UK or Canadian occupation code was done with AI help and then checked by a separate AI pass. For the BLS figures and the exposure scores, an automated check compares the numbers on the site with those files on every change to the dataset.

Researched claims

Everything else that cites a source: older US employment and wages (before 1997, and any figure before 2003 that comes from a source other than the BLS surveys), estimates, tool eras and their dates, historical notes, forecasts from other organisations, facts about named AI tools, and the evidence behind each task. AI research agents (Anthropic's Claude) find the sources and draft these claims, each naming the source it came from. A separate AI pass is meant to check each claim against that source before it is published.

Our own words

Plain-language writing by us: the task labels, the suggested next careers and why they fit, the ways to adapt a task, and summaries of each era. These are our reading of the cited evidence, not statements a source makes.

Where the checking stands

Official statistics are checked automatically, as described above. Researched claims are less certain. An audit in October 2026 found that the AI checking pass had not caught everything: some researched claims are not fully supported by the source they cite. We are re-checking researched claims against their sources and correcting or removing what does not hold up. Until then, treat a researched claim as a lead to follow to its source, not a settled fact.

What each number means

• Employment. How many people did this job in that year, in the country shown. For the United States from 2003 on, it is the national employment estimate from the BLS Occupational Employment and Wage Statistics survey (OEWS). A few recent US points come from the BLS Employment Projections matrix instead, which counts all jobs including the self-employed; those are labelled “BLS Employment Projections”. Earlier years come from the census and other cited surveys.
• Median annual wage. The middle annual pay for the occupation, in that year's money and the country's own currency. When a source only published an average, the label says “Average annual wage” instead, because an average is not a median. Most pre-1990 wage figures are averages.
• Inflation-adjusted wage. For US wages, the line “in 2024 prices” is the same wage adjusted for inflation. We calculate it; see the inflation section below.
• Tool eras. The technology waves that changed how the work is done (the typewriter, the spreadsheet, the large language model), each with start and end years and the sources that date it.
• Cited forecasts. A forecaster's published percentage change in employment by a target year, exactly as that forecaster states it.
• AI exposure measures. Research scores for how much of this work AI could technically do. They are not forecasts of jobs lost; see the exposure section below.
• Task labels. On a curated role, each task carries a plain-language label (“AI is taking this on”, “AI is sitting alongside you here”, “You still lead here”). The label is our editorial reading of the cited evidence for that task, not a measured statistic.

When you move the slider to a year with no cited figure, the readout says which year the number shown actually comes from (“nearest cited” or “latest actual”). We do not fill gaps. Charts plot each cited figure as a separate dot with no line between them, because a line would imply values nobody published.

Stated, estimated, or derived

Every figure falls into one of three groups, and the page treats each one differently.

Stated in a source

Most figures are copied from a named source: a census or survey count, a published median wage, a published forecast, or a research paper's own score. Each links to that source in the page's Sources section.

Estimate (marked “Est.”)

Some early figures are estimates made by the cited source itself, for example a historian's reconstruction for years before a census counted the job separately. They are still cited, but they are not a count, so they show as hollow dots on the charts and carry an “Est.” mark in the readout. We never make estimates of our own.

Derived by us

A few figures are calculated by us from cited inputs, using fixed, documented rules:

• Inflation-adjusted wages (US only), from the cited wage and the official consumer price index.
• Yearly figures from quarterly surveys. Australian employment is the average of a year's quarterly estimates, added up across the survey's state and sex groups to a national total.
• Annual pay from weekly or hourly pay. Where a source only gives weekly or hourly earnings (the current Australian wage, and some historical US figures), we convert it to a year of full-time pay, for example weekly pay times 52. The figure's note says how.
• Sums of occupation groups. When a US occupation has no single match in Australia, the UK or Canada, that country's employment is the sum of the cited counts for the groups that together match it. The wage shown is the cited median of the largest group, never a blend, because medians cannot be added or averaged.
• Exposure scores as percentages. Papers publish exposure as a fraction between 0 and 1; we show it multiplied by 100 and rounded to one decimal place.

How AI exposure is measured

Timelines can show up to three research measures. Each row names its study year and what it measures, and each figure is the one the study published, not a number from a summary or a news story. For Eloundou et al. and Frey and Osborne we read the paper's own published data, and an automated check, run on every change to the dataset, fails if a number on the site drifts from it.

Eloundou et al. (2023), “GPTs are GPTs”

The paper rated the tasks listed in O*NET for whether a large language model could cut the time to do it by at least half with no loss of quality. A task counts fully if the model alone could do that, and half if it would need extra software built on top of the model. We show the paper's beta score: the share of the occupation's tasks exposed on that basis. We use the ratings made by human annotators; the paper also published ratings made by GPT-4, which the row's method note lists for comparison.

Frey and Osborne (2013), “The Future of Employment”

A statistical model, trained on 70 occupations the authors labelled by hand, that gives 702 US occupations a probability of computerisation: the chance that the occupation as a whole is technically automatable over an unstated period the authors suggest is a decade or two. It is a probability for the whole occupation, not a share of its tasks.

Goldman Sachs (2023), “The Potentially Large Effects of Artificial Intelligence on Economic Growth”

The report scored the tasks of US occupations for exposure to automation by generative AI and published the share of work exposed for broad occupational groups only (for example all business and financial operations occupations), not for individual occupations. So a Goldman Sachs row shows the figure for the group the role belongs to, and its label names that group. It is a 2023 estimate, not a forecast for a later year.

What these measures are not

• Not forecasts of jobs lost. Work that becomes faster can grow, not shrink.
• Not a measure of whether employers actually use AI. All three describe technical possibility at the time of the study.
• Not comparable with each other. Eloundou et al. measure a share of this occupation's tasks, Goldman Sachs a share of a whole occupational group's work, and Frey and Osborne a probability for the whole occupation, ten years earlier.

Eloundou et al. and Frey and Osborne used their own occupation codes. We show their score only when the occupation matches ours one to one through the official code crosswalks. When an occupation was split or merged between code versions, we show no score rather than average or guess one.

Forecasts are other people's

We publish no forecasts of our own. Every forecast on a timeline is a figure published by a named organisation, with its target year and a link to where it says so, for the horizon that source published. The figure shown must appear in the cited source; we drop anything a source did not state itself, such as a midpoint of a published range or a number scaled to fit this role.

The main forecast is the BLS Employment Projections ten-year outlook, currently 2025 to 2035. Every timeline whose exact occupation appears in the BLS National Employment Matrix carries one row from it, read from the matrix file itself. When BLS publishes the occupation only inside a combined line, the timeline gets no BLS row: we do not substitute a parent group. Some roles add a forecast from another named body; one published for a broader group is shown only with that group named.

Where forecasters disagree, we show the disagreement. We do not average forecasts, pick a winner, or extend a forecast past its published target year.

Adjusting for inflation

US wages are adjusted with the BLS Consumer Price Index for All Urban Consumers (CPI-U, US city average, all items, series CUUR0000SA0), using annual averages. The base year is 2024: the adjusted wage is the original wage multiplied by the 2024 index and divided by the index for the wage's year, rounded to the nearest dollar.

The CPI-U begins in 1913, so earlier US wages are shown in their original dollars only. Australian, UK and Canadian wages are shown in their own currencies at the prices of the time; we do not adjust them for inflation or convert between currencies.

Countries and what they count

Every timeline has a United States series. Many also carry an Australian, UK or Canadian series. Each country counts differently, so compare the shape of each series, not the levels between countries.

• United States. Occupations use the Standard Occupational Classification (SOC). From 2003 on, employment and median wages come from BLS OEWS, which counts wage and salary jobs and leaves out the self-employed; a few recent points come from the BLS Employment Projections matrix, which includes the self-employed. Earlier figures come from the decennial census (often via IPUMS), the Current Population Survey, historical BLS bulletins, past editions of the Occupational Outlook Handbook, and specialist surveys.
• Australia. Occupations use ANZSCO unit groups. Employment comes from the ABS Labour Force Survey (detailed table EQ08), from 1986. Wages are one current figure: median weekly earnings from the ABS Employee Earnings and Hours survey, as published by Jobs and Skills Australia, times 52. Australia publishes no detailed wage history by occupation, so there is no earlier Australian wage series.
• United Kingdom. Occupations use the UK SOC. Wages (median annual gross pay) and employment (number of employee jobs, which leaves out the self-employed) both come from the ONS Annual Survey of Hours and Earnings, Table 14, most back to 2002.
• Canada. Occupations use the National Occupational Classification (NOC). Figures come from the Statistics Canada Census of Population for the income years 2005, 2015 and 2020: the number of people in the occupation with employment income, and the median employment income of those who worked full year, full time. The 2005 figures count wages and salaries only, so they leave out self-employment. We skip the 2011 National Household Survey, which was voluntary and carries quality warnings.

Where the data comes from

Every figure on a timeline or curated profile is tagged with its source. The sources we draw from most:

Employment and wage figures are drawn from the U.S. Bureau of Labor Statistics. The BLS does not vouch for analyses derived from its data, and we make no claim of BLS endorsement.

Australian figures are based on Australian Bureau of Statistics and Jobs and Skills Australia data, used under CC BY 4.0. Neither agency endorses Future History or its analysis.

UK figures contain public sector information licensed under the Open Government Licence v3.0.

Canadian figures are adapted from Statistics Canada, Census of Population, used under the Statistics Canada Open Licence. This does not constitute an endorsement by Statistics Canada.

How a timeline is built

1. Find the historical series. Employment from census tables and wages from historical bulletins, each figure drafted by an AI research agent from the document it cites.
2. Load the modern series. US employment and wages from 2003 on come straight from the BLS OEWS files, not from research notes.
3. Date the tool eras. Each technology wave that reshaped the role gets start and end years, a description of its effect on the work, and its sources.
4. Add published forecasts and exposure measures. Only figures a named source published, each linked.
5. Check. A second AI pass, separate from the one that drafted the research, is meant to check each researched figure against its source before it is published. Automated checks then require a citation on every figure and block a release if a rule is broken. A citation being present does not prove the source says what we claim; see where the checking stands.

Known limits

• Surveys count different things. OEWS and the UK survey leave out the self-employed, so trades with many self-employed workers look smaller than they are. Older census counts and the BLS Employment Projections matrix include them. A jump between two sources can be a change of survey, not a change in the job.
• Occupation definitions change. Census categories, the US SOC, the UK SOC and Canada's NOC have all been revised over time, and early censuses often grouped several jobs together. Where a definition changed, the figure carries a note.
• Gaps are real. Many series have long stretches with no cited figure, especially wages from the 1940s to the 1960s. We leave gaps as gaps.
• Catch-all occupations have no timeline. The official “All Other” codes bundle many unrelated jobs, so one employment or wage series for them would mislead. They keep a role page instead.
• Coverage outside the US is partial. Australian, UK and Canadian series exist only where an occupation has a faithful match in that country's classification, and their histories are shorter. The Australian survey table we use (EQ08) ended in February 2026; future Australian updates depend on a new classification.
• Exposure measures date quickly. They describe what the technology of 2013 or 2023 could do, not what today's tools can.

How to cite a page

Cite the page and the date you used it, because pages change as the data does. For example:

Future History. “Registered Nurses.” futurehistory.earth/timeline/29-1141. Accessed [day month year].

For a single figure, cite the original source too. Each one is linked in the page's Sources section. Licensing for reuse is on the terms page.

Found a mistake?

The value of Future History collapses if we get a citation wrong. If you spot an error (a wrong employment number, a misdated tool era, a forecast attributed to the wrong source), please flag it.

Use the contact form with the role, the year, and the source you think is wrong, and select “Correction or data issue.” We read every submission and fix what we confirm is wrong. The dataset was last updated 11 October 2026.