How the numbers are computed
Where every number on this site comes from, how it is computed, and what it can't tell you.
How it all fits together
Everything — every chart, graph, or projection on this site — starts from real, measured data. When we provide a projection, we start with data Anthropic Economic Index as of Jun 2026 describing how much of every job's work AI is already being used on, based on real usage, broken out task by task. From a second data set METR benchmark trend 131-day doubling, we know how fast AI capabilities are improving (right now, they're doubling every 131 days). From a third Census BTOS as of Sep 2026, we know how fast businesses are adopting AI. And we know where the effects are showing up first — and it isn't overall employment.
Where effects are showing up first is at entry-level jobs. The Stanford Canaries study Stanford Canaries −3.6%/yr · Aug 2026 tracks employment growth for 22- to 25-year-olds in AI-heavy occupations — and it has turned negative, falling about 3.6% a year, while employment in those same occupations overall is flat. Every study that has looked for AI's effect on total employment has come back with the same result: no measurable effect. The honest picture today is a narrow one — the newest workers in the most exposed jobs are feeling something the broader workforce isn't.
What nobody knows is what happens to jobs after AI takes over part of the work. When AI can do a third of a job's tasks, does employment in that job fall by roughly a third over the long run? By a little? Not at all? That's what the research hasn't settled: some studies find no effect; others allow nearly a one-for-one response.
But that's not very useful to people who are trying to get a sense of what is possible or likely. So the model does the next best thing: it takes today's measured picture of the data and runs it forward — exposure growing as AI capability improves, adoption following its S-curve — and applies the strength of the job effect that the research best supports for that world.
The default view is called today's pace. Every dial sits at its currently measured value: AI keeps improving at its measured rate, businesses keep adopting along their measured curve, and because no study has found job effects yet, the projection holds flat — until adoption crosses the level where studies start picking up job losses, which the young-worker trend puts around 2029. At that point the effect turns on at the research's lower bound, and the bars begin to lengthen.
The other three views move specific dials away from today's measurements — each a different reading of the same research, not a different model. AI speeds up: capability doubles faster, more of AI's use takes work over, and the effect runs at full strength from today. Slower shift: capability doubles more slowly and less of AI's use takes work over — a gentle decline. AI mostly helps: most AI use assists people rather than replaces them, so the effect stays weak and the losses stay small. The same measured exposure runs through all four; what changes is the strength of the link to jobs. This helps you get a sense of the possible.
Every projection is shown with a band of upper and lower bounds that illustrate the possible job impacts (these are called "whiskers"). We show these because we want to be clear: we just don't know how things are going to play out. So our model runs the same measured exposure under every answer — from "jobs don't change" to "they fall roughly in proportion" — and the 10th-to-90th spread of those runs is drawn as the whiskers. The bars are the answer we think the evidence best supports; the whiskers are the honest boundary of what anyone can claim.
Find your number
Use the table below to quickly find an explainer or more detail about the data on this site.
| If you saw it on | The numbers | Where it's explained |
|---|---|---|
| Occupation pages | AI exposure, the ten-year projection, the young-worker view | AI exposure; Young workers; How the projection works |
| Community pages | Local exposure, strain score, adaptive capacity, the ten-year projection | AI exposure; Your local economy; How the projection works |
| National Pulse | Exposure trend, business adoption, the impact card, the stress gauge | AI exposure; AI adoption; Has AI changed employment yet?; Your local economy |
| Scenarios | The projection and its dials, the scenario tracker | How the projection works; The scenario tracker |
| APIs and downloads | Composite scores, displacement risk | Technical reference |
The data underneath
The projection starts from three measurements:
- how much AI is already used in each job
- how fast AI is getting better, and
- how fast businesses are adopting it.
Here's how each of these are sourced and what the site does with it.
AI exposure
Where you'll see it: Occupation pages, Community pages and National Pulse.
Exposure is the share of a job's working time that people already use AI for. It's foundational to understand that Exposure measures what people actually do with AI, not what a model says AI could do.
Where the data comes from. Anthropic's Economic Index publishes how people use Claude on real work tasks. We've stored all six quarterly releases so far, February 2025 through June 2026, about 3.7 million rows. Live values: /api/v2/pulse/national. Because we keep every release as published, we're able to work with exposure as a trend, rather than think of it as a snapshot in time.
How we turn it into a score. We match each task in the data to O*NET, the Labor Department's list of what tasks make up most every conceivable occupation. Then we weight each task by how much of the job's time it takes, using O*NET's own importance and frequency ratings. The result is the share of a job's task time that AI is used for.
We compute that twice: once for automation (people directing AI to do the task) and once for augmentation (people working with AI: learning, iterating, getting feedback).
How we map it to places. The Bureau of Labor Statistics maintains a database of metro areas in the US cross referenced with federal occupation data (BLS OEWS). We keep a fresh copy of this data in our warehouse and we use it to average exposure across every metro area's jobs, weighted by how many people hold each one. For counties, we estimate the job mix: county employment by industry (BLS QCEW), spread across occupations using each industry's national staffing pattern.
How to read the numbers. Communities are shown on an index where 1.0 is the national average. A community at 1.20 is 20% more exposed than average; 0.80 is 20% less. We decided on some thresholds: below 0.8 is Low, 0.8 to 1.25 is Moderate, and 1.25 and up is High. Those bands are for comparing one place to other places, so we can find patterns in the data (like the maps).
Every metro's score is an estimate, so each one is drawn with a band around it: the zone where the true values plausibly sit. The band is built to capture the true value 90% of the time.
The national trend. National Pulse charts the share of U.S. jobs in the most exposed quarter of occupations. Which occupations the data covers shifts from release to release, so we also draw a fixed line: the latest release's most exposed occupations, held constant and traced back through all six releases. That line is essentially flat, so the dips in the main line come from changes in coverage, not changes in use.
How we check it. Our occupation rankings closely match Anthropic's own published rankings (rank correlation 0.85, in other words– pretty good). Our estimated county job mixes reproduce known metro job mixes well (average correlation 0.944).
There's a catch though. Exposure built from those estimates tracks direct metro exposure less closely (0.693 overall, 0.80 for the 50 largest metros). While it's not great, it's the best we could do at estimating local job mixes from national patterns. We hope to have a fix here soon to improve our quality.
What it can't tell you. Exposure measures use, not job loss. A highly exposed job is one where a lot of work already happens with AI. On its own, that says nothing about whether employment in that job will fall.
AI capability
Where you'll see it: the capability speed dial on Scenarios, and inside every projection.
Capability is how fast AI is getting better. We take it from METR's benchmark trend, which right now shows AI's abilities doubling every 131 days. Live value: scenarios model payload. METR measures capability as the length of task an AI system can complete end to end — its "time horizon" — and publishes how fast that horizon doubles. That doubling rate is the trend this site runs on.
How it feeds the projection. Exposure measures what AI is doing in a job today. Capability growth moves that toward what AI could do there, using Eloundou et al.'s published ratings of each job's could-do share. Each doubling closes half the remaining gap. A job where AI could do little stays low, however fast capability grows.
The dial. Today's pace uses the measured 131 days. The dial runs from 89 days to 250. AI speeds up uses a faster doubling (89 days), and Slower shift a slower one (213.5 days).
What it can't tell you. The trend measures what AI can do, not what anyone uses it for. The rule that turns doublings into exposure is our translation, not a measurement.
AI adoption
Where you'll see it: National Pulse and Scenarios.
Adoption is the share of U.S. businesses that say they use AI. It's the S-curve the projection runs on.
Where the data comes from. Every two weeks since September 2023, the Census Bureau's Business Trends and Outlook Survey has asked a sample of U.S. businesses whether they use AI to produce goods or services. In December 2025, Census changed the question to ask whether businesses used AI in any of their business functions in the last two weeks. That rewording is why the line changes shape there. We've collected every published wave, 77 so far. Live value: /api/v2/adoption/national.
Illustrative: intermediate BTOS readings omitted — two measured waves shown (Nov 2025: 17.3%, Sep 2026: 23.8%, the model's anchor).
How to read the numbers. Each figure is the share of businesses answering yes. "Don't know" is never counted as no. Estimates that don't meet Census's own publication standards show as missing, never as zero. Every wave comes with a 90% confidence interval, which we publish alongside it.
Census publishes the figures nationally, by industry, by firm size, by industry and size together, and by state. We publish the same breakdowns. State figures are modeled by Census, and we label them that way.
How we check it. Stanford's Real-time Population Survey asks workers, not businesses, about using generative AI. A share of workers and a share of businesses aren't comparable levels, so we only compare direction: rising, falling or flat. Since the December 2025 rewording, the two agree.
Each line is indexed to 100 at its own first point, so the chart compares direction, never level. The worker series is quarterly; the firm series is biweekly.
What's still in development. The projection uses one national S-curve fitted to these waves. More detailed curves, by industry, firm size and state, are still research inputs. With only three years of data, most of them can't yet find where adoption tops out, so none appear on the site.
What it can't tell you. A business saying it uses AI says nothing about how much of its work changed, or whether any jobs did. The survey only goes back to September 2023, too short for long-run claims.
What's happening now
Two things are true right now. Young workers in the most exposed jobs are losing ground. And no one, including us, can yet find an effect on overall employment. Here's the evidence for both, and the signals we watch for what comes next.
Young workers
Where you'll see it: Occupation pages, and behind the Scenarios projection.
Young workers in the most AI-exposed jobs are losing ground now. Employment growth for 22- to 25-year-olds in high-exposure occupations is falling 3.6% a year as of August 2026 Live value: /api/scenario., while employment in those same occupations overall is flat.
Where the data comes from. The Canaries study from the Stanford Digital Economy Lab, with ADP, tracks employment growth by age in the occupations most exposed to AI.
How we use it. It's one of the eight signals in the scenario tracker, and one of the most heavily weighted. We treat −6% a year as a break. From 2023 to 2025 it ran between −1% and −4%; today it's about 60% of the way to the break. The same trend sets when job effects turn on in the projection, and it drives the young-worker view on occupation pages.
What it can't tell you. It covers one age group in one set of jobs. It shows young workers losing ground, not why, and it says nothing about overall employment.
Has AI changed employment yet?
Where you'll see it: the impact card on National Pulse.
We can't tell yet, which is why the card says Inconclusive. Jobs in the most AI-exposed industries have fallen behind other industries. But the two groups were already drifting apart years before AI arrived, and our test finds the same gap even when we sort industries at random. The test is picking up something other than AI.
What we compare. Employment in the most exposed third of 19 industry sectors against every other industry, quarterly from early 2015 through late 2025, each indexed to 100 at the start. The exposed group is currently wholesale trade; retail trade; information; professional, scientific and technical services; management of companies; and educational services. Live value: /api/v2/impact/gap-series. Groups are set by exposure rank, so the grouping can't shift with the result. The data is the Census Bureau's Quarterly Workforce Indicators, ages 20 and up.
What the chart shows. Behind the line, a band marks the gap's normal range from 2015 through late 2022, −7.9 to −0.1. The gap left that band in late 2025. That alone isn't a verdict, because the gap was already trending down before AI.
The five tests.
| Test | What it asks | What it found |
|---|---|---|
| Overall trend | Is the gap since 2022 real, or noise? | Real by the numbers, but this kind of test is known to produce results this strong even when nothing changed |
| Before COVID | Was the drift already there in 2015–2019? | Yes. The groups were on different paths before AI could have mattered |
| Shuffled exposure | Would any grouping of industries find the same thing? | Yes. All 30 random groupings found the same signal |
| Drop one industry | Is a single industry driving it? | No. Dropping each of the 19 sectors in turn changes nothing |
| Local shocks | Could one county's plant closing or policy change create the result? | No. The math accounts for shocks across 1,742 counties |
What's next. We'll try newer methods built for this problem and re-test with every data release. The full statistics, and how to reproduce every test with public data, are in the Technical reference.
The scenario tracker
Where you'll see it: Scenarios.
The tracker watches eight signals from published research for evidence of which future is taking shape: AI speeds up, Slower shift or AI mostly helps. The projection runs assumptions forward; the tracker reads today's evidence.
Each signal has a published threshold that would mean movement toward one of the three, and a weight for how reliable it is. Heaviest signals first:
| Signal | Source | What would count as movement | Weight | Points toward | ID |
|---|---|---|---|---|---|
| Recession onset | Sahm Rule on BLS unemployment, via FRED | The rule triggers. A recession would make displacement worse. | 0.9 | AI speeds up | L7 |
| Young workers in exposed jobs | Stanford Digital Economy Lab / ADP (Canaries) | Employment growth for ages 22–25 falls past −6% a year (it ran −1% to −4% in 2023–2025) | 0.8 | AI speeds up | L4 |
| Recent-graduate unemployment gap | New York Fed, ages 22–27 vs. everyone | Gap above 2.5 points (historically 0.5 to 1.5) | 0.7 | AI speeds up | C4 |
| AI helping vs. doing | Anthropic Economic Index | More than 60% of AI use is helping rather than doing | 0.7 | AI mostly helps | L6 |
| Layoffs blamed on AI | Challenger, Gray & Christmas monthly reports | Above 40% in any month (15% to 25% in 2024–2025) | 0.6 | AI speeds up | L1 |
| Shift in the occupation mix | Yale Budget Lab tracker | Becomes statistically significant | 0.6 | AI speeds up | C3 |
| Job postings | Indeed Hiring Lab, Feb 2020 = 100 | Index at or above 100 | 0.5 | Slower shift | L3 |
| Layoff rate | BLS JOLTS | Stays above 1.36%, two standard deviations over its 2015–2019 average | 0.5 | AI speeds up | L5 |
The thresholds and how each was set are published at /api/methodology. Current values are at /api/scenario.
We don't publish odds. Internally, each future starts at one-in-three and moves only as signals cross their thresholds. Until we have a model that's been tested against history, we show the evidence, not a verdict.
What it can't tell you. A signal can cross its threshold for reasons that have nothing to do with AI. Waiting for several signals to line up cuts false alarms, but can't eliminate them. And no set of signals proves AI caused anything; interest rates, trade, demographics and the business cycle all move jobs at the same time.
How the projection works
Where you'll see it: Scenarios, and as fixed ten-year views on Occupation and Community pages.
The projection runs one job's employment, or one place's, forward ten years. The same settings always produce the same bars. It runs in your browser, and every setting behind the answer is shown on the page.
What's fixed, and what you can change
Three inputs come straight from data. We do not change these numbers.
- Where adoption is today. 23.8% of U.S. businesses said they used AI in the Census Bureau's September 2026 survey. Live value: /api/v2/adoption/national. The model starts there. The dials change the future.
- How exposed each job is. The share of its work AI is already doing (see AI exposure), and the share AI could do, from Eloundou et al.'s published ratings.
- The shape of the adoption curve. An S-curve fitted to the Census survey's biweekly results.
Everything else is a dial:
| Dial | What it sets | Default and range |
|---|---|---|
| Capability speed | How fast AI's abilities double, from the METR benchmark trend. Each doubling closes half the gap between what AI does on a job today and what it could do there. | 131 days Live value: scenarios model.; 89 to 250 days |
| Adoption ceiling | Where business AI use tops out | Default 100% (the fitted ceiling sits at its upper bound); 0 to 100% |
| Timing | Slides the adoption curve earlier or later | Default: on current trends (0); up to 3 years either way |
| The mix | How much AI use does the task (automation) versus helps with it (augmentation) | Measured national split, about 47/53 Live value: scenarios model. |
| Job effect | How strongly AI use changes employment | Default 0; rises toward 0.25 as adoption passes the turn-on (48.5%); 0 to 1 |
When job effects turn on
Today's pace holds employment flat at first because no study has found a job effect yet. But those studies ran when few businesses used AI. Assuming the result holds forever would be its own claim, so the model lets it expire as adoption grows.
The turn-on is calibrated from the young-worker trend (see Young workers). Over the period we've measured, each added point of business adoption has pushed young-worker employment growth down about 0.10 points. Run forward from today's −3.6% a year, it reaches the −6% break at about 47.7% adoption.
On today's pace, the modeled adoption path crosses that level in January 2029 — the payload pins the turn-on at adoption 48.5% on the monthly grid (a_star_month, the first month at or past the crossing). Below the turn-on, the job effect stays at zero. Above it, the effect rises steadily toward 0.25, the weakest effect the research supports (the bottom of the range in Bick, Blandin & Deming 2024). That keeps today's pace gentler than any of the other three views.
Note: the exact-day logistic reaches 48.5% on ≈19 Apr 2029; the payload pins the turn-on at the 2029-01 monthly grid point (a_star_month), which is what is drawn here. The elasticity (−0.10 canary points per adoption point) rests on one measured adoption interval (Nov 2025 → Sep 2026).
This is an assumption fitted to two measured points, not a finding. If the young-worker decline speeds up, the turn-on moves earlier and the bars grow sooner. If the break never shows up in the data, we take the assumption back out. The full derivation, and the test that would prove it wrong, is in the calibration memo Calibration memo 2026-10-21.
How the bars are built
Each month, the model computes the pressure AI puts on a job: job effect × share of businesses using AI × the job's exposure. Automation pressure lowers employment. Augmentation pressure raises it a little. Changes compound month to month, and nothing is random.
How far jobs respond comes from published studies. At full strength, a full year of automation pressure moves a job's employment by about −30% — the studies' responses span roughly −25% to −35%, and the model runs the range — based on the wage response in the AEI difference-in-differences study. The augmentation counterpart is about +3%, the top of a published GDP estimate (NBER working paper 32487). Both translate wage and GDP findings into employment terms, which is our modeling choice, flagged for review.
The math in your browser matches a Python reference version, and tests pin the expected results.
How the whiskers are built
The whiskers come from 64 runs of the same model. Each run uses one combination of the research's own answers: how strongly jobs respond, how fast AI capability grows, and how strong the job effect is. We draw the 10th-to-90th percentile spread of those runs.
The whiskers are computed for the scenario you're viewing, so they differ between views. They do not move with your custom dial adjustments. They start at zero width today and widen over time. They're solid through 2030, where the studies stop, and lighter after, where we extend the same trends. "No change" always falls inside them.
The young-worker view
The default view covers every worker in a job, because that's what the job studies measure. Occupation pages add a view for young workers: the job's projected path plus the measured young-worker gap (with its own month-to-month spread), scaled by the share of that job's workers who are 20 to 24 (BLS Current Population Survey). It combines measured pieces; it isn't a separate model. Jobs without age data get no young-worker view rather than a guess.
Your area
Pick a metro and the projection uses that place's own mix of jobs, each job's change weighted by how many people do it there (BLS OEWS 2024). Metros without enough local data fall back to the national result and say so.
You can also work backward. "Start from an outcome" lets you set a target, and the page finds the dial settings that would produce it. It shows what would have to be true.
What the projection can't tell you
Three pieces of the model are translations, not measurements: the job response (taken from wage and GDP studies), the rule for how capability growth raises exposure, and the job-effect turn-on. That's why today's pace holds flat until adoption reaches the turn-on, and why "no change" always sits inside the whiskers. We don't publish odds for any view. No scored forecasts have been issued yet; when they are, each one will be published the day it is made and graded against what actually happened on the Forecast Track Record page.
Your local economy
Where you'll see it: Community pages and National Pulse.
These numbers describe how tough a job market is right now, for any reason. They're context, not a measure of AI. Each can move for reasons that have nothing to do with AI, which is why we measure exposure separately.
The local strain score
The strain score turns three local numbers into 0-to-100 scores and averages them. Each uses a fixed scale, so a place's score depends only on its own numbers, never on how other places are doing.
| Input | Source | Scores 0 at | Scores 100 at | Current for |
|---|---|---|---|---|
| Unemployment rate | BLS Local Area Unemployment Statistics, monthly | 3% | 7% | 6 months |
| Job postings | Indeed Hiring Lab index (Feb 2020 = 100) | 100 | 80 | 6 months |
| Layoff notices | WARN filings: workers laid off per 10,000 employed, past 12 months | 0 | 168 | 12 months |
Values in between scale evenly. Values past either end are capped at 0 or 100. The layoff scale tops out at 168, the 99th percentile of metro layoff rates in 2015–2019, so one huge layoff can't read as a trend.
We collect WARN filings through Big Local News (Stanford) and the New York WARN archive, and remove duplicates. The score, the layoff card and the notices list all read the same records, so they can't disagree. A metro needs at least 1,000 jobs on record for its layoff rate to count.
At least two of the three inputs must be current, or the page says "insufficient local data" instead of a score. Below 25 is low strain, 25 to 50 moderate, above 50 high. The score isn't a ranking or a percentile; 100 is just the top of the scale.
Example: Chicago metro, October 2026. Dated snapshot Oct 2026
| Input | Value | Score |
|---|---|---|
| Unemployment rate | 3.67% | 16.8 |
| Job postings index | 86.55 | 67.3 |
| Layoff notices | 30.6 per 10,000 employed | 18.2 |
| Strain score | 34, moderate |
Each input uses its own latest month, so they don't always line up. The page rounds to a whole number.
The national stress gauge
The Labor Market Stress gauge on National Pulse works the same way for the whole country, with one swap. WARN filings aren't collected nationally, so layoffs come from the BLS JOLTS layoff rate instead: 0.8% scores 0, 2.0% scores 100. Unemployment comes from BLS via FRED.
The gauge covers every month with all three inputs: February 2020 through August 2026, 79 months. Live value: /api/v2/stress/national. It peaked at 100 in April 2020, when COVID hit. "Calmer than N% of months since 2020" is a rank within that window, not a probability.
Adaptive capacity
Adaptive capacity is how well a place could absorb economic change. It combines four Census numbers for every county and metro: the share of adults with a bachelor's degree, the share of households with broadband, median household income, and how varied the local industries are. Higher means more room to adapt. It's not an AI prediction. It's available through the APIs and data download.
What it can't tell you. All three measures describe the whole local economy. A high-strain metro may be in an ordinary downturn. A low-strain metro may be quietly losing AI-exposed jobs as people leave and aren't replaced, which never shows up as a layoff.
Limits
Nothing on this site proves AI caused a change in jobs. We show patterns that fit AI displacement, not proof of it. Labor markets respond to interest rates, trade policy, demographics and the business cycle all at once, and we can't separate AI from the rest.
- False alarms happen. A metro can score high on strain because of an ordinary recession. A tracker signal can cross its threshold for reasons unrelated to AI.
- Local survey data is sampled. Census ACS numbers come from samples, not full counts. For smaller places, the margin of error can be large.
- Some values are estimates. Employment projections, some occupation exposure scores and our in-house fallback estimates are modeled, not measured, and labeled "estimated" in our data.
- U.S. only. Every source covers the United States. We have nothing on other countries.
What we don't compute yet
- Whether AI has changed employment. We publish and watch the gap between exposed industries and everything else, but our current test can't isolate AI's effect. See Has AI changed employment yet?
- Scored forecasts. The projection runs measured trends forward under the dials you set. We don't yet publish forecasts for individual jobs or metros that get scored against what actually happens. The tools to do that exist, and their results will appear on Forecast Track Record.
Other data we collect
We collect these for future work. None of them feed a published number yet.
- New businesses: business applications by county (Census Business Formation Statistics)
- Migration: people and income moving between counties (IRS Statistics of Income)
- Local businesses: establishment, employment and payroll counts (Census County Business Patterns and Nonemployer Statistics)
- Data centers: data-center employment by county (BLS QCEW) and national data-center construction spending (Census)
Technical reference
For researchers and API users — select to expand
This section is for researchers and API users. Most of these scores aren't shown on the site; they're served through the APIs and data downloads listed on APIs & downloads.
Confidence labels
Every value in our database carries one of these labels.
| Label | Meaning |
|---|---|
| measured | Comes directly from a live API or data source. Highest confidence. |
| extracted | Parsed from a source that needed transformation, such as a PDF or web page. Good confidence. |
| estimated | A model estimate, heuristic or placeholder. Directional only. |
| computed | Calculated from measured or extracted values. Inherits their confidence. |
| proxy | A stand-in from a related place or aggregate, not a direct observation. |
Scenario tracker weights
Each future's internal weight is a softmax over a log prior plus live evidence:
The prior b is equal across the three futures (33.3 / 33.3 / 33.4) and keeps sparse data from causing big swings. The smoothing factor λ is 1.0, set in config/scenario_thresholds.json. With few signals crossed, the weights sit near the prior. Current evidence totals are at /api/scenario.
Impact test statistics
The series behind every test is at /api/v2/impact/gap-series: 2015Q1 to 2025Q4, 44 quarters, sector-level Census QWI over ages 20–24, 25–54 and 55+. Suppressed cells add no employment, and unscored sectors drop out of both groups.
| Test | Window | Result |
|---|---|---|
| Overall trend (event-study leads test) | 2019Q1–2022Q3, 14 leads | Wald 995.1, p ≈ 1.8×10−203; 1,496,100 rows, 1,742 county clusters |
| Before COVID | 2015Q1–2019Q4, 20 leads | Wald 862.2, p ≈ 8.7×10−170; leads uniformly positive, +0.0015 to +0.0156 (about +0.0095 a quarter) |
| Shuffled exposure | 30 seeded shuffles | 30 of 30 reject at α = 0.01 and 0.10; median p ≈ 1.4×10−68; the real map's p falls inside the shuffled range |
| Drop one industry | 19 sector groups, both windows | All 19 reject; weakest is dropping Information (NAICS 51), p ≈ 7.8×10−33 pre-COVID |
| Clustered errors | All tests | Standard errors clustered across 1,742 counties |
The exposed-vs-unexposed comparison isn't identifiable with this design. A rejection that survives every sector drop and every shuffle is the known signature of spurious pre-trends in two-way fixed-effects estimators (Goodman-Bacon 2021; de Chaisemartin & D'Haultfœuille 2020; Sun & Abraham 2021), not evidence that AI changed employment. Next candidates are modern difference-in-differences estimators and occupation-mix outcomes, pending review.
To reproduce: download the QWI release, build the county-by-quarter panel, and use scripts/wp4/event_study.py (the estimator and clustering) and scripts/wp4/wp42_diag_probeC.py (shuffles with seeds 1–30, and the sector drops) as reference implementations. The run is documented in docs/wp42-diagnosis.md. None of it needs our infrastructure.
Career exposure composite
Each occupation gets an exposure score blended from three published measures, weighted Felten 40%, Eloundou 35%, Eisfeldt 25% (the cognitive weights; a physical weight set waits on O*NET work-context data). The site never shows this blend; the Occupations ranking shows the measures side by side. It's served by the career APIs.
A plain weighted average. When a measure is missing, its weight is spread proportionally over the others. Felten's published data covers 774 SOC codes (805 O*NET occupations); the other 215 occupations use an in-house fallback estimate (heuristic-v1).
| Occupation | Felten | Eloundou | Eisfeldt | Composite |
|---|---|---|---|---|
| Economists (19-3011.00) | 0.9705 | 0.5000 | 0.5692 | 0.7055 |
| Software Developers (15-1252.00) | 0.6800 (fallback) | 0.8684 | missing | 0.7679 |
For Software Developers, effective weights are 53.3% Felten and 46.7% Eloundou. Both examples checked against /api/career/detail/19-3011.00 and /api/career/detail/15-1252.00 on 2026-10-06.
Composite precision
With two or more methodology scores present, the career detail API reports a weighted standard error and 95% interval at risk_assessment.composite_se. The standard error is computed from every methodology in the composite — including the in-house fallback row when it fills a published measure's slot — which is worth knowing when reading an interval that mixes measured and fallback scores.
Weights match the composite's, so a lightly weighted measure counts less. With fewer than two methodology scores present, composite_std_error is null and precision_disclosure explains why. Economists: SE 0.1541, 95% interval 0.4034 to 1.00 (clipped), wide because Felten and Eloundou disagree by 47 points. Software Developers: SE 0.094, 95% interval 0.5837 to 0.9521, from the Felten fallback and Eloundou (n = 2) — checked against the live API on 2026-10-20.
Displacement risk
Exposure is how much a job's tasks overlap with what AI can do. Displacement is actual job loss caused by AI. The site shows exposure only; displacement risk is API-only. None of the three academic measures estimates displacement.
The career detail API's "Peak Displacement Risk" (displacement_risk, from assess_displacement_risk in api/career_scoring.py) is the exposure composite times a multiplier, capped at 1.0. The first matching rule applies:
| BLS 10-year projection | Multiplier |
|---|---|
| Growing more than 10% | 0.7 |
| Shrinking more than 5% | 1.2 |
| Anywhere in between | 1.0 |
Most occupations fall in between, so their displacement risk equals their exposure, and the API flags it. Both examples above do: 0.7055 and 0.7679. A friction multiplier (0.8 for high-friction jobs) was removed until O*NET work-context data arrives; see friction_note in /api/methodology.
A second field, displacement_score, comes from the heuristic-v1 row. It's null for most occupations: suppressed when it's 0.01 or below, or within 0.001 of exposure. When present it carries a "heuristic" badge (Software Developers: 0.0345).
National AI-Impact Index
Served by GET /api/geo/metro-impact, which powers the heatmap:
Stress is the 0-to-100 local strain score. x is the metro's employment-weighted exposure (0 to 1). 0.6685 is the national exposure and 0.83425 the neutral multiplier, so a metro at national exposure keeps its stress score. The 1.8 gain spreads observed metros across most of the 0.5-to-1.0 range. The anchors are fixed constants rather than ranks, so one metro's score never moves when another's data changes. A metro needs its own OEWS occupation data and two current strain inputs to get a score.
National Risk Index
The plain average of the strain score across all tracked metros, 0 to 100. Below 20 is low, 20 to 40 moderate, 40 and up high. API and download only.
Community scores (CM1 and CM2)
CM1, exposure. Each place's employment-weighted measured AI use divided by the national average, so 1.0 is the national average. The older 0-to-1 Felten-based score from ACS tables is retired; don't compare old copies of it to the current index. The full national industry table, every NAICS-3 industry's exposure with the automation and augmentation split, is served by the industry-exposure API.
CM2, adaptive capacity. Four parts, each scaled 0 to 1 and averaged, for 3,135 counties and 923 metro areas from ACS 2024 5-year data:
- Education: share of people 25 and up with a bachelor's degree (B15003)
- Broadband: share of households with a broadband subscription (B28011)
- Income: median household income, capped at $150,000 (B19013)
- Industry diversity: 1 minus the Herfindahl-Hirschman Index of industry employment shares (C24030)
Metro stand-ins for three signals (L1, L4, L6)
Three tracker signals come from national sources we don't have by metro. On metro pages we show stand-ins built from local data, in their own section, and they aren't comparable to the national versions:
- L1, layoffs blamed on AI: built from WARN filings, in WARN-based units
- L4, young workers: the 12-month change in local unemployment × local exposure
- L6, helping vs. doing: the inverted 12-week change in the local postings index
Full descriptions are in the metro_l1_l4_l6_proxies field at /api/methodology.
Collected but not shown
Bachelor's degree share (CM8) is stored from Census ACS but kept off metro panels. It's in the data download and API. A test (tests/test_geo.py::test_cm8_not_in_metro_panels) keeps it off unless deliberately added.
Metro employment composite (MS1) began as a four-metro pilot and now covers all 802 tracked metro and micro areas: nonfarm payrolls in information, professional and business services, and administrative support (thousands). Its label was updated 2026-07-22; the ID and values didn't change.