Overview
The tracker asks one question: which of Anthropic’s three scenarios does the U.S. economy currently resemble? It answers in three steps.
- Trace the scenarios. Each chart’s modest, substantial and extreme paths come from the report.
- Line up the data. Official U.S. statistics are placed on the same footing as the scenario paths.
- Score the gap. Each chart’s latest observation is scored against the paths, and the gauge combines the scores.
1 · Scenario paths
The paths come from Economic Scenarios for Transformative AI (Korinek, Jones, Sacher, Cotter and McCrory, September 2026).
- Labor share and unemployment are traced from the vector lines in Figures 3–4, one point per month from January 2025 to January 2030.
- GDP growth comes from the official explorer’s total GDP paths, using the same year-over-year calculation as the observed data.
- The two knowledge-worker gaps come from the report’s official scenario explorer: the knowledge-worker path minus the other-worker path, in percentage points.
- January 2030 end labels are read from the displayed paths. GDP uses percentage growth in quarterly mean levels; the report instead quotes annual log growth between monthly endpoints.
These are scenarios, not forecasts. The report assigns them no probabilities, and neither does this tracker.
2 · Lining up the data
GDP growth compares each quarter with the same quarter a year earlier, directly, without a 2024 adjustment. The other four charts compare change since 2024, because the model and the official statistics measure levels differently:
aligned = measured − measured 2024 average + shared 2024 level
path = scenario − that scenario’s 2024 average + shared 2024 level
For those four charts, AI already moves each model scenario slightly in 2024, by a different amount in each. Each scenario’s 2024 average comes from the official explorer model, and the shared level is the average of the three, which keeps the paths close to the report’s. Labor share and both unemployment series use all twelve months, matching what they measure; the wage index uses the four quarter-end months it prices.
A full-year average is used, not a single month, so one noisy release cannot shift the whole line. Observed lines begin in 2025. Quarterly data are plotted in the final month of each quarter.
3 · Scoring a chart
Each chart compares its latest observation with the three scenario values over the same period. GDP growth is calculated from quarterly mean levels on both sides; labor share is compared with each scenario’s quarterly average. Every other reading, including the wage index’s quarter-end pay period, is compared with its month:
- 0
- On the modest path
- 50
- On the substantial path
- 100
- On the extreme path
Values between paths are interpolated linearly, and values beyond the outer paths are capped at 0 or 100. Where paths cross or overlap, the closest path sets the score. A score’s color shows the nearest path.
4 · The gauge
The gauge averages the five chart scores, weighted in one of two ways.
- Equal gives each chart 20%.
- Weighted favors charts where the scenarios have already pulled apart. A chart whose paths still sit inside normal year-to-year variation counts for little.
weight ∝ (modest-to-extreme spread ÷ pre-2020 noise)²
| Chart | Spread | Noise | Weight |
|---|
Spread and noise are in each chart’s units at its latest observation. Noise is the 2010–2019 standard deviation of the same comparison.
A score measures resemblance. It is not a probability, and it does not show that AI caused the change.
The charts
GDP growth
- Source
- Real GDP, BEA via FRED · quarterly
- Measure
- Percentage change in quarterly real GDP from the same quarter a year earlier, calculated the same way for observations and scenarios.
- Why
- We use growth because it can be observed directly, while GDP above a no-AI path requires an assumed counterfactual.
- Caveat
- GDP can undercount AI’s output. See the Early Signals methodology below.
Knowledge-worker wage vs. other workers
Anthropic compares knowledge workers’ wages with a hypothetical no-AI path, which we cannot directly observe. We adapt that chart into a gap between knowledge workers and other workers so it can be tracked using observed wages for both groups. This helps account for shared economic movements, but retains the baseline assumptions below and does not isolate AI’s effect.
- Source
- Employment Cost Index by occupation, BLS via FRED · quarterly
- Measure
- Real wage growth for knowledge workers minus real wage growth for other workers, in percentage points.
- Method
- Each group’s wage index is combined at fixed weights, adjusted for inflation, set against a 1.67% a year real-wage trend, and rebased to its own 2024 average.
- Caveat
- The index covers private-industry workers only, so it approximates the model’s broader groups.
Labor share of income
- Source
- Nonfarm business labor share, BLS via FRED · quarterly
- Measure
- The share of income paid to workers, scaled so its 2024 average equals the scenarios’ shared 2024 level, about 59.9%.
Unemployment gap, knowledge vs. others
- Source
- Employment and unemployment by occupation, Current Population Survey via FRED · monthly
- Measure
- The knowledge-worker unemployment rate minus the rate for everyone else, in percentage points. It starts from the model’s pre-AI difference of −2.6 points.
- Groups
- Knowledge work is management, professional, sales and office jobs. Other work is service, construction, maintenance, production and transportation jobs. Unemployed people are grouped by their last job; people with no previous job are left out.
- Method
- The published counts are not seasonally adjusted, so each group is adjusted with fixed monthly factors estimated from 2010–2019.
- Check
- Testing X-13 seasonal adjustment produced an average absolute change of 0.068 percentage points across 20 monthly readings (January 2025–September 2026); we retain the current method.
Unemployment, all workers
- Source
- Unemployment rate, Current Population Survey · monthly, seasonally adjusted
- Gap
- October 2025 was not collected during the federal government shutdown. A solid light-gray segment connects the surrounding observations; the missing month is never scored.
Knowledge work displacement
- Source
- Employment by occupation, Current Population Survey · monthly, not seasonally adjusted
- Measure
- Knowledge work’s share of employed people, comparing three-month averages for the latest complete quarter and the same quarter in 2022.
- Why
- 2022 anchors the comparison around ChatGPT’s launch; matching quarters reduces seasonal differences, and averaging three months smooths monthly noise.
- If it falls
- Displaced is the drop in knowledge-work employment. Crossed over is the part matched by more jobs in other occupations. Still need to move is the part matched by more unemployed former knowledge workers. All three are percentage points of the labor force with a previous job.
- Caveat
- It compares group totals, not individual workers, and is not scored or included in the gauge.