Industry and Economics
Do Robots Destroy Jobs? What the Studies Actually Show
The three most cited studies reach different conclusions because they measure different things. What each found, why they disagree, and what the evidence supports.
The three most cited studies disagree, and the disagreement is informative rather than embarrassing. US regional data finds one additional robot per thousand workers reduces the employment-to-population ratio by roughly 0.2 percentage points. Cross-country data finds robots raised labour productivity growth by about 0.36 percentage points a year without reducing total hours. German data finds no net job loss, with manufacturing losses offset by service gains. All three can be true because they measure different units at different levels.
What each study found
| Study | Unit of analysis | Headline finding |
|---|---|---|
| Acemoglu and Restrepo, US commuting zones | Local labour markets | One more robot per thousand workers lowers the employment to population ratio by about 0.2 pp and wages by about 0.42 % |
| Graetz and Michaels, 17 countries | Industries across countries | Robots contributed about 0.36 pp to annual labour productivity growth, raised wages, and did not significantly reduce total hours, though low-skilled hours fell |
| Dauth and colleagues, Germany | Workers and local markets | No net employment loss; manufacturing job losses offset by service sector gains, with incumbent workers largely protected |
Why the results diverge
- Geographic scope. Local displacement and national reallocation are different measurements. A study of commuting zones finds concentrated losses that a national aggregate absorbs.
- Time period. Displacement appears quickly; the reallocation of labour into new roles takes years, so the answer depends on the window.
- Labour market institutions. Germany's works councils, sectoral bargaining and retraining infrastructure change how adjustment happens. Results from one institutional setting do not transfer automatically.
- What counts as a robot. The underlying data is the IFR installation series, which counts industrial robots and excludes most other automation. Software, conveyors and machine tools do not appear, so the studies measure one slice of automation.
- Direction of causation. Industries that automate may be growing already, which biases naive estimates. The credible studies address this, and the correction is where much of the methodological argument sits.
Putting the coefficients on a real economy
The regional coefficient is easy to misread, so it is worth applying it to concrete numbers.
| Robots per 1,000 workers added | Employment to population effect | Wage effect | Comment |
|---|---|---|---|
| 0.5 | -0.10 pp | -0.21 % | A typical annual increment in a mature market |
| 1.0 | -0.20 pp | -0.42 % | The published coefficient |
| 2.0 | -0.40 pp | -0.84 % | A fast-automating region |
| 5.0 | -1.00 pp | -2.10 % | Beyond the range of the underlying data |
The bottom row carries a warning. Coefficients estimated within one range of variation should not be extrapolated far beyond it, and robot density in the countries studied rose by increments closer to the top two rows than the bottom one. Reading the last line as a forecast is a misuse of the result.
What the evidence does support
| Finding | Confidence | Note |
|---|---|---|
| Robots raise labour productivity | high | Consistent across studies and countries |
| Task composition shifts within jobs | high | Routine physical tasks decline first |
| Effects are locally concentrated | high | Regions with exposed industries bear the adjustment |
| Low-skilled hours fall relatively | moderate | Found in several datasets |
| Net national employment falls | low | Not supported in European data |
| Wage effects | contested | Studies disagree in sign and size |
The gap between rows one and five is the substance of the debate. Higher productivity per worker does not mechanically imply fewer workers; it depends on whether the resulting output growth and reallocation absorb the displaced labour, and that depends on institutions, demand and time.
How to use these numbers
- Quote the unit. The 0.2 percentage point figure is per robot per thousand workers in a US commuting zone. Detached from that unit it means nothing.
- Do not extrapolate across countries. German and US labour markets adjust differently, and the studies show it.
- Do not extrapolate to other automation. The data covers industrial robots only.
- Distinguish level from rate. An economy with high robot density and high employment is not a contradiction; the studies measure changes, not levels.
- Expect the effect to be local. The most robust finding in the whole literature is that adjustment costs concentrate geographically, which is a policy problem rather than a statistical one.
Frequently asked questions
Do robots destroy jobs?
The evidence supports local displacement with contested national effects. US regional data finds one more robot per thousand workers lowers the employment to population ratio by about 0.2 percentage points, while German data finds no net loss because service sector gains offset manufacturing losses.
Why do the studies disagree?
They measure different units at different geographic levels over different periods, in labour markets with different institutions. Local displacement and national reallocation are separate phenomena, and a study designed to capture one will not capture the other.
What is the most robust finding?
That robots raise labour productivity, and that the adjustment costs are locally concentrated. Both recur across datasets and countries, whatever the disagreement about aggregate employment.
Do these results cover all automation?
No. The underlying data is the industrial robot installation series, which excludes software, conveyors, machine tools and most other automation. The studies measure one slice of a much larger phenomenon.
Can I apply US results to my country?
Not directly. Labour market institutions change how adjustment happens, which is exactly why the German results differ from the American ones. Country-specific evidence is required for a country-specific claim.
Sources
- Robots and jobs: evidence from US labor marketsAcemoglu and Restrepo, Journal of Political Economy, the US commuting zone estimates
- Robots at workGraetz and Michaels, Review of Economics and Statistics, the seventeen-country productivity result
- World Robotics report seriesInternational Federation of Robotics, the installation data underlying all of these studies