The best-documented effects of automation on US workers come from two economics papers by Daron Acemoglu and Pascual Restrepo. One found that each additional industrial robot per thousand workers in a local labor market lowered the employment-to-population ratio by about 0.18 to 0.34 percentage points between 1990 and 2007, and the other attributed 50 to 70 percent of changes in the US wage structure since 1980 to automation-driven task displacement. Neither study measured generative AI, and both describe averages, not outcomes for any one worker.
Key takeaways
- A 2020 study in the Journal of Political Economy tied more industrial robots to lower local employment and wages in the US from 1990 to 2007.
- A 2022 Econometrica paper linked automation of routine tasks to a large share of US wage-structure changes.
- Both studies cover earlier technologies, so they do not show what current AI tools will do.
- Results are averages across regions or worker groups, not forecasts for individual jobs.
- Automation means machines performing tasks; AI is one way to do that, not the only one.
Automation and AI are not the same thing
Automation is the use of machines or software to perform a task that people did before. A conveyor-belt sorter, a spreadsheet macro, and a robotic welding arm are all automation, and none of them needs machine learning. AI describes a family of methods that can automate tasks, including ones that are hard to write as fixed rules, such as recognizing speech or drafting text. The distinction matters for evidence: much of what is known about the social effects of automation comes from decades of rule-based machines, and the findings carry over to current AI only with care.
What the robots study measured
In Robots and Jobs: Evidence from US Labor Markets, first circulated by the National Bureau of Economic Research in March 2017 and published in the Journal of Political Economy (volume 128, issue 6) in 2020, the authors compared local labor markets with different exposure to industrial robots between 1990 and 2007. The abstract reports large negative effects that held across specifications: one more robot per thousand workers lowered the employment-to-population ratio by about 0.18 to 0.34 percentage points and wages by 0.25 to 0.5 percent. The design compares places, so the numbers describe regional labor markets, not what happened to each displaced worker.
What the wage-structure paper measured
In Tasks, Automation, and the Rise in US Wage Inequality (NBER working paper, June 2021; Econometrica, 2022), the same authors build a framework in which automation extends the set of tasks that capital performs and displaces workers who had a comparative advantage in them. They attribute between 50 and 70 percent of changes in the US wage structure over four decades to relative wage declines among groups specialized in routine tasks in industries that automated quickly. Offshoring plays a smaller role in their measure, and they find that market power and deunionization do not appear to be major drivers. The estimate depends on the model and on how tasks are classified, which is why the paper's own wording is a range.
The two studies side by side
The table is an orientation aid built from the abstracts, with the limits that apply to each.
| Question | Robots and Jobs | Tasks, Automation, and Wage Inequality |
|---|---|---|
| Technology studied | Industrial robots | Automation of routine tasks in general |
| Period | 1990 to 2007 | About 1980 to 2016 |
| Unit compared | Local labor markets | Worker groups by demographic and education |
| Main reported result | Lower employment and wages where robot use rose | 50 to 70 percent of wage-structure changes linked to task displacement |
| Main limit | Regional averages; ends before current AI | Model-based estimate; routine-task focus |
Why exposure is not the same as job loss
Both papers estimate displacement and wage effects, but neither says that a given occupation will disappear. A job is a bundle of tasks, and automating some of them can change the job without removing it. Reading any statement about "jobs at risk" calls for checking whether it counts tasks that could be automated, jobs that were actually lost, or wage changes, because those are three different measurements. Both studies also concern the United States. Canada has a different industry mix and labor institutions, so the numbers should not be transferred to Canadian provinces without separate evidence.
Questions to ask about a claim on automation
- Does the study measure robots, software, or AI, and in which years?
- Is the result about employment, wages, or the tasks that could be automated?
- Is the unit a region, an industry, or a worker group?
- Does it cover the United States, Canada, or another country?
- Is the figure an estimate with a range, and what are the stated limits?
FAQ
Do these studies show that AI destroys jobs?
No. They examine industrial robots and the broader automation of routine tasks in earlier decades. Generative AI was not part of their data.
Who wrote the studies?
Daron Acemoglu and Pascual Restrepo, who are economists. Both papers appeared first as NBER working papers and were later published in peer-reviewed journals.
Does automation lower wages for everyone?
The wage-structure paper links automation to relative wage declines for groups specialized in routine tasks, not to a uniform fall. Effects differ by worker group.
Is a working paper the same as a published study?
No. A working paper circulates before peer review. Both studies cited here later appeared in journals, and the journal versions are the ones to cite for final numbers.
Related reading
- AI as augmenting automation
- AI and robotics in society
- AI compared with automation
- Where AI is reported to help
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Robots and Jobs: Evidence from US Labor Markets, Acemoglu and Restrepo, NBER Working Paper 23285, 2017; Journal of Political Economy 128(6), 2020
- Tasks, Automation, and the Rise in US Wage Inequality, Acemoglu and Restrepo, NBER Working Paper 28920, 2021; Econometrica, 2022