Automation Anxiety Is Mostly a Story. Evidence Can Correct It.

More than half of workers in the US and Germany believe automation will push unemployment up. Fewer than three in ten believe it will happen to _them_. That gap — between the catastrophe we imagine for everyone else and the risk we actually assign to ourselves — is the most revealing number in a new study, and it points to something hopeful: a large share of automation anxiety is not a forecast. It is a story we absorbed. And stories can be corrected with evidence.

The study is _"The end of work feels near. How do people perceive the impact of digital technologies and automation?"_ by Melanie Arntz, Sebastian Blesse and Philipp Doerrenberg, published in Labour Economics (2026) and hosted by the Institute for Employment Research (IAB). It is unusual among labor-market papers because it does not measure what automation does to jobs. It measures what people _think_ automation does to jobs — and then tests, experimentally, whether telling them the actual research findings changes anything.

It does. But not in the tidy way an optimist would hope.

What people actually believe

The researchers ran a customized survey with 5,147 working-age respondents — 3,066 in the United States and 2,081 in Germany — fielded through YouGov in February and March 2019. Everyone in the sample was in the active labor force, aged 18 to 55.

The picture that comes back is lopsided in a very specific way. More than half of respondents in both countries expect automation to raise the aggregate unemployment rate, while only around 10% expect it to fall. Roughly 90% in both countries expect automation to hit different social groups unequally, with low-skilled workers seen as the biggest losers.

And yet, when the same people are asked whether _they_ are likely to lose their own job to automation within five years, only a little more than a quarter say they are at least somewhat concerned. [Fact]

This is not people being inconsistent for no reason. The authors show these are genuinely distinct dimensions of what they call "automation angst." The correlation between fear of rising aggregate unemployment and fear for one's own job is below 0.1. Fear for the economy and fear about inequality correlate at about 0.3. In other words, "the robots are coming for jobs" and "the robots are coming for _my_ job" are two nearly independent beliefs living in the same head. [Fact]

Anxiety about the economy is partly political, not occupational

Here is the finding that should make anyone rethink how they read automation headlines. Personal risk perception tracks reality reasonably well: it correlates strongly with actual job and workplace characteristics — routine task share, exposure to digital tools, employment history. People are, roughly, decent judges of their own exposure.

Economy-wide fear behaves differently. In the US, it is strongly predicted by political ideology and trust in government, even after controlling for demographics and job characteristics. A left-leaning respondent reports concerns about rising unemployment 0.2 standard deviations higher, and distributional concerns 0.5 standard deviations higher, than a right-leaning one. In Germany, where 24.9% of respondents place themselves at the political extremes versus 50.9% in the US, this ideological patterning is much weaker. [Fact]

So a meaningful slice of "automation will destroy work" is not a labor-market judgment at all. It is a political identity expressing itself through a labor-market question.

The experiment: what happens when you show people the evidence

This is where the study earns its place. Respondents were randomly assigned to receive one of two pieces of real scientific information — drawn from Graetz and Michaels (2018), a peer-reviewed study of industrial robots across 17 countries from 1993 to 2007 — or to a control group that received nothing.

The first treatment told people the study's finding that robots did not significantly reduce overall employment, because productivity gains and lower prices boosted competitiveness. The second told them the study's other finding: robots reduced the employment share of low-skilled workers.

The first treatment worked. Fear of rising aggregate unemployment fell by about 0.15 standard deviations on average, and concern that skilled workers would suffer also declined. Effects were somewhat stronger in the US. [Fact]

Three details make this more than a footnote. First, 87% of treated respondents said they found the information trustworthy and 82% found it helpful — they were not tuning it out. Second, in a follow-up survey four weeks later with 2,225 US participants (about 75% of those recontacted), the researchers found no bouncing back: the corrected perceptions held. Third, the second treatment — the one confirming that automation hurts the low-skilled — barely moved anything, and where it did, it _reduced_ concern for college-educated workers by about 0.1 standard deviations. People already believed the bad distributional news. [Fact]

The asymmetry is the point. Information corrected the belief that was wrong (mass job destruction) and did almost nothing to the belief that was already right (unequal impact). That is exactly what you would expect if aggregate automation angst is, in significant part, a correctable misperception rather than a rational forecast. [Claim]

Now the part optimists skip

The same information did not produce a uniform response, and pretending otherwise would be dishonest.

Treatment effects depended heavily on prior beliefs. The drop in unemployment fear was driven mainly by people who started out pessimistic. But shifts in _personal_ and _distributional_ concern showed up more among people who were _less_ pessimistic to begin with. And when the researchers looked at what happened to policy preferences, the responses were heterogeneous and in places opposing: respondents with optimistic priors reduced their demand for policy intervention after the treatment, previously neutral respondents moved somewhat in the opposite direction, and pessimists — the very group whose fear dropped most — did not change their policy demand at all. [Fact]

There is a further catch. Higher automation angst in the raw survey data was associated with a greater stated willingness to switch occupations and invest in retraining — but also with donating less to charity in a real, incentivized donation decision. Fear was doing two jobs at once: motivating self-protection and eroding solidarity. Remove the fear and you may lose the motivation along with it. [Fact]

And the honest limitation: the survey was fielded in early 2019. The information given to respondents concerned _industrial robots_, not generative AI. Nothing in this study tells you that the same reassurance would hold if you ran it today with an LLM-exposed office worker. Evidence that corrects one narrative does not automatically license optimism about the next one.

What this means for you

The practical lesson is not "relax, nothing will happen." It is narrower and more useful than that.

Separate the two questions. "Will AI destroy jobs in general?" and "Is my job exposed?" are different questions with different answers, and the research shows most people conflate them. Your personal exposure depends on your actual task mix — how routine your work is, how much of it is codified — not on the headline you read this morning. Our data puts automation exposure for customer service representatives far above that of welders, even though both appear in the same "robots are taking over" coverage.

Check the source of the scary number. The widely-quoted claim that roughly 47% of US jobs are at high risk of automation comes from an occupation-level estimate. When the same question is asked task by task, the OECD estimate — from a team including this paper's lead author — lands at about 9% of jobs at high risk across 21 countries. The number changed by a factor of five because the method changed, not because the world did. [Fact]

Use the anxiety, don't be used by it. The finding that worried workers are more willing to retrain is real and worth keeping. The goal is to hold the motivation without the dread: read your own task exposure honestly, retrain against it deliberately, and stop outsourcing your emotional state to aggregate forecasts that the evidence says are wrong more often than they are right.

If you want to see where your own work sits rather than where the narrative says it sits, start with your occupation page — for example accountants, software developers, or paralegals — and compare the task-level breakdown to the story in your head. In this study, that comparison was worth about 0.15 standard deviations of fear. It costs you five minutes.

Sources

• Arntz, M., Blesse, S. & Doerrenberg, P. (2026). "The end of work feels near. How do people perceive the impact of digital technologies and automation?" _Labour Economics_. IAB publication record (DOI: 10.1016/j.labeco.2026.102897, Open Access)
• Arntz, M., Gregory, T. & Zierahn, U. (2016). "The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis", OECD Social, Employment and Migration Working Papers No. 189.

Update History

• 2026-07-14: Initial publication, based on Arntz, Blesse & Doerrenberg (2026), _Labour Economics_, via the IAB publication record.

---

_AI-assisted analysis. This article was drafted with AI assistance and reviewed by a human editor. Survey figures, experimental effect sizes and sample counts are taken directly from the cited study; occupation-level exposure estimates are from our own O\*NET-based dataset._