Most arguments about AI and work start with headcount: how many roles get automated, in which year, according to which bank. A July 2026 paper from Adam Mickiewicz University starts somewhere else. It asks whether replacement anxiety can be turned on by a story about who is in charge — even when nobody says the word “jobs.”
In a lab, it can.
The finding worth keeping
Grobelny, Klakus, Szymański, and Chirkowska-Smolak (arXiv:2608.14609) ran a preregistered 2×2×2 experiment (N = 316) in a partitioned lab. People watched about eighteen minutes of short user-generated videos, then filled a six-item job replacement anxiety scale (“I am afraid that an AI technique or product may replace humans”).
The independent variable was not “AI takes jobs.” It was perceived AI-over-human control versus human-over-AI control. The AI-control reels were about workplace surveillance, unsupervised scaling, decision autonomy, autonomous replication, singularity timelines. The human-control reels were the same topics with oversight, regulation, and governance. The authors are explicit:
Importantly, none of the videos explicitly addressed job replacement.
After the reels, people in the AI-control arm scored higher on replacement anxiety (b = 0.96, robust SE 0.15, p < .001). The manipulation check moved as intended (d = 0.82). The bump survived core self-evaluations, trait anxiety, general negative attitudes toward AI, age, and recent job loss.
That is the piece I would put on a wall: replacement fear can be acquired vicariously from control narratives. You do not need a layoff, a Gartner slide, or even the word “unemployment.” A story that AI is the agent is enough to light the job-fear item.
Why that is not a labor-market paper — and why it still matters
The experiment does not tell you whether displacement is happening. It tells you that a large part of the feeling can be produced upstream of the feeling’s object. Li and Huang’s Integrated Fear Acquisition Theory already said this: most people have not been fired by a model, so the fear has to be learned from other people’s stories. This paper is the first clean randomization of that pathway I have seen.
For anyone who ships, briefs, or posts about AI, the implication is mechanical. “The system decides.” “Humans remain in the loop.” “Unsupervised scaling.” These are not neutral product sentences. They are agency cues. In this sample they moved a replacement-anxiety score even when the videos stayed off the HR topic.
The adjacent result: “useful” is not a sedative
The authors also grafted on the Technology Acceptance Model. Ease of use did nothing in either study. Usefulness did something they did not preregister: it raised the baseline of anxiety. The extra bump from the control narrative got smaller as usefulness went up, because people who already believed AI was highly useful were already anxious. Absolute anxiety peaked when AI was framed as both useful and in charge (M ≈ 4.17 versus 2.88 in the low-usefulness, human-control cell).
Their post-hoc reading is better than the original TAM overlay. Usefulness is closer to severity of the threat than to “people will adopt this happily.” If the machine is good at the work, replacement feels consequential even without a dystopia reel. Pairing “this model is extremely capable” with “don’t worry, humans stay in control” is not a cancellation. In their cells, capability is part of the threat.
What I would not hang on this
- The cover story was “how AI is communicated to the public,” then eighteen minutes of themed reels, then a replacement-anxiety scale. Demand characteristics are not zero.
- The effect is an immediate state. No delay, no job-search, no tool adoption.
- The control scale in Study 1 is thin (α ≈ .62–.63). The anxiety scale is fine (α = .83).
- Study 2 (N = 995) is a same-sitting survey. It supports the association. It does not re-run the causal arrow.
- Polish lab sample, mostly office and intellectual work. Existence proof, not a census of “the workplace.”
So: do not cite this as proof that media caused the labor-market panic. Cite it as proof that agency talk is already a job-fear stimulus, and that usefulness talk does not soothe it.
What to keep
- Job-replacement anxiety can be elicited without job content. Control narratives are sufficient.
- The effect is situational. Trait self-evaluations did not eat it in Study 1.
- Ease of use is the wrong lever if the person is not the user of the system but the possible object of it.
- “AI is useful” raises the floor of fear. It is not a buffer.
- If you write about agents, you are already writing about jobs, whether the copy says so or not.