Framing Artificial Intelligence and Labor Displacement: A Survey Experiment on Policy Preferences in Latin America

How to cite

Scartascini, Carlos. 2025. "Framing Artificial Intelligence and Labor Displacement: A Survey Experiment on Policy Preferences in Latin America." AEA RCT Registry. AEARCTR-0016600. https://doi.org/10.1257/rct.16600-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Carlos Scartascini*

* Primary investigator

Registry ID
AEARCTR-0016600
DOI
https://doi.org/10.1257/rct.16600-1.0
First registered
Last updated
Status
In development
Trial dates
2025-08-20 – 2025-09-26
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/16600

Abstract

We conducted a series of survey experiments to identify individual's public policy preferences regarding AI regulation and AI deployment as a security tool.

Intervention

Survey experiments embedded in an online survey conducted by the Latin American Public Opinion Project (LAPOP). The module on artificial intelligence (AI) was embedded in the national questionnaires as a randomized survey experiment, with respondents assigned to different framing conditions.

Experimental design

The experiment investigates how different framings of the impact of artificial intelligence (AI) on employment affect individual preferences over public policy responses. Building on prior work on automation and technological anxiety, we assess whether respondents’ support for government intervention varies when the disruptive potential of AI is presented with different narratives.

All respondents are first introduced to a neutral explanation of AI, described as a technology capable of learning and making decisions independently. Two follow-up questions measure baseline perceptions: whether AI may replace their job and which social group they believe would be most affected by AI-induced unemployment. Following this, respondents are randomly assigned to one of four treatment arms with equal probability (1/4). Each treatment group is exposed to a brief vignette that frames the impact of AI on employment in one of three ways—an alarmist tone emphasizing mass job loss, a delayed scenario minimizing urgency, or a reallocation narrative stressing transition support (see questions in Appendix A.1). These framings were drawn from real-world headlines and media discourse.

The fourth group serves as a control and receives no framing. The Control group received no additional prompt and proceeded directly to the outcome questions. The Alarmist group was exposed to a headline citing reports from international consulting firms that warned up to 60% of current jobs could be lost to automation. This framing emphasizes the immediacy and severity of AI-related job displacement. In contrast, the Delayed Threat group read a headline which acknowledged the potential for technological disruption but emphasized that such effects are unlikely to materialize in the short term, projecting instead a gradual transformation over the next decade, with limited automation
of existing tasks. Finally, the Job Reallocation group was presented with a headline arguing that AI might eliminate “pointless” or low-value jobs, suggesting that the technology could reallocate labor away from routine tasks and toward more meaningful work, without
necessarily reducing the overall demand for human labor.

All respondents were then asked the same key outcome question: if AI reduces the amount of human labor needed, what public policy should the government pursue? The response options included (1) imposing regulations requiring firms to retain workers, (2) raising taxes on those who remain employed in order to subsidize the unemployed, (3) not intervening in the labor market, or (4) expressing no opinion. This outcome captures respondents’ policy preferences regarding redistribution and regulation in the face of technological change.

By randomly assigning informational framings and holding the outcome measure constant, the design permits identification of causal effects of different narratives on policy attitudes. Moreover, the design allows us to test for heterogeneous effects by education, occupation,
income, and baseline perceptions of job risk.

Design details

Randomization method
All respondents are first introduced to a neutral explanation of AI, described as a technology capable of learning and making decisions independently. Two follow-up questions measure baseline perceptions: whether AI may replace their job and which social group they believe would be most affected by AI-induced unemployment. Following this, respondents are randomly assigned to one of four treatment arms with equal probability (1/4). Each treatment group is exposed to a brief vignette that frames the impact of AI on employment in one of three ways—an alarmist tone emphasizing mass job loss, a delayed scenario minimizing urgency, or a reallocation narrative stressing transition support (see questions in Appendix A.1). These framings were drawn from real-world headlines and media discourse. Randomization done by LAPOP (computer) Control (p ∼ 1/4): Respondents receive only the baseline definition of AI. Alarmist Framing (p ∼ 1/4): Respondents read a headline citing estimates from PwC, McKinsey, and the World Economic Forum warning that up to 60% of current jobs could be eliminated by AI-driven automation. Delayed Threat (p ∼ 1/4). Respondents read a headline from Goldman Sachs emphasizing that AI-driven changes will be gradual, with only 5% of tasks being automatable and effects not expected for at least a decade. Job Reallocation (p ∼ 1/4). Respondents read a headline suggesting that AI will eliminate ”meaningless” jobs,
Randomization unit
Individual
Clusters
The final analytic sample includes 4,147 respondents: 1,006 in Argentina, 1,021 in Brazil, 1,016 in Colombia, and 1,104 in Mexico
Observations
The final analytic sample includes 4,147 respondents: 1,006 in Argentina, 1,021 in Brazil, 1,016 in Colombia, and 1,104 in Mexico
Arms
1. Experiment 1: 1/4 in each arm from each country 2. Experiment 2: 1/2 in each arm from each country
Primary outcomes
1. Policy preferences for labor regulation 2. Policy preferences for the deployment of AI tools for the detection of crime

Cited by

No citation data available for this registration yet.

Keywords