Measuring the Credibility of Nonpolitical Institutions

I will build on previous studies on labor automation (Arntz et al., 2022) to align with existing literature and provide out-of-sample benchmarks. The information will be designed to operate through the updating of beliefs and perceptions, rather than priming. The experimental design will follow a “pre-post” structure with the elicitation of first-stage beliefs and updated posteriors (Stantcheva, 2023). This approach will allow me to examine the credibility and influence of different nonpolitical institutions—such as trade unions, business organizations, and newspapers—on individual decision-making. By systematically varying the sender of the information, I will assess the credibility of these institutions and measure their impact on both beliefs and behavior.

Please see the attached document "Experiment Design & Analysis Plan."

2025-04-14

2025-05-05

Primary end points:
- Change in beliefs (before vs after)
- Policy preferences (especially voting intention and support for reforms)
- Differential impact by sender

1. Belief updating:
- Change in perceived number of jobs at risk (quantitative posterior vs prior)
- Change in perceived impact of AI (qualitative posterior vs prior)

2. Policy preferences:
- Support for job security measures, wage protection, or training programs
- Intended vote in the upcoming labor market reform referendum

3. Sender credibility and perceived trustworthiness:
- Inferred credibility

4. Behavioral intentions:
- Self-reported likelihood of taking action (e.g., changing job, seeking training)

5. Emotional and cognitive reactions (if using open-ended coding or scales):
- Fear, optimism, confusion, etc. in response to the information

Please see the pilot questionnaire in the attached document "Experiment Design & Analysis Plan" for the exact operationalization of outcome measures.

Heterogeneity across trust in political and nonpolitical institutions, ideology, populist attitudes, age, gender, social capital, occupation, contract type, and sector.

The survey will be conducted by the professional polling company SWG through fieldwork for CAWI. The goal is to randomize information m, which will be the same for all treatment groups except the control group, but the source of this information will differ across the groups receiving the treatment. The sample will be divided into five groups of 650 individuals each. The information will be provided in a graphic format resembling a newspaper headline to increase its impact.
• Group A (650 individuals): Receives the information from "major trade unions."
• Group B (650 individuals): Receives the information from "major business associations."
• Group C (650 individuals): Receives the information from "major newspapers."
• Group D (650 individuals): Receives the information without a specified source.
• Group E (650 individuals): Does not receive no information (control group).

Versions of information treatment.
• (Trade Unions): "Major trade unions warn: 66% of jobs at risk due to artificial intelligence. Cgil and Cisl discuss a recent study."
• (Business): "Major business associations warn: 66% of jobs at risk due to artificial intelligence. Confindustria and CNA discuss a recent study."
• (Newspapers): "Major newspapers warn: 66% of jobs at risk due to artificial intelligence. Corriere e Repubblica discuss a recent study."
• (No Source): "66% of jobs are at risk due to artificial intelligence. A recent study discusses this issue."

Note that there is no deception involved, as the information treatments simply state that the specific nonpolitical institutions “discussed” the study in question. In fact, as shown by a press review conducted in preparation for this survey, when a study estimated that 66% of jobs in developed countries could be at risk due to generative AI, major newspapers, trade unions, and business organizations all discussed its findings. Therefore, the statements in the information treatments are accurate.

Please see the attached document "Experiment Design & Analysis Plan."

Randomization automatically performed by the polling company SWG in the CAWI fieldwork.

Survey respondents.

No

Not applicable – treatment is not clustered. Randomization occurs at the individual level (CAWI respondents).

Approximately 3,250 individuals (CAWI respondents) in Italy.

650 individuals – Trade unions as sender
650 individuals – Business associations as sender
650 individuals – Newspapers as sender
650 individuals – Unspecified sender (neutral)
650 individuals – No sender displayed (control group)

Assuming individual-level randomization with 650 respondents per arm, a total sample of 3,250, and no clustering, the minimum detectable effect size (MDES) for main outcomes is approximately 0.155 standard deviations at 80% power and a 5% significance level (two-tailed). This corresponds to a small-to-moderate effect size.