Impact of AI-driven Safety Systems on Worker Well-being

Respondents take part in a survey experiment with vignettes. Everyone first reads a short description of a factory worker and standard workplace safety rules. At this point, after the baseline vignette, the respondents assess the baseline working conditions in terms of safety and job quality. They are then randomly assigned to one of three groups (about 700 respondents per group):

AI-only safety system: The factory introduces a system based entirely on artificial intelligence.

Human-only safety system: The factory introduces a system based entirely on trained human safety supervisors.

Hybrid AI + Human system: The factory introduces a system that combines both approaches, with AI monitoring and human supervisors.

After reading the vignette, respondents answer questions about safety, risks, trust, health, privacy, and job quality. Trust and privacy are only asked after the treatment.

All respondents first read a baseline vignette describing a 40-year-old male assembly line worker in a Dutch manufacturing plant with standard occupational health and safety measures. After answering baseline questions on safety, risk, health, job quality, and wages, respondents are randomly assigned to one of three intervention vignettes:

AI-only safety system (N≈700): The factory introduces a system based entirely on AI. It continuously monitors the factory floor with cameras and smart wearables (e.g., helmets with sensors), detects risks such as fatigue, overheating, sudden movements, or faulty equipment, and issues real-time safety alerts and recommendations.

Human-only safety system (N≈700): The factory introduces a system based entirely on trained safety supervisors. Supervisors walk the factory floor, detect risks such as fatigue or hazards, and provide immediate safety alerts and recommendations.

Hybrid AI + Human system (N≈700): The factory introduces a system that combines both approaches. AI continuously monitors the floor, while human supervisors conduct their own checks and review AI alerts before issuing recommendations.

Study 1: Safety, Risk, and Trust

Outcomes: perceived safety, risk, stress, and physical demand; expected accidents, physical health problems, and mental health problems (all measured pre- and post-intervention, analyzed as within-subject changes).

Placebo: perceived societal importance of the job (pre- and post-intervention).

Post-treatment only: trust in the system, perceived effectiveness, and respect for privacy and dignity (between-subject comparisons).

Study 2: Job Quality and Fair Wage Perceptions

Outcomes: job satisfaction, meaningfulness, and perceived fair wages (measured before and after the safety system vignette, enabling within-subject analysis).

Design details
Randomization is preloaded by Centerdata to ensure balanced groups by age, gender, education, income, employment status, and urban/rural location. The survey was programmed and tested by Centerdata and fielded via the LISS panel in October and ends by November, 2025. Average completion time is about nine minutes. The respondents answer short comprehension checks after each vignette but can continue answering the questionnaire if they fail the checks (they have two attempts to answer correctly).

2025-10-06

2025-11-02

We preregister two sets of primary outcomes:

Study 1: Safety, Risk, and Trust

Perceived job safety, riskiness, stress, and physical demand (measured pre- and post-vignette).

Expected likelihood of workplace accidents, physical health problems, and mental health problems (measured pre- and post-vignette).

Placebo outcome: perceived societal importance of the job (measured pre- and post-vignette).

Trust in the safety system, perceived effectiveness, and respect for privacy and dignity (measured post-randomization only).

Study 2: Job Quality and Fair Wage Perceptions

Job satisfaction, job meaningfulness, and perceived fair wages (measured pre- and post-vignette).

All outcomes are self-reported on Likert-type or categorical response scales and are derived directly from the survey instrument programmed in the LISS panel.

All primary outcomes come directly from the survey questions. Most are measured on 7-point Likert scales and will be analyzed as continuous variables.

Safety perceptions: Four items on perceived safety, riskiness, stress, and physical demand. Each item will be analyzed separately; exploratory factor analysis may be used to check if they can be combined into an index.

Expected health risks: Three items on the likelihood of workplace accidents, physical health problems (e.g., neck/back pain), and mental health problems (e.g., stress, fatigue). Each will be analyzed separately; an index may be constructed if internal consistency is high.

Placebo outcome: One item on the perceived societal importance of the job.

Trust and effectiveness: Four post-treatment items measuring trust in the system to prevent accidents, improve mental health, improve physical health, and provide targeted recommendations.

Privacy and dignity: Two post-treatment items on whether the system respects workers’ privacy and dignity.

Job quality outcomes: Three items on job satisfaction, job meaningfulness, and perceived fair wages (categorical salary bands). Job satisfaction and meaningfulness are Likert outcomes; fair wage is ordinal and will be analyzed as both categorical and continuous (midpoints).

Where outcomes are measured both before and after the vignette, we will compute change scores (post–pre) as the main outcome.
We will standardize the outcomes for ease of interpretation.

Moderators: We will explore heterogeneity of treatment effects by baseline characteristics available in the LISS panel, including age, gender, education, income, employment status, and urban/rural residence. We will also test heterogeneity by personality traits.

Engagement measures: Response times, survey completion, and item non-response will be examined and may be used as indicators of participant engagement with the vignette.

Open comments: Any qualitative comments provided at the end of the survey will be coded and may be used descriptively to assess respondents’ reactions to the vignettes and treatments.

This study is a survey experiment with vignettes conducted in the Dutch LISS panel. A total of 2,100 adult respondents are randomly assigned in advance to one of three groups of roughly equal size: (1) an AI-only safety system, (2) a Human-only safety system, or (3) a Hybrid AI + Human system. All respondents first read a short description of a factory worker under standard conditions. They then read one of the three safety system vignettes depending on their assigned group.

Some outcomes (safety, risk, health expectations, job satisfaction, meaningfulness, and fair wages) are measured both before and after the vignette, allowing for within-subject comparisons of changes. Other outcomes (trust in the system, perceived effectiveness, and respect for privacy and dignity) are measured once, after the vignette, and are compared between groups. Randomization is implemented by Centerdata to ensure balance across demographic characteristics.

This project uses a survey experiment with vignettes implemented in the probability-based Dutch LISS panel. A total of 2,100 adult respondents are pre-randomized by Centerdata into three groups of about 700 each:

AI-only safety system: Factory introduces continuous AI monitoring with cameras and smart wearables, which detect risks (e.g., fatigue, overheating, equipment failure) and provide automatic alerts.

Human-only safety system: Factory introduces monitoring by trained safety supervisors who walk the floor, detect risks, and provide safety alerts.

Hybrid AI + Human system: Factory introduces both AI monitoring and human supervisors, with supervisors also reviewing AI alerts before issuing recommendations.

Survey procedure:

Respondents first read a baseline vignette describing a male factory worker with standard occupational health and safety rules.

They answer baseline questions on safety, risk, stress, physical demand, accident and health risks, societal importance of the job, job satisfaction, meaningfulness, and fair wages.

They are then exposed to one of the three intervention vignettes.

Post-vignette, they answer the same questions on safety, risks, health, societal importance, job satisfaction, meaningfulness, and wages (enabling within-subject comparisons).

They also answer additional post-treatment questions on trust in the system, perceived effectiveness, and whether the system respects workers’ privacy and dignity (analyzed as between-subject outcomes).


Comprehension checks:
After reading the assigned vignette, respondents complete one multiple-choice question to verify they understood which safety system was described. Respondents are given two chances to answer correctly. If they fail both times, they can continue the survey, but in the analysis, we will test robustness to excluding these cases. Results will be reported both including and excluding respondents who did not pass the comprehension check.

Design features:

Randomization is preloaded by Centerdata to ensure balance across age, gender, education, income, employment status, and urban/rural location.

Survey completion time averages nine minutes. Respondents can return to the vignette text while answering to reduce measurement error.

Engagement will be monitored through response times, item non-response, and survey completion.

This design allows estimation of both within-subject treatment effects (changes in safety, risk, health, job quality, and wages) and between-subject treatment effects (trust, effectiveness, privacy, and dignity).

Randomization was conducted in advance by Centerdata using computerized assignment. Respondents were pre-randomized by a computer into one of three treatment arms (AI-only, Human-only, Hybrid AI+Human). The randomization was stratified to ensure balanced distributions across key socio-demographics (age, gender, education, income, employment status, urban/rural location). After assignment, Centerdata checked the distributions with crosstabs and χ² tests and confirmed that respondents in the three groups do not differ systematically on these baseline characteristics

The unit of randomization is the individual respondent. Each panel member in the LISS sample was pre-randomized by computer into one of the three treatment arms (AI-only, Human-only, Hybrid AI+Human). No clustering or higher-level randomization was applied.

No

2100 observations

2100 respondents

2100 respondents, 700 control, 700 treated in condition AI, 700 treated in condition AI+human

With approximately 2,100 respondents (≈700 per treatment arm), randomization at the individual level, and no clustering, the minimum detectable effect size for the main outcomes is about 0.18 SD for two-sided tests at α = 0.05 and power = 0.80. This corresponds to detecting differences of about 18% of a standard deviation in perceived safety, risk, and job quality outcomes.