Making Workers AI Ready: Behavioral Interventions and AI Adoption Among Workers

This study examines whether informational interventions and transparency can reduce resistance toward artificial intelligence (AI) in the workplace and increase openness to AI adoption. Many workers report feeling stressed or threatened by AI, even though existing evidence suggests limited negative effects on overall employment or mental health. Such perceptions can slow the adoption of AI tools and reduce engagement with reskilling opportunities.

Following the approach of Celebi et al. (2026), participants complete an initial screening Survey 0 (S0) designed to identify valid, attentive respondents before entering the main study. We will focus on US based workers aged 18-65. This screening stage includes attention checks, comprehension questions, and basic demographic measures to ensure that participants are reading instructions carefully and engaging thoughtfully with the survey content. Only participants who successfully meet the screening criteria are invited to continue to Survey 1 (S1).

This prescreening approach helps improve data quality and ensures that subsequent intervention effects are measured among participants who are genuinely engaged with the study materials.

Participants who pass the data quality checks are invited to S1, where they are randomly assigned to one of four conditions in a 2x2 factorial design crossing two independent interventions.

The first is an informational intervention highlighting evidence that the use of AI tools can increase productivity, designed to encourage hands-on engagement with generative AI tools in a low-stakes context.

The second is an AI transparency policy statement, presented before participants complete a productivity task, which explicitly informs participants that they are free to use or not use AI tools and that this choice has no consequences for their payment or future study participation. This transparency statement is designed to reduce concerns about being monitored or judged for their AI use decisions, thereby lowering social and psychological barriers to experimentation with AI tools. Participants assigned to the transparency condition also complete a comprehension question to verify their understanding of the policy statement.

Participants will also be invited to complete a follow-up Survey 2 (S2), contingent on funding. The purpose of S2 is to measure the longer-run effects of the interventions on AI-related attitudes, AI adoption, perceptions, well-being, health, and productivity-related outcomes.

2026-04-21

2026-05-26

1) Task performance
Productivity in the writing or coding task, measured by quality ratings and completion time. A standardized productivity index will be constructed from these measures.
2) Task experience
Task-specific self-efficacy, perceived realism, and enjoyment.
3) Behavioral outcomes
AI adoption during the task, measured by whether participants choose to use a generative AI tool when its use is optional; persistence and engagement in completing the task, including time spent on the task and time allocated across different task components (brainstorming, drafting, and editing); and self-reported AI use outside the study.
We will adjust for multiple hypothesis testing by organizing outcomes into pre-specified families and controlling the false discovery rate (FDR) within each family using the Benjamini-Hochberg procedure (Benjamini and Hochberg, 1995; Anderson, 2008).

As secondary outcomes, we will also analyze participants' perceptions of AI's broader effects on work, productivity, health, and well-being. We also measure expected changes in participants' use of generative AI tools over the next six months. Post intervention, we also measure perceived benefits of AI (Humlum and Vestergaard, 2025) and predictions for labor market effects (e.g., displacement, if any).
We will also explore heterogeneous treatment effects across a set of pre-specified baseline characteristics. These include psychological and attitudinal measures such as general attitudes toward AI (ATTARI-12), AI reliability, trust in AI, growth mindset, reskilling intentions, feelings of guilt, ethical, privacy, and transparency concerns, imposter syndrome, and sense of agency. We will also examine heterogeneity by beliefs about AI's labor market impacts across workers of different ages and genders, as well as by prior work experiences, including displacement, working hours, and collaboration at work. In addition, we will consider heterogeneity across individual characteristics such as, for instance, occupation, industry, education, field of degree, age, gender, socioeconomic background, prior AI exposure and familiarity, loss aversion, risk preferences, ambiguity aversion, time preferences, interpersonal trust, trust in science, Big Five personality traits, cognitive skills, work experience, decision involvement, firm type, and firm size.

This study uses a randomized online experiment to test whether informational interventions and transparent AI-use policies can increase AI adoption and influence both task outcomes and workers’ experience with AI.

Stage 0 (S0): Screening for valid subjects

Following the approach of Celebi et al. (2026), participants first complete a screening survey designed to identify valid and attentive respondents before entering the main study. This stage includes attention checks, comprehension questions, and basic demographic measures to ensure that participants are reading instructions carefully and engaging thoughtfully with the survey content. Only participants who successfully pass the screening criteria are invited to continue to Survey 1 (S1). Based on previous pilots, we expect approximately 3,000 participants to pass the S0 screening and complete S1.

Survey 1 (S1): 2x2 factorial design

In S1, participants are randomly assigned to one of four conditions generated by crossing two independent interventions in a 2x2 factorial design. Randomization is implemented automatically when participants begin the survey.

The first factor is an informational AI-productivity intervention. Participants assigned to this treatment receive a brief message highlighting evidence from prior research showing that AI tools can increase productivity. This intervention is designed to encourage hands-on engagement with AI in a low-stakes context. Participants in the control condition do not receive this information.

The second factor is a transparent AI-use policy statement. Participants assigned to the transparency condition see the following statement before completing the task: “In this task, you may choose whether or not to use artificial intelligence (AI) tools (e.g., ChatGPT, Claude, Copilot, Gemini, Perplexity, or similar AI tools).

In many workplaces today, employees are allowed to use AI tools to support their work, while remaining responsible for the final output. Similarly, in this task, you are free to decide whether to use AI tools or not.

Your choice will not affect your payment or your participation in this study. “
This statement is designed to reduce concerns about being monitored or judged for AI-use decisions, thereby lowering social and psychological barriers to experimentation. Participants in the control condition do not receive this statement. Participants in the transparency condition also complete a comprehension question to verify that they have read and understood the policy statement.

The crossing of these two factors produces four experimental conditions: (A) Information + Transparency, (B) Information only, (C) Transparency only, and (D) Control, with approximately 750 participants per cell.

Survey structure and outcome measurement

The survey is structured to distinguish pre-treatment moderators from post-treatment outcomes. Before treatment, participants complete a battery of attitudinal, behavioral, and psychological measures, with question order randomized across participants. These include AI-related attitudes (ATTARI-12), baseline self-efficacy, growth mindset, trust in AI, trust in technology, trust in science, empowerment, sense of agency, feelings of guilt, ethical concerns, privacy and transparency attitudes, beliefs about AI's labor market impacts by age and gender, loss aversion, risk preferences, ambiguity aversion, time preferences, Big Five personality traits, cognitive skills, and collaboration and interdependence. These variables are measured before treatment to reduce priming and are used as pre-specified moderators in the heterogeneity analysis.Following exposure to the assigned intervention(s), participants complete an occupation-specific incentivized productivity task, such as a writing or coding task, with optional access to a generative AI tool.Primary outcomes are organized into three families. First, task performance outcomes include productivity, measured by quality ratings and completion time, combined into a standardized productivity index. Second, task experience outcomes include task-specific self-efficacy, perceived realism, and enjoyment. Third, behavioral outcomes include AI adoption during the task, persistence and engagement, time spent, time allocation across different task components (brainstorming, drafting, and editing), and self-reported AI use outside the study.Post-intervention, we also measure perceived benefits of AI (Humlum and Vestergaard, 2025) and predictions for labor market effects (e.g., displacement, if any). As secondary outcomes, we will analyze participants' perceptions of AI's broader effects on work, productivity, health, and well-being, as well as self-reported well-being and health (life satisfaction, job satisfaction, mental health) measured post-treatment in S1 and again at S2. We also measure expected changes in participants' use of generative AI tools over the next six months.Sociodemographic and employment characteristics are also collected after treatment, as these are stable individual attributes that cannot be affected by the intervention.Following S1, participants will be invited to complete a follow-up Survey 2 (S2), contingent on funding. The purpose of S2 is to measure the longer-run effects of the interventions on AI-related attitudes, AI adoption, perceptions, well-being, health, and productivity-related outcomes. S2 will also assess whether the effects observed in S1 persist over time.

Not available

Participants will be individually randomized using the randomization functions embedded in the online survey platform (e.g., Qualtrics or Prolific assignment tools). Randomization will be implemented automatically at the time participants begin Survey 1 (S1) to ensure allocation concealment.
In Survey 1 (S1), participants will be randomly assigned to one of four conditions resulting from the 2x2 factorial design crossing two independent factors: (1) whether they receive the informational AI-productivity intervention or not, and (2) whether they receive the AI policy transparency statement or not. Assignment uses even-presentation block randomization to ensure roughly equal allocation across the four conditions.

In Survey 1 (S1), each participant is independently assigned to one of four conditions resulting from the 2x2 factorial design crossing of the informational intervention and the AI policy transparency statement.

No

N/A

We expect approximately 3,000 participants to pass the S0 screening and proceed to complete S1.

Stage 0 (S0):

Initial screening: ~ up to 9,000 participants targeting 3000 in S1
Expected valid participants proceeding to and completing S1 (based on previous pilots): ~3,000

Survey 1 (S1) – Four experimental conditions (2x2 factorial design):

Condition A (Information + Transparency): ~750 participants
Condition B (Information only): ~750 participants
Condition C (Transparency only): ~750 participants
Condition D (Pure Control): ~750 participants
Total: ~3,000 participants

Power calculations focus on detecting treatment effects in Survey 1, using main effects from the 2x2 factorial design. We set α = 0.05 (two-sided) and a target power of 0.80. For main effects, each treatment arm is compared against its pooled counterpart (e.g., all information participants vs. all no-information participants), yielding N = 1,500 per comparison. For interaction effects (e.g., the effect of the information within a given transparency condition), the relevant comparison involves N = 750 per arm. Power calculations assume AI use at around 50% at baseline and a standard deviation of 1.2 points on Likert-scale outcomes, both in line with previous pilots. Under these assumptions, the minimum detectable effect (MDE) for main effects is 5.1 percentage points for binary outcomes and 0.123 points on a 6-point Likert scale. For interaction comparisons, the MDE is 7.23 percentage points for binary outcomes and 0.174 points on the Likert scale.