The Effects of Artificial Intelligence on Career Choice

Participants complete an online survey about artificial intelligence, education, and career choices. During the survey, they are randomized to one of two pieces of information about the potential effects of AI on labor-market opportunities. In a second treatment arm, they are randomized to information about the AI exposure of educational or vocational fields they are considering. The experiment studies how this information affects beliefs about AI, expectations about future career prospects, and preferences over educational and vocational fields.

2026-06-01

2026-12-31

The posterior-belief outcomes allow us to test whether the information interventions shift expectations about the future exposure of fields to AI and about the corresponding future career prospects.
We will also test whether information about AI changes educational and career preferences. The key field-choice outcome compares the two fields each respondent is considering. For example, if a respondent is considering both gardening and economics, we test whether information showing that gardening is less exposed to AI than the respondent initially believed, relative to economics, makes the respondent more likely to prefer gardening over economics. In addition to the stated preferences from the survey, we will also attempt to link respondents to administrative records on realized educational choices, including applications, rankings, enrollment, and field of study.

The secondary outcomes help us understand the channels through which information about AI affects educational and career choices. For example, we can distinguish whether changes in field preferences are driven by beliefs about earnings, job stability, fulfillment, social impact, family approval, or work-life balance.

Participants are invited by email to complete an online survey. The target population consists of individuals over the age of 18 in Denmark who are about to graduate from high school or who graduated from high school within the past three years. Some respondents (e.g. those who are already enrolled in college) are routed out of the main experimental module and complete a shorter survey about general AI use and attitudes. Additionally, we will need to exclude some subjects for reasons that are unobservable ex-ante (e.g., those that by the time they start our survey had already submitted their educational choices).

The study cross-randomizes two treatments.

First, respondents are randomly assigned with equal probability to one of two general AI labor-market narratives:
- Labor-augmenting information.
- Labor-replacing information.

Before this information is provided, respondents report their beliefs about whether AI has led to better or worse job opportunities in fields exposed to AI. After this information is provided, respondents report their expectations about whether AI will lead to better or worse job opportunities in AI-exposed fields by 2030.

Second, respondents are randomly assigned with equal probability to one of two field-specific AI-exposure information conditions:
- Control group: no field-specific AI-exposure information.
- AI-exposure information treatment: information about the current estimated share of tasks that can already be performed autonomously by AI in each of the two fields the respondent is considering.

Before this information is provided, respondents report their beliefs about the current AI exposure of each of the two fields they are considering. After this information is provided, respondents report their expectations about the AI exposure of each field by 2030 and about the labor-market prospects of each field by 2030.

The survey elicits prior and posterior beliefs, allowing us to study heterogeneity by baseline beliefs and to test whether belief updating mediates the treatment effects. If pertinent, we will estimate the causal effects of beliefs using the same two-stages least square model from Cullen & Perez-Truglia (JPE, 2022).

We will also explore other forms of heterogeneity, such as by gender and by AI proficiency.

Not available

Randomization done by the survey software

Individuals

No

We will send email invitations to over 100,000 individuals, but we do not know what the response rate will be

Same as cluster

- In the first treatment arm, we randomly assign 50% of the subjects to the labor-augmenting information and the other 50% to the labor-replacing information.
- In the second treatment arm, we randomly assign 50% of the subjects to receive information about the current AI-exposure in the two fields of study and the other 50% to no information.