Generative AI Use and Loneliness

We run a survey experiment where the treatment group is nudged to have daily conversations about personal matters with a generative AI tool.

The study uses a two-wave online panel with individual randomization (1:1). In Wave 1, we collect sociodemographic and baseline outcomes from all participants, who are then randomly assigned to treatment or control. Treatment participants are instructed to converse daily with a generative AI tool for 28 consecutive days, recording privately the number of days they complied. Control participants continue their usual AI use and privately note days when personal-topic conversations occurred. In Wave 2 (≈4 weeks later), we collect the same outcome measures from all participants as well as the number of days they actually used AI for personal matters. The research team never accesses or records conversation content.

2026-01-26

2026-02-28

Life satisfaction (0–10 scale, single-item).
3-item UCLA Loneliness Scale (UCLA - 3): constructed from 3 items: lack of companionship, exclusion, isolation.
Self-reported loneliness.
Affect (two items): happiness yesterday, depression yesterday.

All outcomes will be standardized in the main regression tables. Regarding the 3-item UCLA Loneliness Scale (UCLA-3), Each item (‘lack of companionship,’ ‘feeling left out,’ ‘feeling isolated from others’) is answered on a 3-point Likert scale (1 = hardly ever/never, 2 = some of the time, 3 = often). Scores from the three items will be summed to create an index ranging from 3 to 9, with higher scores indicating greater loneliness. For analysis, this index will be standardized (mean 0, standard deviation 1) so that effect sizes are expressed in standard deviation units.

Number of days spent chatting about personal matters with an AI (first stage).
Frequency of use of AI for different matters (first stage)
Relationship satisfaction (romantic partner, if applicable).
Family relations satisfaction.
Friendship satisfaction.
Objective loneliness (meals alone, meetings with friends/relatives)
Perceived social support: whether the respondent has someone they can rely on.
Valuation of AI (willingness-to-accept to abstain to use for personal reasons for 1 week).
Pleasant or unpleasant personal conversations with AI.

Here, we merge both what is referred to as "first stage and secondary outcomes" in the PAP. If any inconsistencies arise between the information in the AEA RCT Registry and this PAP pdf file, the latter shall take precedence.

Participants are recruited from an online panel in France and complete two survey waves spaced four weeks apart. The unit of randomization and analysis is the individual. After Wave 1, participants are randomly assigned in equal proportions to treatment or control. Outcomes are measured again in Wave 2 using the same instruments, allowing comparison across arms while controlling for baseline measures. Randomization is implemented by the survey platform.

This is a two-wave randomized controlled trial with equal assignment at the individual level. Wave 1 collects baseline covariates and outcome measures. Randomization is simple 1:1 (treatment vs. control) within the survey platform. Wave 2 occurs approximately 28 days later with identical outcomes measures.

Randomization performed by the survey company.

Individual

No

10,000 survey respondents

10,000 survey respondents

5,000

Two-arm RCT, equal allocation, two-sided α = 0.05. We expect 20% attrition between baseline and endline surveys. Total N = 8,000. Primary outcomes are standardized (unit: SD). Under these assumptions, the minimum detectable effect (MDE) at 80% power is 0.06 SD. Accordingly, the power to detect an effect of 0.10 SD (a benchmark often used in survey experiments) exceeds 99%. These estimates are conservative, since they do not incorporate baseline outcome adjustment (which would reduce residual variance).

PAP