Keywords: AI, LLM, web survey, data quality, gender attitudes

Self-administered and online surveys lack the human interaction present in interviewer-administered surveys. Recent advancements in generative large language models allow for using AI avatars in surveys. AI avatars hold promises to enhance survey-taking experience. At the same time, incorporating the human element can introduce social desirability bias, especially on sensitive topics, leading to potential interviewer effects. In this study, we ask whether the gender of the avatar impacts the measurement of the gender attitudes of the respondent. We use web survey data collected in Hungary in October 2023 from an opt-in panel. In a pre-registered experiment, respondents were randomly assigned to one of the three experimental conditions at the beginning of the completion. Group A received a condition where a male AI avatar read the survey question. Group B received a similar condition but with a female avatar. Group C received the question in the standard, non-video mode (control group). We use https://app.heygen.com/home to generate the AI videos. We used measures from the R11 ESS gender module. The results will be presented at the conference. The experiment can help in understanding the benefits and caveats of using AI avatars in online surveys. In particular, we contribute by examining whether AI avatars are prone to interviewer-effect-type biases, similar to those observed in human interactions during interviewer-administered surveys.