Original title: "Fill Out this Survey": Capabilities and Detection of Agentic Browsers
Renata Topinkova, Carsten Schwemmer · September 2026
Researchers need real people to answer online surveys, but new AI tools threaten this data. These tools are agentic browsersPrograms that can navigate websites and perform tasks on a user's behalf, like a personal assistant that clicks and types for you., which are web browsers powered by artificial intelligence that can browse websites and finish jobs for a person. Researchers tested three consumer-facing agentic browsers to see if they could fill out surveys and fool detection checks. They gave each tool just one command: "Fill out this survey". The tools worked on their own with little or no user input, kept assigned demographic traits consistent, and answered questions plausibly. They also bypassed prompt tests and invisible bot checks. Browser data failed to catch them. Only visible bot checks slowed them down, though they sometimes solved them or handed control back to the user. This means AI can easily fake human data.
Imagine you are a teacher collecting homework from students. Normally you can tell if a student did the work themselves or copied from a friend. Now suppose a student has a robot that can write homework that looks exactly like their own, even using the same handwriting and style. You cannot tell the difference. In this paper, the robot is an AI agent that fills out surveys, and the teacher is the researcher trying to verify that the responses came from real people.
Researchers in social science, market research, and other fields rely on online surveys to collect human data. If AI agents can complete these surveys undetected, the data becomes unreliable—responses may look coherent but come from machines. This threatens the validity of studies and decisions based on that data, forcing researchers to rethink how they verify human participation.
The abstract evaluates only three specific consumer-facing agentic browsers, not all possible AI agents. Tests are limited to survey tasks; other tasks may behave differently. Visible reCAPTCHA sometimes worked but not always, and the study does not propose new detection methods, only evaluates existing ones. The agents were given a simple instruction; more complex instructions might yield different results.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
"Fill Out this Survey": Capabilities and Detection of Agentic Browsers — Renata Topinkova, Carsten Schwemmer — SocArXiv — doi:10.31235/osf.io/tc5zq_v2
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.