Initial Attempts to Detect or Screen Out AI Responses Prove Elusive in the Age of Agentic AI

Abstract

OBJECTIVE: Although screening for bots and/or using costly panel services for recruiting participants online has become increasingly necessary, such efforts may no longer ensure the validity of data collected online. Newly released agentic AI models, such as the ChatGPT agent, have the ability to complete surveys relatively indistinguishably from humans. METHODS: The current paper outlines efforts that the body image, weight, and eating disorders (BIWED) lab has undergone to screen for and detect AI data completion reliably and validly. RESULTS: There are some tasks that ChatGPT agents do not perform identically to human responders (e.g., video tasks, online games, open-ended responses, and reCAPTCHA). We present the methods that have been the most successful at identifying AI agent survey completion. DISCUSSION: We discuss potential solutions, field-wide concerns, and future directions for the field more broadly.

Ähnliche Arbeiten