Identifying Bots Through LLM-Generated Text in Open Narrative Responses: A Proof-of-Concept Study

Online survey participants are frequently recruited through social media platforms, opt-in online access panels, and river sampling approaches. Such online surveys are threatened by bots that shift survey outcomes and exploit incentives. In this proof-of-concept study, we advance the identification of bots driven by Large Language Models (LLMs) through the prediction of LLM-generated text in open narrative responses. We conducted an online survey on same-gender partnerships, including three open narrative questions, and recruited 1512 participants through Facebook. In addition, we utilized two LLM-driven bots, each of which responded to the open narrative questions 400 times. Open narrative responses synthesized by our bots were labeled as containing LLM-generated text (“yes”). Facebook responses were assigned a proxy label (“unclear”) as they may contain bots themselves. Using this binary label as ground truth, we fine-tuned prediction models relying on the “Bidirectional Encoder Representations from Transformers” (BERT) model, resulting in an impressive prediction performance: The models accurately identified between 97% and 100% of bot responses. However, prediction performance decreases if the models make predictions about questions they were not fine-tuned with. Our study contributes to the ongoing discussion on bots and extends the methodological toolkit for protecting the quality and integrity of online survey data.