ESRA 2027 sessions by theme

Back to Overview of Sessions

Session Details

The rapid development of large language models (LLMs) has turned a longstanding data quality concern into a potentially existential one. Demonstrations show that LLM-powered agents can complete surveys while maintaining coherent personas, producing internally consistent responses, and evading quality checks designed for humans. If researchers cannot trust that their respondents are human, the evidential foundation of survey research itself is called into question. Public opinion measurement, official statistics, and the behavioural sciences all depend on that foundation. Yet capability is not prevalence,  demonstrating that an agent can defeat detection is not evidence that agents are being deployed within the panels researchers actually use. Adjacent threats, including LLM-assisted human responding, traditional script-based bots, and systematic variation in human data quality across recruitment ecosystems, are frequently conflated. The field currently lacks a shared map of where the genuine risks to inference reside. Without one, it risks responding to the wrong threat with the wrong tools.

This session aims to build that map. We bring together researchers working on AI-related threats to online survey research, spanning conceptual, empirical, and methodological contributions. We welcome taxonomies that distinguish the threat landscape’s distinct actors and behaviours, from autonomous agents and LLM-assisted humans to conventional bots and low-effort respondents. We invite empirical estimates of prevalence, particularly cross-platform studies assessing whether infiltration varies by recruitment infrastructure, work on detection and mitigation strategies, and critical examinations of what current methods can and cannot distinguish. We equally encourage analyses of how fraud in online surveys has changed over time in response to technological progress and forward-looking frameworks for future-proofing survey research as agent capabilities advance further. We welcome contributions from academic, commercial, and official statistics settings. By replacing generalised alarm with precise, evidence-based understanding, the session seeks to help