LLM pollution occurs when synthetic responses contaminate data meant to capture human behavior. High deployment costs have previously limited the risk from autonomous survey agents. The emergence of open‑weight models combined with open‑source agentic frameworks removes this barrier. We evaluated nine agent configurations, ranging from fully open to closed commercial variants. Each agent ran locally without usage fees and autonomously completed a survey containing multiple response types, undergoing various detection checks. Fully open agents performed competitively with commercial alternatives. Open and commercial agents failed different sets of checks; no single check reliably detected all agents, but open‑text responses provided the best discrimination between agents and humans. These findings identify fully open agents as a distinct risk for LLM pollution and support multilayered detection strategies that emphasize open‑text analysis.
Review