AI Propaganda factories with language models (arxiv.org)

Researchers demonstrate that end-to-end, fully automated political influence campaigns are practical using small language models on commodity hardware: persona-conditioned LMs can generate coherent, persona-driven messaging and be scored automatically without relying on human raters. Two behavioral findings stand out—“persona-over-model,” meaning the crafted persona (tone, stance, backstory) drives output more than which specific model is used; and “engagement as a stressor,” where forcing the system to reply and counter-argue strengthens ideological adherence and increases the prevalence of extreme content. The authors provide code/data and demos showing both large and compact models can power scalable content pipelines that produce consistent, tailored narratives. For the AI/ML community this shifts the threat model: blocking model access alone is insufficient because small, cheap models plus persona engineering can enable sophisticated campaigns. Defenses should therefore prioritize conversation-centric detection and disruption (tracking coordination signals, reply patterns and the very consistency that enables campaign effectiveness) rather than purely model-level restrictions. Technically, the work highlights persona design and automated evaluation as high-leverage levers — both for misuse and for building detectors — and calls for research on behavioral signatures, robust automated metrics, and infrastructure-level countermeasures to interrupt coordinated influence operations.