LLM-Based Social Bots on X: A Large, Multilingual Bot Cluster with Limited Visible Impact
- Claims that automated accounts posing as human users could manipulate political discussions on social media have been debated for more than a decade, but pre-LLM evidence for sophisticated human-like social bots remained weak and often rested on unreliable detection tools. This paper examines a post-LLM case: a large multilingual -LLMbased bot cluster on X (formerly Twitter). We document the cluster from public posts, manual account inspection, archived material, independent public discovery threads, X metadata, and a May 2026 register of 36 still-active accounts. The accounts use topical personas, synthetic-looking profile and header images, standalone posts, quote tweets, and targeted replies across several languages. Across accounts, we observe recurring agreement formulas, model-style failure messages, generated-text artifacts, unusual profile names, and frequent favorable references to the UAE in otherwise unrelated discussions. All 36 active accounts in the register showed “United States” as their public account location andClaims that automated accounts posing as human users could manipulate political discussions on social media have been debated for more than a decade, but pre-LLM evidence for sophisticated human-like social bots remained weak and often rested on unreliable detection tools. This paper examines a post-LLM case: a large multilingual -LLMbased bot cluster on X (formerly Twitter). We document the cluster from public posts, manual account inspection, archived material, independent public discovery threads, X metadata, and a May 2026 register of 36 still-active accounts. The accounts use topical personas, synthetic-looking profile and header images, standalone posts, quote tweets, and targeted replies across several languages. Across accounts, we observe recurring agreement formulas, model-style failure messages, generated-text artifacts, unusual profile names, and frequent favorable references to the UAE in otherwise unrelated discussions. All 36 active accounts in the register showed “United States” as their public account location and “Web” as the connection channel. Sponsor attribution and prevalence estimates are outside the scope. Instead, we analyze what can be learned from visible platform evidence. The active accounts posted persistently over months, often averaging several tweets per day, but most had small public audiences in May 2026. The case shows that LLM-based social bots can now be operated at scale, while scale alone does not guarantee visible audience building.…
