Detection of LLM-powered bots using image classification
Abstract
In the rapidly changing landscape of online social interactions, the presence of automated accounts, or bots, has always posed a significant challenge to maintaining platforms where the information posted is authentic and reliable. The emergence of large language models (LLMs) may exacerbate this problem, as researchers have recently found families of automated accounts that use generative artificial intelligence to produce their posts. This paper focuses on this new type of bot, particularly on detecting them. Using a new detection technique that relies on image classification to distinguish between human and automated accounts, we demonstrate remarkable efficiency in identifying bots whose posts are generated by large language models. Our research improves the results of previous work on the detection of bot accounts powered by generative artificial intelligence.
