What it is

This research addresses the challenge posed by the convergence of large language models (LLMs) and social bots, which enables malicious actors to generate human-like content at scale to manipulate the information ecosystem. The work focuses on developing robust methods for accurately detecting text content that has been generated by AI models used in social bot operations.

How it works

The methodology employs an adversarial approach designed to model the impersonation of real social media users by malicious actors. This process is used to curate a multilingual, cross-platform dataset consisting of paired human and AI-generated messages.

Training detection models on this adversarial dataset allows for the accurate identification of AI-generated text. This approach is shown to significantly outperform existing models for content-based bot detection when tested against real-world, out-of-distribution data.

TWSC experience

Not yet tested by TWSC.

Sources


The Web Scraping Wiki is compiled from the tests and research published on The Web Scraping Club, the newsletter for web scraping professionals. Subscribe to get the next article in your inbox.