Partially supported by a NULab Seedling Grant.
Large language models (LLMs) can generate human-like text and excel in various natural language processing tasks, catalyzing a range of applications. However, the potential misuse of LLMs for generating misinformation and harmful content is a growing concern. For instance, studies have shown that models like the GPT series can produce news articles indistinguishable from human-generated ones and vast amounts of compelling mis/disinformation with little human involvement.
This project investigates the capability of LLMs to impersonate humans on social media. Recent studies show that LLMs can effectively role-play, exhibit diverse characteristics when interacting with humans, and capture political views when fine-tuned with social media data. These findings suggest that LLMs can mimic humans, although a systematic analysis is lacking. To close this gap, this project tests ChatGPT’s ability to emulate individual Twitter users with different traits and studies how humans perceive and react to synthetic personas and content.
The results will clarify whether and how LLMs can adequately mimic social media users, which could facilitate cost-effective and ethically responsible experiments and serve as an alternative method for generating social media data. From a trust and security perspective, the findings can provide invaluable insights into the potential misuse of LLMs for deceptive purposes. Furthermore, analyzing human reactions to these abuses will assist in identifying vulnerabilities and formulating effective countermeasures.
Principal Investigator
Kai-Cheng Yang, Postdoctoral Researcher, Network Science Institute
