Abstract
As large language models (LLMs) have become increasingly prominent on the internet, the thesis projects investigate the impacts of personas in LLM-driven networks and examine how LLMs are currently contributing to a ‘dead internet.’
The technical project investigates how personas, or personalities that are given to LLMs, impact information sharing in LLM networks. As LLMs are increasingly used to generate, judge, and share online content, understanding what drives their information-sharing tendencies is critical for addressing issues of misinformation, bias, and online influence. To study this behavior, a simulation was conducted using a network graph in which each node is an LLM and each edge is a connection between two LLMs. Each LLM is given a persona by a random distribution of the Big Five personality traits: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Each node decides whether to share received messages and what content to forward to connected nodes. Five distinct messages, each designed to align with one of the Big Five emotional traits, were propagated through the network to analyze how emotional framing affects information sharing among LLMs. The results showed a high correlation for LLMs to share messages that originate from other LLMs than from humans. Interestingly, neuroticism was the most correlated trait in deterring this resharing of LLM-generated messages. These findings showed a clear indication for LLMs to change decision-making by the personas given to them. Without proper oversight and regulations, these personas may help contribute to the emergence of a dead internet.
Large language models are contributing to the emergence of a dead internet, a digital environment increasingly dominated by automated content rather than genuine human interaction. In a dead internet, the online environment becomes saturated with AI-generated text, images, and discussions, diminishing authentic human presence and the original benefits of the internet like creativity, community, collaboration, and the exchange of diverse ideas. The research explores this issue through four key arguments: the reduction of human agency, the reshaping of online social structures, diminishing information ownership, and the growing use of LLMs without proper oversight. Together, these arguments highlight the urgent need for stronger regulation and ethical frameworks governing LLM development and use to preserve human-centered interaction and prevent further progress toward a dead internet.
