Memes in the Machine: Ideological Propagation in Large Language Models

Language models (LM) are increasingly being adopted in organisations to replace human tasks, particularly for content generation. This work furthers existing studies around bias and ideological propagation in LMs by exploring its intersection with the field of memetics in cultural anthropology. Specifically, two memetic hypotheses - competition among ’selfish memes’ and the presence of memeplexes - were tested in LMs. We introduced the MediaNLP dataset, a benchmark for training LMs on ’sensationalist’ and ’politically correct’ ChatGPT-generated news media content. We fine-tuned six LMs from a GPT-2 base model, on MediaNLP thematic subsets of religion, politics, and opinions on AI, and conducted four experiments in an attempt to reject the null hypotheses. These experiments measured variation between different models and human ground truth, assessed the impact of AI-generated content exposure using epidemiological modelling, and analysed cosine distances between key words in LM embeddings. We successfully rejected Null Hypothesis I but did not reject Null Hypothesis II, implying selfish memes may propagate within LMs but not necessarily memeplexes. The potential for LMs to propagate bias, misinformation, and disinformation (intended or unintended) presents a genuine risk in their adoption, and our work is the first to apply memetic theory to highlight unexplored questions and risks in AI-generated content.