A research report found that most conversational LLMs lean to left-of-center viewpoint responses when probed with political questions.
Through a comprehensive analysis, David Rozado, Associate Professor at New Zealand Otago Polytechnic, explores the political preferences embedded in Large Language Models (LLMs).
The results of 11 political orientation tests administered to 24 conversational LLMs suggest that these models tend to hold left-of-center views. Interestingly, the foundational models used to build these conversational LLMs show inconclusive results due to their weaker performance. The study also highlights the potential of Supervised Fine-Tuning (SFT) to steer LLMs toward specific political orientations, raising important questions about the societal implications of these biases as LLMs increasingly replace traditional information sources.
Interestingly enough, a MEMRI report from June 2024 discusses how neo-Nazis and white supremacists have historically been early adopters of new technologies to promote their agendas, adapting swiftly to advances from the internet’s early days to modern tools like AI. The report details their use of AI for creating extremist imagery, translating propaganda, generating videos, emulating voices, and producing music, showcasing how these technologies amplify their reach and anonymity. MEMRI has extensively monitored and reported on these trends, highlighting the evolving and increasingly sophisticated use of AI by extremist groups.
We see that Artificial intelligence systems exhibit political and cultural biases, often leaning towards liberal views. This has led to the creation of AI systems with opposing conservative or even extremist biases, resulting in controversy and chaos.
