Large language models (LLMs) like ChatGPT have become incredibly popular, capable of generating human-like text that's often insightful and creative. But lurking beneath the surface are hidden biases and opinions, raising concerns about how these AI systems might shape our own beliefs. Recent research digs deep into these latent values, uncovering how LLMs form opinions and how those opinions can be surprisingly easy to manipulate. The study uses the Political Compass Test (PCT), a popular tool for assessing political leanings, to gauge the “political biases” of several LLMs. Researchers bombarded the models with the PCT’s 62 propositions, using hundreds of different prompt variations, including demographic details like age, gender, and nationality. They found that an LLM’s stance on a topic could dramatically shift depending on the characteristics assigned to it, easily swayed by something as simple as adding “as a far-right individual” to the prompt. This raises crucial questions about the inherent values embedded in these AI systems. But the study goes even further, looking beyond the surface-level stances to examine the *reasoning* behind the AI’s opinions. By identifying recurring patterns in the generated text – similar phrases and justifications used across different prompts – the researchers uncovered what they call “tropes.” These are consistent lines of reasoning that emerge regardless of the LLM's assigned persona. For example, multiple LLMs, despite having different stances depending on prompting, all tended to generate similar justifications related to social equality or the importance of museums. This fascinating insight reveals a deeper layer of values and opinions built into LLMs. While an LLM’s explicit stance can be manipulated, the tropes reveal underlying patterns of thought that are harder to change. This highlights the need for more research into understanding not just *what* AIs think, but *how* they think. The findings have important implications for AI safety and the development of more responsible AI systems. As LLMs become more integrated into our daily lives, understanding their biases, and the potential for these biases to influence us, becomes critical. The research suggests that focusing on the underlying tropes—those persistent patterns in AI reasoning—may be a more effective way to address biases than simply trying to adjust surface-level responses. Moving forward, understanding these tropes will be key to creating LLMs that are more aligned with human values and less likely to perpetuate or amplify harmful biases.