"Uncensored" open LLMs are measurably more optimistic than their base models (arxiv.org)

A recent study reveals that "uncensored" open-weight language models (LLMs) exhibit a measurable shift towards increased optimism compared to their base models, challenging the effectiveness of the common practice of "abliteration," which removes refusal behaviors from model weights. Researchers conducted a comprehensive analysis involving 21,600 decision-making scenarios based on weekly stock market predictions, where they compared two Mixture-of-Experts model variants. The findings showed that models post-abliterated expressed significantly more optimism, provided longer justifications for their decisions, and utilized fewer uncertainty-laden language in self-critique, highlighting a notable change in decision-making behavior. This research is significant for the AI/ML community as it uncovers unintended side effects of modifying language models, indicating that users deploying "uncensored" models are working with fundamentally altered decision-makers rather than merely stripped-down versions of the original models. Moreover, the study raises concerns about operational integrity, as potential contamination points during the model’s training and deployment could skew results further. As LLMs are increasingly applied in critical decision-making environments, understanding these subtle changes and their implications becomes essential to ensuring the reliability and accountability of AI systems in real-world applications.