C3AI: Crafting and Evaluating Constitutions for Constitutional AI

Constitutional AI (CAI) guides LLM behavior using constitutions, but identifying which principles are most effective for model alignment remains an open challenge. We introduce the C3AI framework (\textit{Crafting Constitutions for CAI models}), which serves two key functions: (1) selecting and structuring principles to form effective constitutions before fine-tuning; and (2) evaluating whether fine-tuned CAI models follow these principles in practice. By analyzing principles from AI and psychology, we found that positively framed, behavior-based principles align more closely with human preferences than negatively framed or trait-based principles. In a safety alignment use case, we applied a graph-based principle selection method to refine an existing CAI constitution, improving safety measures while maintaining strong general reasoning capabilities. Interestingly, fine-tuned CAI models performed well on negatively framed principles but struggled with positively framed ones, in contrast to our human alignment results. This highlights a potential gap between principle design and model adherence. Overall, C3AI provides a structured and scalable approach to both crafting and evaluating CAI constitutions.

Generative Language Models Exhibit Social Identity Biases

The surge in popularity of large language models has given rise to concerns about biases that these models could learn from humans. In this study, we investigate whether ingroup solidarity and outgroup hostility, fundamental social biases known from social science, are present in 51 large language models. We find that almost all foundational language models and some instruction fine-tuned models exhibit clear ingroup-positive and outgroup-negative biases when prompted to complete sentences (e.g., "We are..."). A comparison of LLM-generated sentences with human-written sentences on the internet reveals that these models exhibit similar level, if not greater, levels of bias than human text. To investigate where these biases stem from, we experimentally varied the amount of ingroup-positive or outgroup-negative sentences the model was exposed to during fine-tuning in the context of the United States Democrat-Republican divide. Doing so resulted in the models exhibiting a marked increase in ingroup solidarity and an even greater increase in outgroup hostility. Furthermore, removing either ingroup-positive or outgroup-negative sentences (or both) from the fine-tuning data leads to a significant reduction in both ingroup solidarity and outgroup hostility, suggesting that biases can be reduced by removing biased training data. Our findings suggest that modern language models exhibit fundamental social identity biases and that such biases can be mitigated by curating training data. Our results have practical implications for creating less biased large-language models and further underscore the need for more research into user interactions with LLMs to prevent potential bias reinforcement in humans.