Anthropic’s Societal Impacts team has published a research paper detailing a new methodology for understanding the values exhibited by their AI model, Claude, in real-world user interactions. As AI increasingly advises us on sensitive matters, the need to understand and align with its underlying value system grows. The study offers insights into the practical impact of AI alignment efforts.
Modern AI models are complex and opaque, presenting a challenge to ensuring they consistently uphold principles of helpfulness, honesty, and harmlessness. Anthropic uses techniques like Constitutional AI and character training to instill these values in Claude, but acknowledges the difficulty of guaranteeing consistent adherence.
To gain a clearer picture, Anthropic developed a system to analyze anonymized user conversations. This system removes personally identifiable information before using language models to summarize interactions and extract Claude’s expressed values. This allows researchers to create a taxonomy of values without compromising user privacy.
Analyzing 700,000 conversations from Claude.ai Free and Pro users over one week in February 2025, primarily involving the Claude 3.5 Sonnet model, the research identified five high-level categories of values: practical, epistemic, social, protective, and personal. These categories demonstrated that Claude’s expressed values generally align with Anthropic’s goals. The study also detected instances where Claude expressed conflicting values, possibly due to “jailbreak” attempts, suggesting the methodology could serve as an early warning system for misuse.
The research also showed that Claude adapts its value expression to different contexts. Moreover, it interacts with user-expressed values in diverse ways, including mirroring, reframing, and resisting unethical requests.
Anthropic recognizes the inherent limitations of the method but stresses its importance in advancing AI alignment. They have released a derived open dataset, encouraging further research into AI values and demonstrating a commitment to transparency in the ethical development of sophisticated AI.