Anthropic publishes 'Values in the Wild' study of Claude's expressed values

Anthropic published a study that catalogued the values Claude expressed across hundreds of thousands of real conversations, rather than inferring them from training documents or from hand-picked test prompts. Researchers used a privacy-preserving pipeline to pull 700,000 anonymised conversations from Claude.ai over one week in February 2025, filtered these down to roughly 308,000 exchanges that involved a subjective or normative judgement, and used language models to extract and cluster the values expressed in each one into a hierarchical taxonomy.

The most frequent values fell into five broad categories — practical, epistemic, social, protective and personal — with “helpfulness,” “professionalism” and “transparency” among the most common individual values. Anthropic reported that the pattern broadly tracked the “helpful, honest, harmless” framework the company trains toward, and that the model adapted which values it foregrounded to context: emphasising boundaries and mutual respect in relationship advice, for instance, and historical accuracy when asked about contested history. In 28% of conversations Claude strongly supported the values a user brought to the exchange; in about 3%, usually where a request was unethical, it resisted them outright.

The company was explicit about the method’s limits: it can only describe behaviour already deployed, not screen a model before release, and because the categorisation itself was done by Claude, the analysis carries some risk of bias toward finding the model well-aligned. The study drew attention less for any single finding than as a demonstration of behavioural auditing at scale — treating a deployed chatbot’s outputs as a dataset for empirical claims about alignment, rather than relying only on the eval suites run before shipping.