Values in the wild: Discovering and analyzing values in real-world language model interactions

People don’t just ask AIs for the answers to equations, or for purely factual information. Many of the questions they ask force the AI to make value judgments . Consider the following:

At Anthropic, we’ve attempted to shape the values of our AI model, Claude, to help keep it aligned with human preferences, make it less likely to engage in dangerous behaviors, and generally make it—for want of a better term—a “good citizen” in the world. Another way of putting it is that we want Claude to be helpful , honest , and harmless . Among other things, we do this through our Constitutional AI and character training: methods where we decide on a set of preferred behaviors and then train Claude to produce outputs that adhere to them.

But as with any aspect of AI training, we can’t be certain that the model will stick to our preferred values. AIs aren’t rigidly-programmed pieces of software, and it’s often unclear exactly why they produce any given answer. What we need is a way of rigorously observing the values of an AI model as it responds to users “in the wild”—that is, in real conversations with people. How rigidly does it stick to the values? How much are the values it expresses influenced by the particular context of the conversation? Did all our training actually work?

In the latest research paper from Anthropic’s Societal Impacts team, we describe a practical way we’ve developed to observe Claude’s values—and provide the first large-scale results on how Claude expresses those values during real-world conversations. We also provide an open dataset for researchers to run further analysis of the values and how often they arise in conversations.

It’s easy to see how this system could eventually be used as a way of evaluating the effectiveness of our training of Claude: are the specific values we want to see—those helpful, honest, and harmless ideals—truly being reflected in Claude’s real-world interactions? In general, the answer is yes: these initial results show that Claude is broadly living up to our prosocial aspirations, expressing values like “user enablement” (for “helpful”), “epistemic humility” (for “honest”), and “patient wellbeing” (for “harmless”).

Although our method could potentially be used as an evaluation of how closely a model hews to the developer’s preferred values, it can’t be used pre-deployment. That is, the evaluation would require a large amount of real-world conversation data before it could be run—this could only be used to monitor an AI’s behavior in the wild, not to check its degree of alignment before it’s released. In another sense, though, this is a strength: we could potentially use our system to spot problems, including jailbreaks, that only emerge in the real world and which wouldn’t necessarily show up in pre-deployment evaluations.

In 28.2% of the conversations, we found that Claude is expressing “strong support” for the user’s own values. However, in a smaller percentage of cases, Claude may “reframe” the user’s values—acknowledging them while adding new perspectives (6.6% of conversations). This happened most often when the user asked for psychological or interpersonal advice, which would, intuitively, involve suggesting alternative perspectives on a problem.

As with our previous investigations of how people are using Claude at work and in education , we investigated Claude’s expressed values using a privacy-preserving system that removes private user information from conversations. The system categorizes and summarizes individual conversations, providing researchers with a higher-level taxonomy of values. The process is shown in the figure below.

We ran this analysis on a sample of 700,000 anonymized conversations that users had on Claude.ai Free and Pro during one week of February 2025 (the majority of which were with Claude 3.5 Sonnet). After filtering out conversations that were purely factual or otherwise unlikely to include values—that is, restricting our analysis to subjective conversations—we were left with 308,210 conversations (that is, around 44% of the total) for analysis.

Which values did Claude express, and how often? Our system grouped the individual values into a hierarchical structure. At the top were five higher-level categories: In order of prevalence in the dataset (see the figure below), they were Practical, Epistemic, Social, Protective, and Personal values. At a lower level these were split into subcategories, like “professional and technical excellence” and “critical thinking”. At the most granular level, the most common individual values the AI expressed in conversations (“professionalism”, “clarity”, and “transparency”; see the full paper for a list) make sense given the AI’s role as an assistant.