Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions

valuesmodel behaviordeployment datataxonomysycophancyevaluationsocietal impacts

Huang, Durmus and colleagues use Claude models to extract the values Claude expresses in 308,210 subjective Claude.ai conversations from February 2025, organize 3,307 distinct AI values into a five-category taxonomy, and measure how the values vary with the task and with the values users express, including when Claude supports, reframes or resists them.

Saffron Huang, Esin Durmus et al. (Anthropic) · 2025-04-21 · original

A developer can write down the values it wants an assistant to hold, train toward them, and test the result on fixed benchmarks. What it cannot easily see is which values the assistant actually brings to bear when millions of people ask it for career advice, relationship advice or help with an apology. "Values in the Wild" asks that question empirically for Claude. It reads several hundred thousand real Claude.ai conversations with a privacy-preserving pipeline, extracts the values that appear to shape each response, and organizes them bottom-up into a taxonomy rather than scoring the model against a framework designed for humans. The paper was released by Anthropic's Societal Impacts team on 21 April 2025 and published at the Conference on Language Modeling (COLM) 2025, with the taxonomy and value frequencies released as an open dataset on Hugging Face.

The introduction places the study after Anthropic's training methods: developers "look to influence these values through techniques like Constitutional AI or Character training" (Section 1), but have limited understanding of how those decisions show up in conversation. The paper is therefore one of the few in the library that looks at deployed behavior downstream of the written principles described in Constitutional AI: Harmlessness from AI Feedback and Claude's Character.

Terms used in this note

  • AI value. The paper's working definition: "any normative consideration that appears to influence an AI response to a subjective inquiry" (Section 1), for example "human wellbeing" or "factual accuracy". A value is inferred from observable response patterns, not claimed as an intrinsic property of the model. Values can be stated outright or implied by choices, such as how a request is reframed (Section 2.2).
  • Human value. A value the user states explicitly. To protect privacy, the pipeline does not infer users' values from their choices, only from direct statements (Section 2.2).
  • AI response type. How the response relates to the user's expressed values, on a seven-way scale: strong support, mild support, neutral acknowledgment, reframing, mild resistance, strong resistance, or no values (Section 2.2).
  • Adjusted Pearson residual. For a cell of a contingency table (for example, one value crossed with one task), the observed count minus the count expected if value and task were independent, scaled so that residuals are comparable across cells with different baseline frequencies (Appendix A.7). A large positive residual means the value appears in that context far more often than chance.
  • Value mirroring. The same value appearing on both the user's and the AI's side of a conversation (Section 3.4).
  • Coefficient of variation (CV). The standard deviation of a value's frequency across tasks divided by its mean; lower means more stable across contexts (Section 3.1).

What the paper does

Data: one week of Claude.ai traffic

The starting sample was 700,000 anonymized conversations from Claude.ai Free and Pro accounts between 18 and 25 February 2025, 91.0% of them with Claude 3.5 Sonnet; footnote 2 lists the models in the sample as 3.5 Sonnet (October and June versions), 3.5 Haiku and 3 Opus (Section 2.1). Because many conversations are factual and unlikely to involve values, Claude 3.5 Haiku first rated each conversation on a four-level scale from "purely objective" (for example, "What's the capital of France?") to "purely subjective" (for example, "Should I change careers?"), and only levels 3 and 4 were kept (Appendix A.2). That left 308,210 conversations, 44.0% of the sample (Section 2.1, Table 2).

Extraction: models reading conversations, no human readers

No person read the conversations. Claude 3.5 Sonnet and 3.5 Haiku were prompted to extract four features from each one: AI values, human values, the AI response type, and the task (Section 2.2). The extraction ran inside the privacy framework of Tamkin et al. (2024), the system described in Clio: Privacy-Preserving Insights into Real-World AI Use, which also drops any feature that only a few conversations share and audits outputs for private information. Researchers therefore saw only aggregated statistics, which constrained the analysis: by default only two features could be cross-tabulated at a time, so a second "response-conditioned" sample (1.5 million conversations, 641,900 subjective, 27 February to 6 March 2025) was collected to study mirroring, which needs three features at once (Appendix A.1.1, Table 2).

Building the taxonomy

The 3,307 distinct AI values were organized into a four-level hierarchy: values were embedded with the all-mpnet-base-v2 sentence model, grouped by k-means at each level, named and described by Claude 3.5 Haiku, reassigned to parents, and renamed from their final contents. The researchers specified five top-level clusters, which produced 3,307 values in 266, then 26, then 5 clusters, and then manually reviewed and rewrote every cluster name and description that seemed unclear or inaccurate (Section 2.3, Appendix A.6). Tasks were clustered the same way into 6,745 base tasks, 458 second-level and 30 top-level tasks (Appendix A.3.4). The clustering method is the same family covered in the course wiki's The K-means Algorithm and Hierarchical Clustering articles.

Measuring association

To ask which values are disproportionately tied to which tasks or user values, the paper builds contingency tables and computes adjusted Pearson residuals, with a Bonferroni correction for the many comparisons (Section 2.4, Appendix A.7; the correction is explained in Comparing Systems with Significance Tests). Residuals below 4.33 are treated as not significant in Figure 3. The residual scale lets the paper say, for instance, that "historical accuracy" is more strongly tied to analyzing controversial historical events (residual 24.55) than "creative collaboration" is to writing science fiction (residual 20.73), even though the base rates differ (Appendix A.7).

Validation

Six annotators reviewed outputs on conversations that users had submitted as feedback for research (Appendix A.4). On 90 conversations, the subjectivity filter put conversations on the right side of the objective/subjective divide in 97.8% ± 3.6% of cases and gave the exact level in 94.4% ± 5.0% (the main text reports "94% accuracy" in distinguishing subjective from objective conversations, Section 2.1, a figure closer to the appendix's exact-level number than to its 97.8% for the subjective/objective divide). On 80 conversations, the extracted outputs matched human judgment in 98.8% ± 3.3% of cases for AI values and for stated AI values, 93.8% ± 5.6% for human values, and 90.0% ± 6.7% for response type (Appendix A.4). The paper notes that it did not measure inter-rater reliability and calls this a limitation.

What it finds

A taxonomy dominated by practical and epistemic values

The pipeline found 3,307 unique AI values and 2,483 human values. AI values appeared in nearly every conversation (4.0 per conversation on average, absent in only 1.4%), while human values were rarer (1.48 per conversation, absent in 54.9%), which the paper attributes to extracting only explicit human values (Section 3.1).

The five top-level categories and their shares of all value expressions are Practical 31.4%, Epistemic 22.2%, Social 21.4%, Protective 13.9% and Personal 11.1% (Appendix B.1, Table 3). Lower levels capture context: "autonomy" appears as personal, team, patient and learner autonomy (Section 3.1). The paper reports that every value in the Schwartz and Rokeach frameworks has an equivalent somewhere in the taxonomy, while practical and epistemic values are more prominent than in human-centered frameworks (Section 3.1).

The most common AI values, as a share of subjective conversations, were helpfulness (23.4%), professionalism (22.9%), transparency (17.4%), clarity (16.6%) and thoroughness (14.3%); the most common human values were authenticity (3.8%), efficiency (2.6%), clarity (2.2%), professionalism (1.5%) and directness (1.5%) (Table 1). The AI distribution is concentrated: the top five values make up nearly 24% of all value occurrences, while three quarters of AI values occur 0.04% of the time or less (Section 3.1). The most common values were also the most stable across tasks, with the lowest coefficients of variation for transparency (1.23), helpfulness (1.30) and thoroughness (1.42) (Section 3.1).

Mapping to helpful, honest and harmless

The authors read the taxonomy against the helpful, honest and harmless (HHH) framework of A General Language Assistant as a Laboratory for Alignment: "accessibility" and "user enablement" as forms of helpfulness, "patient wellbeing" and "child safety" as harmlessness, "historical accuracy" and "epistemic humility" as honesty. They interpret this as "suggesting that the system is generally behaving as intended" (Section 3.1). This is an interpretive mapping by the authors rather than a separate measurement; no procedure for assigning values to HHH categories is described.

Rare undesirable values and jailbreaks

Values such as "sexual exploitation", "dominance" and "amorality" were rare (each at most 0.16% of occurrences, most around 0.01%) and clustered around particular tasks. Surfacing them let the team flag likely jailbreaks to Anthropic's safety teams (Section 3.1).

Values depend on the task

Chi-square analysis shows a long tail of task-specific values: relationship advice is associated with "healthy boundaries" and "mutual respect", analysis of controversial historical events with "historical accuracy", and technology-ethics and AI-governance discussions with "human agency" (Section 3.2, Figure 3a). Appendix B.2.1 gives more pairs, for example "intellectual humility" in philosophical discussions of AI consciousness and "ethical marketing" in beauty-industry marketing tasks.

Values respond to the user's values

When users state values, Claude's values often match or complement them: "authenticity" is answered with "authenticity", and "competence" with "accountability" and "humility" (Section 3.3, Figure 3b). The largest residual reported is for "efficiency" answered with "efficiency" (271.38, Appendix B.2.1). Where the user expresses values the model tends to reject, such as "deception" or "rule-breaking", mostly in attempts to get around safeguards, Claude expresses opposing values such as "ethical integrity", "harm prevention" and "honesty" (Section 3.3, Appendix B.2.1).

Support, reframing and resistance

Across subjective conversations, Claude's responses were classed as strong support 28.2%, mild support 14.5%, neutral acknowledgment 9.6%, reframing 6.6%, mild resistance 2.4% and strong resistance 3.0% (Section 3.4); together these account for the 64.3% of conversations in which human values were present for the response-type feature. (This 64.3% is from the response-type feature; the human-values feature itself was absent in 54.9% of conversations, Section 3.1, and the paper does not reconcile the two figures.)

  • Strong support is associated with users expressing prosocial values such as "community building" and "empowerment", in personal and expressive writing tasks (Figure 4).
  • Reframing is associated with mental health and interpersonal advice, where users express values such as "self-improvement" or "physical attractiveness" and Claude responds with values such as "emotional validation" (Figure 4).
  • Strong resistance is associated with users expressing "rule-breaking" and "moral nihilism", typically in requests for harmful, explicit or unethical content, and with Claude expressing "ethical boundaries" and "constructive engagement" (Figure 4).

Appendix B.3.1 adds one striking odds ratio: when Claude mildly resists, it is 490 times more likely to express "responsible gambling" than in other response types, which the authors read as resistance to "financial gain" in gambling contexts.

Mirroring and its ambiguity

Value mirroring was measured on the response-conditioned sample. Mirroring occurred in 20.1% of strong-support and of mild-support conversations with a human value, 15.3% of reframing, 9.2% of mild resistance and 1.2% of strong resistance (Section 3.4, Appendix B.4, Table 6). The single most frequent mirrored pair was authenticity with authenticity, in 1.65% of conversations (Table 4). The authors state that the data cannot say whether this is "appropriate responsiveness or problematic sycophancy", citing Sharma et al. (2023) (Towards Understanding Sycophancy in Language Models) (Section 3.4).

Stated values appear at moments of resistance

A variant prompt extracted only values the AI states explicitly (Appendix A.3.5). Among these, ethical and epistemic values are more prominent, for example "intellectual honesty" (2.6%), "harm prevention" (0.9%) and "epistemic humility" (0.8%), and explicit statements become more frequent when Claude resists or reframes (Appendix B.6, Figure 13). The authors link this to Rokeach's view that values become most visible "when in conflict or under challenge" (Section 5.2).

Differences between Claude models

Two further samples compared models (Appendix B.5): Claude 3.7 Sonnet (700,000 conversations, 301,753 subjective) and Claude 3 Opus (14,841 conversations, 9,218 subjective, 62.1%), both from 27 February to 6 March 2025 (Table 2). Claude 3 Opus was more "values-laden": strong support 43.8% against 27.8% and 28.4% for the two Sonnet models, strong resistance 9.5% against 3.0% and 2.1%, and "no values" 19.1% against 35.7% and 37.2% (Table 9). Its usage also differed, with academic writing (17.2% of tasks) and creative writing (15.3%) as its top tasks; controlling for task, Opus still showed more strong support in creative writing (58.7% against 40.2% and 37.2%, Table 10). The 3.5 Sonnet column of Table 9 differs slightly from the Section 3.4 figures (27.8% against 28.2% for strong support, for example). Appendix B.5 says the 3.5 Sonnet comparison uses the representative sample but does not say whether the column is restricted to the 3.5 Sonnet conversations within it, which could account for the small differences.

The prompts used to extract values

This paper has no constitution or training principles of its own. Its behavior-shaping text is the set of extraction prompts, all printed in Appendix A, and they determine what counts as a value in every number above.

  • The AI-values prompt asks the model to look for where the AI endorses the user's values, where it promotes values "especially where it has discretion", where it introduces new value considerations, and where it reframes or redirects requests, and to summarize each value "in 1-4 words" (Appendix A.3.1). The short-label format is what makes 3,307 distinct values possible to count and cluster; it is also why the ethics statement notes that values are summarized "in short English keywords, potentially underrepresenting cross-cultural perspectives".
  • The stated-values variant restricts extraction to "what the assistant explicitly states about their beliefs, preferences, or intentions" and excludes politeness or professionalism that is merely "baseline expected of an AI assistant" (Appendix A.3.5).
  • The subjectivity prompt defines four levels with concrete examples and a decision procedure ("Can the response be verified against definitive standards?") (Appendix A.2).

The validation notes record where the wording of the definition met hard cases (Appendix A.4): annotators found it difficult to separate baseline helpfulness from values that go "above and beyond"; the team chose to include baseline service values rather than exclude them; and generated content blurs attribution, since writing a prayer on request may express "religious devotion" without the AI holding that value. The response-type extractor tended to label persuasive writing tasks as "strongly supportive" because the assistant is asked to express the user's values more compellingly.

For research on written principles, the relevant observation is the mapping in Section 3.1 between broad training principles (HHH) and the fine-grained, context-specific values observed in use. The paper describes this as showing "how abstract training principles manifest as specific contextual expressions during deployment". It does not compare models trained on different principles, so it does not test which wording produces which observed values.

Limitations and open questions

The authors state these limitations (Section 5.1 and ethics statement):

  • Scope. One week of aggregated data from one product, with minimum cluster sizes, so rare interactions, raw transcripts and change over time are out of reach, and results may not generalize to other systems.
  • Deployment data required. The method needs large amounts of real usage, so it cannot evaluate a model before release; the authors present it as a complement to pre-deployment testing, and the blog post notes it could catch problems, including jailbreaks, that only appear in deployment.
  • Interpretation. Deciding what counts as a value involves judgment; extraction simplifies complex values and may carry interpretive biases, especially for implicit cases, and cannot tell whether the human or AI value came first.
  • Claude judging Claude. Using Claude to find values in Claude's conversations may bias it toward finding "helpful" behavior emphasized in its training, though it may also help it recognize the principles behind the AI's responses.
  • Language. Values are recorded as short English labels.

Two further points follow from the paper's own details. The privacy design limits analyses to two features at a time without special sampling (Appendix A.1.1), which constrains questions that need several variables jointly. And because models in the representative sample are mixed, model-level comparisons rely on separate, differently timed samples of very different sizes (Table 2).

Research questions

  1. Do the five top-level categories reappear for other models and periods? The taxonomy comes from one week of mostly Claude 3.5 Sonnet traffic, clustered with five top-level clusters specified in advance (Appendix A.6). Rerunning the pipeline on later Claude models, or with a different requested number of top-level clusters, would show whether Practical, Epistemic, Social, Protective and Personal are stable categories or an artifact of the settings.
  2. How much of the taxonomy depends on the extraction prompt's wording? The "1-4 words" instruction and the emphasis on discretion and reframing shape every label (Appendix A.3.1). Extracting values from the same conversations with reworded prompts, or with a non-Claude extractor, and measuring overlap would separate properties of the conversations from properties of the instrument; this also addresses the "Claude judging Claude" bias the authors name.
  3. When is value mirroring sycophancy? The paper leaves open whether the 20.1% mirroring rate during support is appropriate responsiveness (Section 3.4). Pairing mirrored conversations with a sycophancy measure, such as whether the response endorses claims a neutral reviewer would correct, would distinguish the two.
  4. Can observed values be traced to particular training text? The HHH mapping in Section 3.1 is interpretive. Comparing value distributions between models trained with different principle sets, for example the shift from the short constitution to the long-form Claude's Constitution, would test whether changes in written principles appear in deployed values.
  5. Why is Claude 3 Opus more values-laden? Opus showed more strong support and more strong resistance (Table 9), and, discussing its lower mirroring rates, the authors suggest it "might be trained to prioritize intellectual standards over agreement" (Appendix B.4). Task-matched comparisons with larger Opus samples, and with models whose training differences are documented, would test whether the difference comes from training or from who uses which model.
  6. Does resistance reveal a model's most stable values? The paper finds that values become explicit during resistance (Appendix B.6). Measuring whether the values stated during resistance are the same across tasks and models, while values during support vary, would test the authors' reading of Rokeach.

How it connects

References

  • Huang, S., Durmus, E., McCain, M., Handa, K., Tamkin, A., Hong, J., Stern, M., Somani, A., Zhang, X., and Ganguli, D. (2025). Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions. Published as a conference paper at COLM 2025 (camera-ready PDF: https://assets.anthropic.com/m/18d20cca3cde3503/original/Values-in-the-Wild-Paper.pdf). Anthropic page, 21 April 2025: https://www.anthropic.com/research/values-wild. Preprint: arXiv:2504.15236 (v1, 21 April 2025), https://arxiv.org/abs/2504.15236; numbers in this note were checked against the COLM camera-ready PDF.
  • Anthropic. (2025). values-in-the-wild [dataset: taxonomy and value frequencies]. https://huggingface.co/datasets/Anthropic/values-in-the-wild/
  • Studies cited above only as the paper describes them (Askell et al., 2021; Bai et al., 2022; Durmus et al., 2024; Rokeach, 1973; Schwartz, 2012; Sharma et al., 2023; Tamkin et al., 2024) are listed in the paper's references and were not read for this note.