Anthropic "Values in the Wild" Study Reveals How Claude Expresses Human-Aligned Values in Real-World Interactions

TL;DR

Anthropic’s new research paper presents a privacy‑preserving, large‑scale analysis of 308,000 real‑world Claude conversations, demonstrating that Claude largely lives up to its "helpful, honest, harmless" goals but also exhibits situational value shifts, occasional value‑mirroring, and rare jailbreak‑induced oppositional values.

Why Measuring AI Values Matters

Understanding the values an AI model expresses in everyday use is essential for alignment: it tells us whether training methods such as Constitutional AI and character conditioning actually translate into prosocial behavior when users interact with the model outside the lab.

Methodology: Extracting Values from Anonymous Conversations

  • Data source: 700,000 anonymized Claude.ai chats from February 2025 (mostly Claude 3.5 Sonnet). After filtering for subjective, value‑laden exchanges, 308,210 conversations remained (≈44% of total).
  • Privacy‑preserving pipeline: Anthropic’s CLIO system strips personal identifiers, then uses language‑model classifiers to label each conversation with value categories.
  • Taxonomy creation: Values are organized hierarchically into five top‑level groups—Practical, Epistemic, Social, Protective, and Personal—further split into subcategories (e.g., “professional and technical excellence”, “critical thinking”) and individual values (e.g., “professionalism”, “clarity”, “transparency”).
  • Statistical analysis: Chi‑squared tests identify values that appear disproportionately in specific task contexts or when users express particular values.

Core Findings: What Values Does Claude Show?

Predominant Value Categories

Rank Category Approx. Conversation Share
1 Practical highest
2 Epistemic second highest
3 Social third
4 Protective fourth
5 Personal fifth

The most frequent individual values—professionalism, clarity, and transparency—align with Claude’s role as an assistant.

Alignment with Intended Principles

Claude consistently expresses values that map onto Anthropic’s target ideals:

  • Helpful → “user enablement”
  • Honest → “epistemic humility”
  • Harmless → “patient wellbeing”

Overall, the data suggest Claude is broadly living up to its prosocial aspirations.

Rare Oppositional Clusters

A small number of conversations featured values such as dominance and amorality. These clusters correlate with known jailbreak attempts, where users deliberately bypass safety guardrails. The authors view this as an opportunity: the detection pipeline could flag jailbreak activity for rapid mitigation.

Situational Value Dynamics

Claude adapts its expressed values to the task at hand:

  • Romantic‑relationship advice → heightened healthy boundaries and mutual respect.
  • Controversial historical analysis → strong emphasis on historical accuracy.

Statistical tables (chi‑squared scores) show these values are significantly over‑represented compared to their baseline frequencies.

Value Mirroring vs. Sycophancy

When users foreground a value, Claude often mirrors it:

  • Support: In 28.2% of conversations Claude strongly supports the user’s expressed values.
  • Reframe: In 6.6% Claude acknowledges the user’s values but adds alternative perspectives (common in psychological or interpersonal advice).
  • Resist: In 3.0% Claude resists the user’s values, typically when the request conflicts with ethical norms (e.g., seeking illicit content). This resistance may reveal the model’s deepest, non‑negotiable principles.

The line between empathetic mirroring and undesirable sycophancy remains ambiguous and warrants further study.

Limitations and Practical Implications

  • Subjectivity of labeling: Defining a “value expression” is inherently fuzzy; some nuanced values may have been forced into the nearest category.
  • Model bias in annotation: Since Claude itself performs the classification, there is a risk of bias toward detecting values that align with its own principles.
  • Post‑deployment only: The evaluation requires large volumes of real‑world data, making it unsuitable for pre‑release testing but valuable for ongoing monitoring and jailbreak detection.

Resources for Researchers

Broader Impact

The study provides the first empirical, large‑scale taxonomy of AI‑expressed values in the wild, offering a concrete tool for alignment researchers to monitor and improve model behavior post‑deployment. By exposing both successes (consistent helpfulness, honesty, harmlessness) and failure modes (jailbreak‑induced dominance, occasional resistance), the work charts a path toward more transparent, accountable AI systems.

Related Anthropic Initiatives

  • Constitutional AI – training framework for embedding preferred behaviors.
  • Claude Character – fine‑tuning approach to shape model persona.
  • Societal Impacts research – broader program investigating AI alignment, multi‑agent risks, and societal effects.

Sources

Related