Anthropic Opens Real Claude Usage Data to Researchers

Episode 1 · Anthropic Opens Real Claude Usage Data, 250K-Conversation Study Reveals Human-AI Collaboration Patterns (2026-08-27, 13 posts)

Anthropic has announced it is opening privacy-protected real-world Claude usage data to external researchers and institutions for the first time, provided as aggregated telemetry through the Anthropic Insights tool (formerly known as Clio). This marks the first time research of this kind—previously confined to AI labs' internal efforts—has been made available externally. Jack Clark noted that AI systems are so complex that relying solely on internal lab evaluation is arrogant and mistaken; outside teams must be able to study AI's real-world impact. During the pilot, three external research groups already used this data to complete independent analyses.

Confirmed

• Form of data release: Aggregated, anonymized, privacy-protected telemetry data via Anthropic Insights (formerly Clio), with the first pilot round aimed at external researchers and institutions.
• Pilot outcomes: Three external research groups completed independent studies using the data; among them, Diyi Yang's team at Stanford conducted a large-scale study of human-AI collaboration.
• Study scale: Based on 249,834 real Claude conversations.
• Findings on task criticality: Contrary to the old view that "users only delegate low-stakes tasks to AI," users frequently delegate tasks in critical, hard-to-reverse domains such as law and finance; moreover, the higher the stakes, the less willing users are to adapt to or criticize AI output, and the more they shift toward trying to understand it.
• Collaboration patterns: Human-led collaboration accounted for 72%; AI exhibited teaching behavior in 67% of conversations.
• Friction findings: 49.7% of conversations contained some form of friction; some friction was productive, leading to clarification or learning; users actively attempted recovery in 78.7% of friction conversations. The most effective recovery strategies included questioning the model's reasoning, asking for simplified output, and correcting factual errors, while simply apologizing proved far less effective.

Why it matters

• Diyi Yang's team concluded that although the study initially focused on hypotheses around task criticality, agency, and friction, all conclusions ultimately pointed to the importance of the "human" factor: improving AI capabilities alone is not enough to achieve strong human-AI synergy.
• Externalizing this data breaks AI labs' monopoly on usage data, offering a new paradigm for independently evaluating AI's real societal impact.

• Anthropic opens privacy-preserved Claude data to external researchers — AnthropicAI · 2026-08-27
• Anthropic pilots external access to privacy-preserved Claude data — Diyi_Yang · 2026-08-27
• Users Less Likely to Critique AI in High-Stakes Tasks — EchoShao8899 · 2026-08-27
• Anthropic enables research on human-AI collaboration using real data — Diyi_Yang · 2026-08-27
• Users Delegate High-Stakes Tasks to AI at Surprisingly High Rates — EchoShao8899 · 2026-08-27
• Study Finds Friction in 49.7% of Human-AI Conversations, Reveals Effective Recovery Strategies — EchoShao8899 · 2026-08-27
• Anthropic Study: Human-AI Collaboration at Scale Across 250k Conversations — EchoShao8899 · 2026-08-27
• Anthropic opens Claude usage data to independent researchers — EchoShao8899 · 2026-08-27
• Anthropic opens pilot for external researchers to access privacy-preserved Claude data — mattbeane · 2026-08-27
• Study on 250k Conversations: How Humans Collaborate with AI — Diyi_Yang · 2026-08-27
• Anthropic study: 72% of AI conversations remain human-led — Diyi_Yang · 2026-08-27
• Stanford Study: AI-Augmented Work Dominates in 249k Real Claude Conversations — EchoShao8899 · 2026-08-27
• Anthropic Opens Privacy-Preserved Claude Data for Research — Diyi_Yang · 2026-08-27

Episode 2 · Anthropic Analyzes 400K Claude Code Sessions: Domain Expertise Drives Better Results (2026-08-28, 2 posts)

Anthropic analyzed roughly 400,000 Claude Code sessions from about 235,000 users over seven months, finding that humans handle 70% of planning decisions while AI handles 80% of execution. The study also shows domain experts achieve notably higher success rates with the tool.

• Study of 400k Claude Code sessions: Domain expertise beats coding skills — alex_verem · 2026-08-28
• Anthropic Study: Domain Experts See Higher Success Rates with Claude Code — alex_verem · 2026-08-28