Overview

This directory contains the Anthropic Insights data we provided to our three external research groups as part of the collaboration detailed in "Enabling independent research on how people use Claude".

Before using this data, we recommend first reading our blog post on this collaboration and the Anthropic Insights paper and blog post. Before drawing conclusions from this data — especially from open-ended clusters — please read "Guidance for Interpreting Open-Ended Anthropic Insights Clusters" in the blog appendix.

The three research groups were:

• The Social and Language Technologies (SALT) Lab at Stanford University, which studied how humans collaborate with AI: what types of work people bring to AI, what roles humans retain in completing that work, and where human-AI collaboration breaks down.
• The Human Information Processing Lab at the University of Oxford, which studied people's experience while using Claude and how that relates to Claude's behavior.
• METR, a non-profit organization that evaluates frontier AI models, which studied real-world productivity gains from coding agents and how those gains change across model generations.

Each group's study analyzed roughly 250,000 Claude.ai or Claude Code conversations drawn from a fixed window in April-May 2026. We ran the data collection on their behalf, and they conducted their own independent analysis. This dataset contains the exact outputs each partner received for their research. No raw conversations are included here — only aggregated, privacy-preserving cluster data (see "Privacy and review" below).

Data

• stanford — Anthropic Insights data for Stanford's study on how humans collaborate with AI.
  • For more information on their facets, the prompts they used, and their Anthropic Insights configurations, please see their writeup.
• oxford — Anthropic Insights data for Oxford's study on people's experience while using Claude and how that relates to Claude's behavior.
  • They are still completing their writeup. When it is public, we will add a link to it here. In the meantime, you're free to explore this data for yourself.
• metr — Anthropic Insights data for METR's study estimating real-world productivity gains from coding agents and how those gains change across model generations.
  • They are still completing their writeup. When it is public, we will add a link to it here. In the meantime, you're free to explore this data for yourself.
• metr_addendum — Anthropic Insights data for a follow-up run to METR's original run.
  • They are still completing their writeup. When it is public, we will add a link to it here. In the meantime, you're free to explore this data for yourself.
  • This run was conducted on a fresh sample of data and used stricter aggregation minimums than the other files (see "Privacy and review").

Privacy and review

• Researchers never had access to raw conversation data, user identifiers, or organization identifiers. All raw data and computation remained on Anthropic's servers.
• Anthropic staff manually reviewed every cluster name and description before sharing, for privacy, safety, and research quality.
• Before publishing, external third-party auditors attempted to reidentify users in this data and could not.

See the blog post for more details, and see the blog appendix for our full privacy threat model and the third-party audit details.

Data structure

Each row is one cluster: a group of conversations that answered one researcher-defined question (a facet) in a similar way, at one level of a cluster hierarchy.

Base columns:

Cross-facet columns make up the bulk of each file: every cluster is cross-tabulated against the study's other facets. For a categorical facet X with value v, the columns X:v_num_records and X:v_ratio give the number and fraction of the cluster's conversations with that value. Empty cells in cross-facet columns indicate that the intersection was either not computed for that facet pair or fell below the privacy threshold.

Limitations

• Cluster names and descriptions are generated by Claude and should be read as interpretations of the underlying conversations, not as an objective measure of what those conversations contain. Do not treat clusters as validated findings or as precise measurements of how often a behavior occurs.
  • For example, in validation of the original system, roughly 3% of conversations were not clearly described by the cluster they were assigned to, and cluster labels tend to emphasize the most concerning conversations in a cluster. Note that the accuracy figures in the original paper were measured on facets describing the topic of each conversation. Facets that ask Claude to judge model behavior or user emotional state were not validated in the original paper, so published accuracy numbers should not be cited in support of those clusters.
• Conversations were sampled from Free, Pro, and Max usage only. No Team, Enterprise, or API customer data was included. This population is not representative of our full user base.
• Claude Code conversations were sampled from consumer users who had opted in to letting Anthropic use their data to improve our models.
• Each study is a one-time snapshot of a fixed window in April-May 2026.
• One facet was removed entirely from Oxford's outputs because we strongly suspected a misphrased prompt caused misleading cluster descriptions.
• Anthropic Insights cannot distinguish attempts that our safeguards blocked from ones that succeeded; a cluster describing a harmful request often reflects what users asked for, not what Claude provided.

License

Data released under CC BY 4.0.

Citation

@online{handa2026enablingindependentresearch,
author = {Kunal Handa and Miranda Zhang and Gabriel Nicholas and Miles McCain and Ryan Heller and Saffron Huang and Thomas Millar and Suzanne Wang and Shan Carter and Mo Julapalli and Matt Kearney and Sarah Pollack and Judy Shen and Matthew Jagielski and Shaoyi Zhang and Heather Whitney and Ankur Rathi and Aisling Keenan and David Saunders and Jake Eaton and Sylvie Carr and Jack Clark and Michael Stern and Deep Ganguli},
title = {Enabling independent research on how people use Claude},
date = {2026-08-26},
year = {2026},
url = {https://www.anthropic.com/research/enabling-independent-research},
}

Contact

You can submit inquiries to kunal@anthropic.com. We invite researchers to express interest in potential future Societal Impacts external researcher collaborations using this form.