Anthropic Opens Privacy-Preserving Claude Usage Data to Independent Researchers

A new model for independent AI-usage research

Anthropic published results on August 26, 2026 from a pilot designed to give outside researchers more independence when studying how people use frontier AI systems in the real world.

Three external groups—the Stanford Social and Language Technologies Lab, the University of Oxford's Human Information Processing Lab, and METR—designed their own studies using Anthropic Insights, the company's privacy-preserving system for analyzing usage patterns. Anthropic ran the data collection on their behalf, and the groups independently analyzed roughly 250,000 Claude.ai or Claude Code conversations from April and May 2026.

Anthropic says it limited its contractual review rights to privacy, misuse-enabling information, confidential company information and research accuracy. It otherwise allowed partners to publish findings even when those findings might be inconvenient for the company. Aggregate data from the projects is also being released publicly.

What the first studies found

The Stanford SALT Lab examined human-AI collaboration. Anthropic's summary says more than half of the conversations in that study involved people delegating consequential work to Claude, with professional legal and financial guidance among the areas where consequential use appeared frequently.

At the same time, people generally remained in a directing role. Anthropic reports that in nearly three-quarters of conversations, users set the direction while Claude assisted, and users usually adapted the model's output rather than using it verbatim. The study also found that some friction in collaboration can be productive by forcing users to clarify their intent and refine instructions.

Oxford's Human Information Processing Lab is studying how users' emotional states relate to Claude's behavior. Its early results suggest patterns in model behavior and user experience move together—for example, warmer model behavior appearing alongside more positive user states. The full write-up was still in progress at publication time.

METR is studying productivity changes across generations of coding agents using Claude Code data. Anthropic describes those results as preliminary, with an early indication that more capable models may save users more time. METR has not yet published its final analysis.

Why the access model matters

Real-world AI usage data is concentrated inside model companies. External researchers can study public datasets, but those datasets may overrepresent casual or self-selected use. Company-authored usage reports have the opposite limitation: they reflect real traffic but usually answer questions chosen by the company itself.

Anthropic's pilot attempts to create a middle layer. Researchers can choose their own questions while analysis happens through a privacy-preserving interface rather than through unrestricted access to raw private conversations.

That structure does not make the work fully independent in the same way as open access to raw data. Anthropic still operates the analysis environment and applies privacy and safety constraints. The company also reports that small portions of categories or conversations were altered or removed in each study when privacy or policy issues arose, and that researchers were informed about those changes.

What comes next

Anthropic says it conducted an additional privacy audit and is considering how to expand the program while preserving privacy, safety and research quality. It has opened an expression-of-interest process for researchers who could use access to Anthropic Insights for studies that are difficult to conduct with public datasets alone.

If this model scales, it could become a useful template for AI accountability: independent question-setting, auditable privacy constraints, publication freedom and public aggregate datasets. The key test will be whether future access becomes broad enough, reproducible enough and sufficiently independent to support robust external scrutiny across both beneficial and harmful real-world AI use.