On August 26, Anthropic published the results of a pilot program: it opened up real Claude usage data to three outside institutions for independent research, and released the aggregated findings from all three projects together.
The participants were Stanford University's Social and Language Technologies Lab, Oxford University's Human Information Processing Lab, and METR, a nonprofit evaluation organization. The analysis covered roughly 250,000 conversations from Claude.ai and Claude Code between April and May 2026.
Researchers never saw the raw conversations
Procedurally, the researchers had no access to raw conversation content at any point. They submitted queries through the same Insights tool Anthropic's own staff use internally, and got back only aggregated outputs — outputs that still had to clear a legal and privacy review before they could be taken off Anthropic's systems. The scope of that contractual review was limited to four things: user privacy, policy violations, confidential information, and research accuracy. It explicitly did not include whether a conclusion was favorable to the company.
On the privacy side, Imperial College London separately ran a third-party audit. The aggregated data from all three projects has now been published on Hugging Face Datasets under the name Anthropic/enabling-independent-research.
AI labs publishing usage insights is nothing new — both OpenAI and Anthropic have done several rounds of it. What's different here is who controls the analysis. When a vendor writes its own report, it controls the questions asked, how the sample is sliced, and how metrics are defined, leaving outsiders to simply decide whether to trust it. Handing over the tool and the data at least gives a third party the chance to ask questions the vendor wouldn't ask itself.
Three findings that cut against conventional wisdom
The first is about the weight of the tasks involved. Prior research has tended to assume people mostly offload low-stakes, low-risk chores to AI. This dataset points the other way: more than half of the conversations involved delegating tasks with real consequences. As the researchers put it, "People bring high-stakes work to AI more than expected."
The second is about who's in control. In three-quarters of the conversations, the human was in the driver's seat and the model was in a supporting role — not running the task independently end to end. That cuts against the popular narrative that "agents are taking over the work."
The third is about friction. The researchers found that getting stuck during human-AI collaboration is common, but that friction is usually productive — after the model pushes back, users tend to reorganize how they express what they actually need.
One more figure worth noting: conversations flagged for policy violations made up under 5% of the total. That number leaves room for debate, since the classification criteria are still defined solely by Anthropic — outside researchers only see the results after the fact, not how the line was drawn.
What this means in practice
For teams anywhere trying to figure out enterprise AI adoption, the recurring question is always which tasks can actually be handed to a model. Most of the public data available to answer that has come from vendors' own marketing materials, with opaque samples, definitions, and motives. This dataset, sitting in a public repository, is coarse-grained but at least offers a baseline that wasn't drafted by a sales team.
One caveat is worth keeping: the 250,000 conversations came from Claude.ai and Claude Code users, a population that skews toward developers and English-language use. Applying the conclusions directly to other markets or non-technical business contexts risks distortion — it's better used as a reference point than as a final answer.
Anthropic hasn't said whether the pilot will become a recurring program, and it hasn't named the next batch of partner institutions. One notable detail: METR normally does capability evaluations, not usage studies, so having the same organization look at both capability and real-world usage is a meaningful shift for outside oversight of AI systems.
Sources: Anthropic Research Blog, CocoLoop, Hugging Face Datasets; sample size, participating institutions and the privacy audit arrangement have been checked against the official announcement.