On August 26, 2026, Anthropic officially announced the results of a pilot program that, for the first time, gave outside research institutions access to real Claude usage data. Research teams at Stanford University and the University of Oxford independently analyzed roughly 250,000 Claude.ai and Claude Code conversations from April–May 2026 to study how people actually delegate work to AI and collaborate with it.
What exactly did Anthropic announce?
Anthropic opened its own real Claude usage data to three outside research institutions for the first time, and published their findings as-is. Researchers accessed the data through Anthropic Insights (formerly named Clio) — the same privacy-preserving analysis tool Anthropic's own teams use to study usage patterns, which lets researchers see patterns aggregated across many conversations without anyone reading individual conversation transcripts. Each of the three institutions designed its own research questions independently; Anthropic only ran the data collection on their behalf and had no say in the analysis itself.
Why open real usage data to outsiders at all?
Understanding how AI actually affects society requires more than analysis run by the companies that hold the data. Until now, outside researchers had only two options.
- Lab-published analyses — reflect real usage, but tend to answer the lab's own questions rather than the researcher's.
- Public datasets — researchers can study them freely, but they skew toward casual use and may not reflect how most people actually use AI.
Anthropic judged that neither option was sufficient for independent research on real-world AI usage, so it ran this pilot to give outside institutions direct access instead.
Which institutions took part?
Three institutions with different research focuses joined the pilot.
| Institution | Affiliation | Research focus |
|---|---|---|
| SALT Lab (Social and Language Technologies Lab) | Stanford University | How humans collaborate with AI |
| Human Information Processing Lab | University of Oxford | How Claude's behavior relates to how people feel while using it |
| METR | Non-profit that evaluates frontier AI models | Third partner in this pilot |
All three institutions worked with aggregate data — data summarizing overall patterns rather than individual conversations — covering roughly 250,000 Claude conversations from April–May 2026. The source material Anthropic published does not include a separate breakdown of METR's specific findings, so this article only notes METR's participation as the third partner.
What did Stanford's SALT Lab find about how much people trust AI with important work?
Over half of Claude conversations involved people delegating "consequential" work — work that affects others or is hard to undo — to AI. That runs counter to prior assumptions.
- People bring higher-stakes work to AI than expected — prior research suggested people mostly delegate low-accountability tasks and keep consequential work for themselves. SALT Lab found the opposite: over half of conversations involved consequential work, especially when people sought professional guidance on legal or financial questions.
- People usually direct and oversee the work — in nearly three-quarters of conversations, people set the direction while Claude assisted, and people usually adapted Claude's output rather than using it verbatim. Still, how much people understood or learned from what Claude produced varied a lot from person to person.
- Friction is common — and often productive — friction (unclear requests, misunderstood intent, output that isn't quite right) is a normal part of working with AI. The time people spent seeing how Claude attempted a task, spotting where their request was unclear, and iterating led to better results — it pushed people to clarify what they actually wanted.
What did the Oxford team find about Claude's tone and human emotion?
How Claude behaves and how people feel while using it are closely linked. Oxford's Human Information Processing Lab studied how people's emotional state connects to Claude's behavior during a conversation.
The researchers also found that patterns of absorption, frustration, and enjoyment in Claude conversations closely resembled patterns found in a separate study of everyday internet browsing — suggesting how people experience AI isn't dramatically different from how they experience other digital activity. The Oxford team's full writeup is still being completed; according to the official announcement, a link will be added once it's public.
How was privacy protected?
Researchers never saw individual conversation transcripts — only patterns aggregated and anonymized by Anthropic Insights. Anthropic also limited its own contractual review rights to four areas.
- User privacy
- Information that could help people violate usage policies
- Anthropic's confidential information
- Research accuracy
Beyond those four areas, Anthropic had no say in the findings, and researchers were free to publish results even when inconvenient for Anthropic. The company also said it conducted an additional privacy audit — a review confirming its privacy protections actually held — covering all data shared with third-party researchers.
What can you actually take away from this if you use Claude?
Beyond the statistics, these findings point to some practical habits worth adopting.
- Friction isn't a sign you're doing it wrong — if you don't get the answer you want on the first try and need several rounds of back-and-forth, that's likely normal, not inefficiency. The research found this kind of friction tends to produce better results, not worse ones.
- You should keep your hands on the wheel — most successful collaboration in the study involved people setting direction and reviewing or editing Claude's output rather than using it verbatim. Reviewing before trusting is a habit the data itself supports.
- The more consequential the task, the more it needs a final check — since many people already bring legal, financial, and other high-stakes questions to AI, it's worth double-checking the output yourself rather than trusting it outright for anything hard to undo.
Frequently asked questions
Q. What exactly is Anthropic Insights?
It's the privacy-preserving analysis tool Anthropic uses to study Claude usage patterns. It shows patterns aggregated across many conversations without anyone reading individual transcripts. It was previously named Clio.
Q. Where did the data for this research come from?
It came from roughly 250,000 Claude.ai and Claude Code conversations between April and May 2026, provided to researchers as aggregated data rather than individual transcripts.
Q. Where can I read the full findings?
SALT Lab published a separate writeup with its full results, and Anthropic also released the aggregate data from each project. The original announcement is available on Anthropic's official research page.
Q. Does this apply to users outside the US too?
The research wasn't broken down by country — it covers overall Claude.ai and Claude Code usage patterns. That said, the conclusion that people should set direction and review AI output is a general usage principle, not one specific to any single country.
Anthropic said it plans to scale this pilot further and has set up an expression-of-interest form for researchers who want to work with it in the future. The original announcement is available on Anthropic's official research page.