Anthropic Opens Claude Data to AI Research

Anthropic has allowed external researchers for the first time to study real Claude conversations without accessing the original chats. Three groups analyzed aggregated data from around 250,000 conversations and published initial findings on how people use AI.

The pilot program ran between April and May 2026. It involved Stanford University's SALT Lab, Oxford's Human Information Processing Lab and METR, a nonprofit organization that evaluates advanced AI models.

The researchers formulated their own questions, and Anthropic processed the data through Anthropic Insights, a tool that replaces individual conversations with categories and percentages. This meant the teams could not read specific chats or identify their authors.

People use AI for important tasks

The Stanford team studied how people collaborate with Claude. One of its findings challenges a common assumption: that most users keep sensitive decisions to themselves and leave routine tasks to AI.

Instead, more than half of the conversations analyzed included tasks with important consequences. This use was especially common when people asked for professional guidance, particularly on legal or financial matters.

That does not mean Claude made all those decisions. In nearly three out of four conversations, the person set the direction and used Claude as support. Users also typically adapted the responses rather than copying them directly.

The collaboration created friction too: responses that did not fit, unclear instructions or results that required several corrections. According to the study, that friction was not always a problem. Reviewing what the AI had understood could help people clarify their goals and improve the final result.

Claude's behavior influences the experience

The Oxford team analyzed how people feel when using Claude and how that relates to the system's behavior. Its preliminary findings point to specific patterns:

• When Claude responded in a warm tone, people tended to express themselves more positively.
• When it rejected a request or showed disagreement, users often insisted or argued.
• When it offered more unconventional responses, people appeared to engage more intellectually.
• When it simply helped with the task, the conversation tended to reflect satisfaction.

The researchers also found similarities between using Claude and browsing the internet in everyday life. States such as concentration, frustration and enjoyment appeared in similar ways in both environments. The full study is still being prepared.

New models could save more time when programming

METR examined conversations from Claude Code, Anthropic's assistant for programming tasks. Its analysis is not yet complete, but the initial results indicate that newer models could save more time than earlier ones.

To estimate this, the team compared how much time Claude calculated a task would have taken without AI with the time it took to complete using different model versions. Those estimates showed a reasonable correlation with the actual times recorded in an earlier study with developers.

METR now wants to study whether this method can also measure how much AI speeds up researchers' work. The question is becoming more important as systems begin to take part in scientific and technical development tasks.

More open access, but difficult to scale

Anthropic says the researchers were able to publish their findings even when they were uncomfortable for the company. Their review rights were limited to protecting privacy, preventing information that could facilitate abuse, safeguarding internal secrets and correcting accuracy errors.

The system also underwent a privacy audit by Imperial College London. Less than 5% of the categories and conversations in each study had to be modified or removed because they included details about prohibited uses or ways to bypass safety measures.

The problem is that the process is slow and expensive. Researchers cannot review the original conversations to check whether a question was properly formulated. And a category that works with a public dataset can produce misleading results when applied to real-world Claude use.

Anthropic has published the aggregated data from all three projects and is evaluating how to expand the program. For you, the significance is that the AI debate could depend less on what companies themselves say and more on independent studies of how these systems are used in practice. The next step will be to see whether this access can grow without putting privacy or researchers' freedom at risk.